Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	README.md
#	examples/server/README.md
#	examples/speculative/speculative.cpp
#	flake.lock
#	ggml/src/CMakeLists.txt
#	scripts/sync-ggml.last
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo
2024-11-14 21:40:52 +08:00
parent bfa118ee45 2a82891a85
commit df080b074d
41 changed files with 132106 additions and 130341 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/count-equal.cu
+1 -1
View File
@@ -44,7 +44,7 @@ void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t ne = ggml_nelements(src0);
GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int");
const int64_t dne = GGML_PAD(ne / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
const int64_t dne = GGML_PAD((ne + 4*nsm - 1) / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
CUDA_CHECK(cudaMemsetAsync(dst_d, 0, ggml_nbytes(dst), stream));
ggml/src/ggml-metal.m
+3 -6
View File
@@ -596,17 +596,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
ctx->kernels[i].pipeline = nil;
}
/*
GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
(int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
(int) kernel->pipeline.threadExecutionWidth); \
*/
#define GGML_METAL_ADD_KERNEL(e, name, supported) \
if (supported) { \
struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
kernel->pipeline = [device newComputePipelineStateWithFunction:metal_function error:&error]; \
GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
(int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
(int) kernel->pipeline.threadExecutionWidth); \
[metal_function release]; \
@@ -3046,6 +3041,8 @@ static void ggml_metal_encode_node(
bool use_vec_kernel = false;
// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
// for now avoiding mainly to keep the number of templates/kernels a bit lower
if (ne01 >= 4 || (ne00%128 != 0)) {
switch (src1->type) {
case GGML_TYPE_F16:
ggml/src/ggml-metal.metal
+73 -51
View File
@@ -2942,6 +2942,7 @@ kernel void kernel_flash_attn_ext(
half smax = -INFINITY;
// load the mask in shared memory
#pragma unroll(Q)
for (short j = 0; j < Q; ++j) {
device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*nb31);
@@ -2968,7 +2969,7 @@ kernel void kernel_flash_attn_ext(
// we can read directly from global memory
device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
#pragma unroll
#pragma unroll(D8)
for (short i = 0; i < D8; ++i) {
k8x8_t mk;
simdgroup_load(mk, pk + i*8, nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
@@ -2989,7 +2990,7 @@ kernel void kernel_flash_attn_ext(
simdgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll
#pragma unroll(4)
for (short k = 0; k < 4; ++k) {
k8x8_t mk;
@@ -3067,7 +3068,7 @@ kernel void kernel_flash_attn_ext(
s8x8_t mm;
simdgroup_load(mm, ss + 2*C, TS, 0, false);
#pragma unroll
#pragma unroll(D8)
for (short i = 0; i < D8; ++i) {
simdgroup_multiply(lo[i], mm, lo[i]);
}
@@ -3082,7 +3083,8 @@ kernel void kernel_flash_attn_ext(
if (is_same<vd4x4_t, v4x4_t>::value) {
// we can read directly from global memory
device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
#pragma unroll
#pragma unroll(D8)
for (short i = 0; i < D8; ++i) {
v8x8_t mv;
simdgroup_load(mv, pv + i*8, nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
@@ -3103,7 +3105,7 @@ kernel void kernel_flash_attn_ext(
simdgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll
#pragma unroll(4)
for (short k = 0; k < 4; ++k) {
v8x8_t mv;
@@ -3196,6 +3198,7 @@ kernel void kernel_flash_attn_ext(
simdgroup_load(ms0, ss + 2*C, TS, 0, false);
simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
#pragma unroll(D8)
for (short i = 0; i < D8; ++i) {
o8x8_t t;
@@ -3356,8 +3359,8 @@ kernel void kernel_flash_attn_ext_vec(
const short D4 = D/4;
const short D16 = D/16;
const short NW = N_SIMDWIDTH;
const short NL = NW/4;
const short SH = 2*C; // shared memory per simdgroup
const short NL = NW/4; // note: this can be adjusted to support D%64 == 0 and D%32 == 0
const short SH = 2*C; // shared memory per simdgroup
const short T = D + nsg*SH; // shared memory size per query in (half)
@@ -3413,6 +3416,7 @@ kernel void kernel_flash_attn_ext_vec(
// load the queries from shared memory into local memory
q4x4_t mq[D16/NL];
#pragma unroll(D16/NL)
for (short ii = 0; ii < D16; ii += NL) {
mq[ii/NL] = sq4x4[ii + tx];
}
@@ -3448,26 +3452,33 @@ kernel void kernel_flash_attn_ext_vec(
// Q*K^T
{
// each simdgroup processes 1 query and 4 keys
// each simdgroup processes 1 query and 4 (NW/NL) keys
for (short cc = 0; cc < C/4; ++cc) {
qk_t mqk = 0.0;
qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 };
device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
#pragma unroll
#pragma unroll(D16/NL)
for (short ii = 0; ii < D16; ii += NL) {
const short i = ii + tx;
k4x4_t mk;
deq_k(pk + i/nl_k, i%nl_k, mk);
mqk +=
dot(mq[ii/NL][0], mk[0]) +
dot(mq[ii/NL][1], mk[1]) +
dot(mq[ii/NL][2], mk[2]) +
dot(mq[ii/NL][3], mk[3]);
// note: this is less precise than the version below
//mqka[0] += dot(mq[ii/NL][0], mk[0]);
//mqka[1] += dot(mq[ii/NL][1], mk[1]);
//mqka[2] += dot(mq[ii/NL][2], mk[2]);
//mqka[3] += dot(mq[ii/NL][3], mk[3]);
mqka[0] += dot((float4) mq[ii/NL][0], (float4) mk[0]);
mqka[1] += dot((float4) mq[ii/NL][1], (float4) mk[1]);
mqka[2] += dot((float4) mq[ii/NL][2], (float4) mk[2]);
mqka[3] += dot((float4) mq[ii/NL][3], (float4) mk[3]);
}
qk_t mqk = mqka[0] + mqka[1] + mqka[2] + mqka[3];
// simdgroup reduce
// [ 0 .. 7] -> [ 0]
// [ 8 .. 15] -> [ 8]
@@ -3512,7 +3523,7 @@ kernel void kernel_flash_attn_ext_vec(
ss[tiisg] = vs;
// O = diag(ms)*O
#pragma unroll
#pragma unroll(D16/NL)
for (short ii = 0; ii < D16; ii += NL) {
lo[ii/NL] *= ms;
}
@@ -3522,13 +3533,12 @@ kernel void kernel_flash_attn_ext_vec(
// O = O + (Q*K^T)*V
{
#pragma unroll
for (short cc = 0; cc < C/4; ++cc) {
device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
const s4x4_t ms(ss[4*cc + ty]);
#pragma unroll
#pragma unroll(D16/NL)
for (short ii = 0; ii < D16; ii += NL) {
const short i = ii + tx;
@@ -3645,7 +3655,7 @@ kernel void kernel_flash_attn_ext_vec(
half, half4, half4x4, \
half4x4
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64>) flash_attn_ext_vec_t;
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 128>) flash_attn_ext_vec_t;
template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 128>;
#if defined(GGML_METAL_USE_BF16)
@@ -6317,8 +6327,8 @@ kernel void kernel_mul_mm(device const uchar * src0,
const uint im = tgpig.z;
// if this block is of 64x32 shape or smaller
short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
short n_rows = (ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
short n_cols = (ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
// a thread shouldn't load data outside of the matrix
short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
@@ -6326,9 +6336,10 @@ kernel void kernel_mul_mm(device const uchar * src0,
simdgroup_T8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 c_res[8];
for (int i = 0; i < 8; i++){
c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
simdgroup_float8x8 mc[8];
for (short i = 0; i < 8; i++){
mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
}
short il = (tiitg % THREAD_PER_ROW);
@@ -6339,7 +6350,7 @@ kernel void kernel_mul_mm(device const uchar * src0,
uint offset0 = (i12/r2)*nb02 + (i13/r3)*nb03;
ushort offset1 = il/nl;
device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*nb01 + offset0) + offset1;
device const float * y = (device const float *)(src1
+ nb13 * i13
+ nb12 * i12
@@ -6353,13 +6364,13 @@ kernel void kernel_mul_mm(device const uchar * src0,
threadgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll(16)
for (int i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+ (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+ (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
}
*(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
*(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2+nl-1)/nl : x;
@@ -6368,27 +6379,27 @@ kernel void kernel_mul_mm(device const uchar * src0,
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup T * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
threadgroup T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
#pragma unroll(4)
for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
#pragma unroll(4)
for (int i = 0; i < 4; i++) {
simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
for (short i = 0; i < 4; i++) {
simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
}
simdgroup_barrier(mem_flags::mem_none);
#pragma unroll(2)
for (int i = 0; i < 2; i++) {
simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
for (short i = 0; i < 2; i++) {
simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
}
lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
lsma += BLOCK_SIZE_M/SG_MAT_ROW * SG_MAT_SIZE;
lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE;
#pragma unroll(8)
for (int i = 0; i < 8; i++){
simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
for (short i = 0; i < 8; i++){
simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
}
}
}
@@ -6396,25 +6407,36 @@ kernel void kernel_mul_mm(device const uchar * src0,
if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg & 1)) \
+ (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
for (short i = 0; i < 8; i++) {
simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
threadgroup float * temp_str = ((threadgroup float *) shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
for (short i = 0; i < 8; i++) {
simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
if (sgitg == 0) {
for (int i = 0; i < n_rows; i++) {
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
*(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
device float * D = dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*ne0 + im*ne1*ne0;
device float4 * D4 = (device float4 *) D;
threadgroup float * C = temp_str + (j*BLOCK_SIZE_M);
threadgroup float4 * C4 = (threadgroup float4 *) C;
int i = 0;
for (; i < n_rows/4; i++) {
*(D4 + i) = *(C4 + i);
}
i *= 4;
for (; i < n_rows; i++) {
*(D + i) = *(C + i);
}
}
}
ggml/src/ggml-sycl.cpp
+4
View File
@@ -4350,6 +4350,10 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
if (op->op == GGML_OP_MUL_MAT) {
a = op->src[0];
b = op->src[1];
if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
// TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
return false;
}
} else {
a = op->src[2];
b = op->src[1];
ggml/src/ggml-sycl/norm.cpp
+3 -3
View File
@@ -8,7 +8,6 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
const int nthreads = item_ct1.get_local_range(2);
const int nwarps = nthreads / WARP_SIZE;
assert(nwarps % WARP_SIZE == 0);
sycl::float2 mean_var = sycl::float2(0.f, 0.f);
for (int col = tid; col < ncols; col += block_size) {
@@ -55,7 +54,6 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
int end = start + group_size;
const int nthreads = item_ct1.get_local_range(2);
const int nwarps = nthreads / WARP_SIZE;
assert(nwarps % WARP_SIZE == 0);
start += item_ct1.get_local_id(2);
int nreduce = nwarps / WARP_SIZE;
@@ -144,7 +142,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
const int tid = item_ct1.get_local_id(2);
const int nthreads = item_ct1.get_local_range(2);
const int nwarps = nthreads / WARP_SIZE;
assert(nwarps % WARP_SIZE == 0);
float tmp = 0.0f; // partial sum for thread in warp
for (int col = tid; col < ncols; col += block_size) {
@@ -202,6 +199,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols,
}
else {
const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
const sycl::range<3> block_dims(1, 1, work_group_size);
/*
DPCT1049:17: The work-group size passed to the SYCL kernel may exceed
@@ -244,6 +242,7 @@ static void group_norm_f32_sycl(const float* x, float* dst,
}
else {
const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
const sycl::range<3> block_dims(1, 1, work_group_size);
/*
DPCT1049:18: The work-group size passed to the SYCL kernel may exceed
@@ -290,6 +289,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
}
else {
const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
const sycl::range<3> block_dims(1, 1, work_group_size);
/*
DPCT1049:19: The work-group size passed to the SYCL kernel may exceed
ggml/src/ggml-sycl/outprod.cpp
+2 -1
View File
@@ -1,4 +1,5 @@
#include <sycl/sycl.hpp>
#include <oneapi/mkl.hpp>
#include "outprod.hpp"
@@ -39,7 +40,7 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, const ggml_tensor* sr
try {
// Perform matrix multiplication using oneMKL GEMM
oneapi::mkl::blas::gemm(*stream,
oneapi::mkl::blas::column_major::gemm(*stream,
oneapi::mkl::transpose::nontrans, src1_op,
ne0, ne1, ne01,
alpha,
ggml/src/ggml-vulkan-shaders.cpp
+129821 -128921
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-vulkan-shaders.hpp
+582 -573
View File
File diff suppressed because it is too large Load Diff
ggml/src/ggml-vulkan.cpp
+220 -478
View File
@@ -106,6 +106,15 @@ struct vk_matmul_pipeline_struct {
typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
struct vk_matmul_pipeline2 {
vk_matmul_pipeline2() {
f16acc = std::make_shared<vk_matmul_pipeline_struct>();
f32acc = std::make_shared<vk_matmul_pipeline_struct>();
}
vk_matmul_pipeline f32acc;
vk_matmul_pipeline f16acc;
};
struct vk_device_struct;
typedef std::shared_ptr<vk_device_struct> vk_device;
typedef std::weak_ptr<vk_device_struct> vk_device_ref;
@@ -161,11 +170,11 @@ struct vk_device_struct {
vk_matmul_pipeline pipeline_matmul_f32;
vk_matmul_pipeline pipeline_matmul_f32_f16;
vk_matmul_pipeline pipeline_matmul_f16;
vk_matmul_pipeline pipeline_matmul_f16_f32;
vk_matmul_pipeline2 pipeline_matmul_f16;
vk_matmul_pipeline2 pipeline_matmul_f16_f32;
vk_pipeline pipeline_matmul_split_k_reduce;
vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
vk_matmul_pipeline pipeline_matmul_id_f32;
vk_matmul_pipeline pipeline_matmul_id_f16;
@@ -183,9 +192,10 @@ struct vk_device_struct {
vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_acc_f32;
vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
vk_pipeline pipeline_mul_f32;
vk_pipeline pipeline_div_f32;
vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat;
vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat;
vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat;
vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat;
vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
vk_pipeline pipeline_upscale_f32;
vk_pipeline pipeline_scale_f32;
@@ -196,6 +206,7 @@ struct vk_device_struct {
vk_pipeline pipeline_pad_f32;
vk_pipeline pipeline_repeat_f32;
vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
vk_pipeline pipeline_norm_f32;
vk_pipeline pipeline_group_norm_f32;
vk_pipeline pipeline_rms_norm_f32;
@@ -722,6 +733,12 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
std::lock_guard<std::mutex> guard(compile_count_mutex);
assert(compile_count > 0);
compile_count--;
// "Progress bar" for shader compiles
static uint32_t total_compile_count = 0;
if ((total_compile_count++ % 10) == 0) {
std::cerr << ".";
}
}
compile_count_cond.notify_all();
}
@@ -1200,38 +1217,31 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
static void ggml_vk_load_shaders(vk_device& device) {
VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
std::cerr << "ggml_vulkan: Compiling shaders";
// mulmat
std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
std::initializer_list<uint32_t> warptile_m = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
std::initializer_list<uint32_t> warptile_s = { std::max(device->subgroup_size, 16u), 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
std::vector<uint32_t> l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq;
std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms;
uint32_t l_align, m_align, s_align;
std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
std::initializer_list<uint32_t> warptile_mmq_m = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
std::initializer_list<uint32_t> warptile_mmq_s = { std::max(device->subgroup_size, 16u), 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
s_warptile = { std::max(device->subgroup_size, 16u), 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
std::array<uint32_t, 3> m_wg_denoms = { 64, 64, 1 };
std::array<uint32_t, 3> s_wg_denoms = { 32, 32, 1 };
l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
s_warptile_mmq = { std::max(device->subgroup_size, 16u), 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
uint32_t l_align = 128;
uint32_t m_align = 64;
uint32_t s_align = 32;
l_wg_denoms = {128, 128, 1 };
m_wg_denoms = { 64, 64, 1 };
s_wg_denoms = { 32, 32, 1 };
l_align = 128;
m_align = 64;
s_align = 32;
device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
@@ -1249,7 +1259,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
std::vector<std::future<void>> compiles;
auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align) {
{
// wait until fewer than N compiles are in progress
uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1263,411 +1273,95 @@ static void ggml_vk_load_shaders(vk_device& device) {
};
if (device->fp16) {
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
// Create 6 variants, {s,m,l}x{unaligned,aligned}
#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
#undef CREATE_MM
} else {
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
// Create 6 variants, {s,m,l}x{unaligned,aligned}
#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
#undef CREATE_MM
}
// mul mat vec
@@ -1759,13 +1453,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16_norepeat, "add_f16_f32_f16_norepeat", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
ggml_vk_create_pipeline(device, device->pipeline_div_f32_norepeat, "div_f32_norepeat", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -1817,6 +1519,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
for (auto &c : compiles) {
c.wait();
}
std::cerr << "Done!" << std::endl;
}
static vk_device ggml_vk_get_device(size_t idx) {
@@ -2299,10 +2002,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
return ctx->device->pipeline_matmul_f32_f16;
}
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
return ctx->device->pipeline_matmul_f16_f32;
return ctx->device->pipeline_matmul_f16_f32.f32acc;
}
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
return ctx->device->pipeline_matmul_f16;
return ctx->device->pipeline_matmul_f16.f32acc;
}
if (src1_type != GGML_TYPE_F32) {
@@ -2326,7 +2029,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
return nullptr;
}
return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
return ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
}
static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -3061,18 +2764,34 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) {
if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
return ctx->device->pipeline_cpy_f32_f32;
static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to) {
// Choose "contiguous copy" shader if src/dst are contiguous
bool contig = ggml_is_contiguous(src) && (!dst || ggml_is_contiguous(dst));
if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
if (contig) {
return ctx->device->pipeline_contig_cpy_f32_f32;
} else {
return ctx->device->pipeline_cpy_f32_f32;
}
}
if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
return ctx->device->pipeline_cpy_f32_f16;
if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
if (contig) {
return ctx->device->pipeline_contig_cpy_f32_f16;
} else {
return ctx->device->pipeline_cpy_f32_f16;
}
}
if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
return ctx->device->pipeline_cpy_f16_f16;
if (src->type == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
if (contig) {
return ctx->device->pipeline_contig_cpy_f16_f16;
} else {
return ctx->device->pipeline_cpy_f16_f16;
}
}
std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
GGML_ABORT("fatal error");
}
@@ -3082,6 +2801,15 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
const int tensor_type_size = ggml_type_size(tensor->type);
const uint32_t ne = ggml_nelements(tensor);
std::array<uint32_t, 3> elements;
if (ne > 262144) {
elements = { 512, 512, CEIL_DIV(ne, 262144) };
} else if (ne > 512) {
elements = { 512, CEIL_DIV(ne, 512), 1 };
} else {
elements = { ne, 1, 1 };
}
const vk_op_unary_push_constants pc = {
(uint32_t)ne,
@@ -3091,7 +2819,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
0.0f, 0.0f,
};
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 });
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
}
static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -3147,7 +2875,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
if (mmp == nullptr) {
if (qx_needs_dequant) {
// Fall back to dequant + f16 mulmat
mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16);
}
@@ -3176,12 +2904,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
vk_pipeline to_fp16_vk_1 = nullptr;
if (x_non_contig) {
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
} else {
to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
}
if (y_non_contig) {
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
@@ -3361,10 +3089,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
vk_pipeline to_fp16_vk_0 = nullptr;
vk_pipeline to_fp16_vk_1 = nullptr;
if (x_non_contig) {
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type);
}
if (y_non_contig) {
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type);
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
@@ -3630,9 +3358,19 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")");
if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) {
if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1 &&
// detect 0213 permutation, and batch size of 1
src0->nb[0] <= src0->nb[2] &&
src0->nb[2] <= src0->nb[1] &&
src0->nb[1] <= src0->nb[3] &&
src1->nb[0] <= src1->nb[2] &&
src1->nb[2] <= src1->nb[1] &&
src1->nb[1] <= src1->nb[3] &&
src0->ne[3] == 1 &&
src1->ne[3] == 1) {
ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
} else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1) {
} else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
!ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
} else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -3708,7 +3446,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
if (mmp == nullptr) {
if (qx_needs_dequant) {
GGML_ABORT("fatal error");
}
@@ -3735,12 +3473,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
vk_pipeline to_fp16_vk_1 = nullptr;
if (x_non_contig) {
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
} else {
to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
}
if (y_non_contig) {
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
@@ -3928,10 +3666,10 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
vk_pipeline to_fp16_vk_0 = nullptr;
vk_pipeline to_fp16_vk_1 = nullptr;
if (x_non_contig) {
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type);
}
if (y_non_contig) {
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type);
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
@@ -4068,20 +3806,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return nullptr;
case GGML_OP_ADD:
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_add_f32;
return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f32_norepeat : ctx->device->pipeline_add_f32;
}
if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
return ctx->device->pipeline_add_f16_f32_f16;
return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16;
}
return nullptr;
case GGML_OP_MUL:
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_mul_f32;
return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32;
}
return nullptr;
case GGML_OP_DIV:
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_div_f32;
return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_f32_norepeat : ctx->device->pipeline_div_f32;
}
return nullptr;
case GGML_OP_CONCAT:
@@ -4138,7 +3876,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
case GGML_OP_CPY:
case GGML_OP_CONT:
case GGML_OP_DUP:
return ggml_vk_get_cpy_pipeline(ctx, src0->type, dst->type);
return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
case GGML_OP_NORM:
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
return ctx->device->pipeline_norm_f32;
@@ -4271,7 +4009,6 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
case GGML_OP_DIV:
case GGML_OP_CONCAT:
case GGML_OP_UPSCALE:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SIN:
case GGML_OP_COS:
@@ -4470,7 +4207,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
const uint32_t OH = is_2D ? dst->ne[2] : 1;
const uint32_t OW = dst->ne[1];
const uint32_t batch = src1->ne[3];
const uint32_t batch = src1->ne[is_2D ? 3 : 2];
elements = { OW * KW * KH, OH, batch * IC };
} break;
@@ -4915,7 +4652,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
const uint32_t OW = dst->ne[1];
const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
const uint32_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
const uint32_t batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
const uint32_t pelements = OW * KW * KH;
@@ -5022,10 +4759,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->a_s;
shname = "F32_F16_ALIGNED_S";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->a_s;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_s;
shname = "F16_F32_ALIGNED_S";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->a_s;
p = ctx->device->pipeline_matmul_f16.f32acc->a_s;
shname = "F16_ALIGNED_S";
} else {
GGML_ABORT("fatal error");
@@ -5038,10 +4775,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->a_m;
shname = "F32_F16_ALIGNED_M";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->a_m;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_m;
shname = "F16_F32_ALIGNED_M";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->a_m;
p = ctx->device->pipeline_matmul_f16.f32acc->a_m;
shname = "F16_ALIGNED_M";
} else {
GGML_ABORT("fatal error");
@@ -5054,10 +4791,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->a_l;
shname = "F32_F16_ALIGNED_L";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->a_l;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_l;
shname = "F16_F32_ALIGNED_L";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->a_l;
p = ctx->device->pipeline_matmul_f16.f32acc->a_l;
shname = "F16_ALIGNED_L";
} else {
GGML_ABORT("fatal error");
@@ -5077,10 +4814,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->s;
shname = "F32_F16_S";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->s;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->s;
shname = "F16_F32_S";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->s;
p = ctx->device->pipeline_matmul_f16.f32acc->s;
shname = "F16_S";
}
} else if (shader_size == 1) {
@@ -5091,10 +4828,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->m;
shname = "F32_F16_M";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->m;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->m;
shname = "F16_F32_M";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->m;
p = ctx->device->pipeline_matmul_f16.f32acc->m;
shname = "F16_M";
}
} else if (shader_size == 2) {
@@ -5105,10 +4842,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
p = ctx->device->pipeline_matmul_f32_f16->l;
shname = "F32_F16_L";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16_f32->l;
p = ctx->device->pipeline_matmul_f16_f32.f32acc->l;
shname = "F16_F32_L";
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
p = ctx->device->pipeline_matmul_f16->l;
p = ctx->device->pipeline_matmul_f16.f32acc->l;
shname = "F16_L";
}
}
@@ -5437,13 +5174,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
vk_pipeline p;
std::string shname;
if (shader_size == 0) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
} else if (shader_size == 1) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
} else if (shader_size == 2) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
} else {
GGML_ASSERT(0);
@@ -5453,13 +5190,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
if (k != kpad) {
if (shader_size == 0) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
shname = std::string(ggml_type_name(quant)) + "_S";
} else if (shader_size == 1) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
shname = std::string(ggml_type_name(quant)) + "_M";
} else if (shader_size == 2) {
p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l;
p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
shname = std::string(ggml_type_name(quant)) + "_L";
} else {
GGML_ASSERT(0);
@@ -6804,6 +6541,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
if (a->ne[3] != b->ne[3]) {
return false;
}
if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) ||
!(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) {
return false;
}
return true;
} break;
case GGML_OP_GET_ROWS:
ggml/src/llamafile/sgemm.cpp
+608
View File
@@ -106,6 +106,10 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__MMA__)
typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED FUSED MULTIPLY ADD
@@ -1026,6 +1030,600 @@ class tinyBLAS_Q0_AVX {
};
#endif // __AVX__
//PPC Implementation
#if defined(__MMA__)
#define SAVE_ACC(ACC, ii, jj) \
__builtin_mma_disassemble_acc(vec_C, ACC); \
for (int I = 0; I < 4; I++) { \
for (int J = 0; J < 4; J++) { \
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
} \
} \
template <typename TA, typename TB, typename TC>
class tinyBLAS_PPC {
public:
tinyBLAS_PPC(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) {
int64_t i, j;
float *aoffset = NULL, *boffset = NULL;
float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
aoffset = const_cast<float*>(a);
boffset = vec;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5);
C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6);
C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7);
C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
__builtin_vsx_disassemble_pair(c5, &C5);
__builtin_vsx_disassemble_pair(c6, &C6);
__builtin_vsx_disassemble_pair(c7, &C7);
__builtin_vsx_disassemble_pair(c8, &C8);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergeh(c5[0], c6[0]);
t4 = vec_mergeh(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_mergel(c5[0], c6[0]);
t4 = vec_mergel(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergeh(c5[1], c6[1]);
t4 = vec_mergeh(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+32);
vec_xst(t6, 0, boffset+36);
vec_xst(t7, 0, boffset+40);
vec_xst(t8, 0, boffset+44);
t1 = vec_mergel(c1[1], c2[1]);
t2 = vec_mergel(c3[1], c4[1]);
t3 = vec_mergel(c5[1], c6[1]);
t4 = vec_mergel(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+48);
vec_xst(t6, 0, boffset+52);
vec_xst(t7, 0, boffset+56);
vec_xst(t8, 0, boffset+60);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 64;
i--;
} while(i > 0);
}
if (cols & 4) {
vector float c1, c2, c3, c4, c5, c6, c7, c8;
vector float t1, t2, t3, t4, t5, t6, t7, t8;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
c5 = vec_xl(0, aoffset5);
c6 = vec_xl(0, aoffset6);
c7 = vec_xl(0, aoffset7);
c8 = vec_xl(0, aoffset8);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_mergeh(c5, c6);
t4 = vec_mergeh(c7, c8);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_mergel(c5, c6);
t4 = vec_mergel(c7, c8);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
}
j--;
} while(j > 0);
}
if (rows & 4) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4;
vector float c1[2], c2[2], c3[2], c4[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergel(c1[0], c2[0]);
t4 = vec_mergel(c3[0], c4[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergel(c1[1], c2[1]);
t4 = vec_mergel(c3[1], c4[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 32;
i--;
} while(i > 0);
}
if (cols & 4) {
vector float c1, c2, c3, c4;
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
if (rows & 3) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
if (cols & 4) {
vector float c1, c2, c3, c4 = {0};
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
}
void KERNEL_4x4(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[4], vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
for (int l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
}
void KERNEL_4x8(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[8], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[1], (vec_t)vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], (vec_t)vec_B[4]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[2], (vec_t)vec_B[5]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
}
void KERNEL_8x4(int64_t ii, int64_t jj) {
vec_t vec_A[8], vec_B[4], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[3], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[4], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[5], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii+4, jj);
}
void KERNEL_8x8(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[16], vec_C[4];
acc_t acc_0, acc_1, acc_2, acc_3;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
for (int l = 0; l < k; l+=8) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
for(int x = 0; x < 16; x+=2) {
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
__builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
}
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
SAVE_ACC(&acc_2, ii+4, jj);
SAVE_ACC(&acc_3, ii+4, jj+4);
}
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
int m_rem = MIN(m - m0, 16);
int n_rem = MIN(n - n0, 16);
if (m_rem >= 16 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if(m_rem >= 8 && n_rem >= 16) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 8) {
mc = 4;
nc = 8;
gemm<4,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 4) {
mc = 8;
nc = 4;
gemm<8,4>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 4) {
mc = 4;
nc = 4;
gemm<4,4>(m0, m, n0, n);
} else if ((m_rem < 4) && (n_rem > 4)) {
nc = 4;
switch(m_rem) {
case 1:
mc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
mc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
mc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else if ((m_rem > 4) && (n_rem < 4)) {
mc = 4;
switch(n_rem) {
case 1:
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else {
switch((m_rem << 4) | n_rem) {
case 0x43:
mc = 4;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x42:
mc = 4;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x41:
mc = 4;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x34:
mc = 3;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x33:
mc = 3;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x32:
mc = 3;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x31:
mc = 3;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x24:
mc = 2;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x23:
mc = 2;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x22:
mc = 2;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x21:
mc = 2;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x14:
mc = 1;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x13:
mc = 1;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x12:
mc = 1;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x11:
mc = 1;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
vec_t vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
vec_t vec_A[4], vec_B[4];
for (int l=0; l<k; l+=4) {
if (RN >= 4 && RM == 1) {
float* a = const_cast<float*>(A+(ii)*lda+l);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
vec_A[0] = (vec_t)vec_xl(0,a);
vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
} else {
READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
}
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
__builtin_mma_disassemble_acc(vec_C, &acc_0);
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
}
}
}
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (RM == 4 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_4x4;
} else if (RM == 4 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_4x8;
} else if (RM == 8 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_8x4;
} else if (RM == 8 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_8x8;
}
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
(this->*kernel)(ii, jj);
}
}
const TA *const A;
const TB *const B;
TC *C;
TA *At;
TB *Bt;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
#endif
} // namespace
/**
@@ -1114,6 +1712,16 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
ith, nth};
tb.matmul(m, n);
return true;
#elif defined(__MMA__)
if (k % 8)
return false;
tinyBLAS_PPC<float, float, float> tb{
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
ggml/src/vulkan-shaders/acc.comp
+7 -2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_binary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = gl_GlobalInvocationID.x;
if (idx >= p.ne) {
@@ -15,10 +17,13 @@ void main() {
const uint oy = (src1_i - (oz * p.nb02)) / p.nb01;
const uint ox = src1_i % p.nb01;
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
} else {
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]));
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
}
}
ggml/src/vulkan-shaders/add.comp
+20 -5
View File
@@ -1,14 +1,29 @@
#version 450
#extension GL_EXT_shader_16bit_storage : require
#include "types.comp"
#include "generic_binary_head.comp"
const uint num_threads = 256;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
uint idx = get_idx();
if (idx >= p.ne) {
return;
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 2;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
idx += num_threads;
}
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
}
ggml/src/vulkan-shaders/clamp.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
ggml/src/vulkan-shaders/concat.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_binary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
const int dim = p.param3;
ggml/src/vulkan-shaders/contig_copy.comp
+42
View File
@@ -0,0 +1,42 @@
#version 450
#include "types.comp"
#include "generic_unary_head.comp"
#extension GL_EXT_control_flow_attributes : require
const uint num_threads = 128;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
uint idx = get_idx();
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 4;
// fast path for when all four iterations are in-bounds
if (idx + (num_iter-1)*num_threads < p.ne) {
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
#ifndef OPTIMIZATION_ERROR_WORKAROUND
data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
#else
data_d[p.d_offset + idx] = data_a[idx];
#endif
idx += num_threads;
}
} else {
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
#ifndef OPTIMIZATION_ERROR_WORKAROUND
data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
#else
data_d[p.d_offset + idx] = data_a[idx];
#endif
idx += num_threads;
}
}
}
ggml/src/vulkan-shaders/copy.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
ggml/src/vulkan-shaders/cos.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
ggml/src/vulkan-shaders/div.comp
+18 -5
View File
@@ -3,12 +3,25 @@
#include "types.comp"
#include "generic_binary_head.comp"
const uint num_threads = 256;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
uint idx = get_idx();
if (idx >= p.ne) {
return;
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 2;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
idx += num_threads;
}
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
}
ggml/src/vulkan-shaders/generic_binary_head.comp
+32 -24
View File
@@ -1,4 +1,5 @@
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_control_flow_attributes : require
layout (push_constant) uniform parameter
{
@@ -10,43 +11,50 @@ layout (push_constant) uniform parameter
float param1; float param2; int param3;
} p;
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
// true if src0/src1 are the same shape and the indices can be reused without additional modulus
layout(constant_id = 0) const bool norepeat = false;
uint get_idx() {
return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
}
uint src0_idx(uint idx) {
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
uint fastmod(uint a, uint b) {
if ((b & (b-1)) == 0) {
return a & (b-1);
}
return a % b;
}
uint fastdiv(uint a, uint b) {
return (a < b) ? 0 : (a / b);
}
void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
const uint i02_offset = i02*p.ne01*p.ne00;
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
i01 = (idx - i03_offset - i02_offset) / p.ne00;
i00 = idx - i03_offset - i02_offset - i01*p.ne00;
}
uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
}
uint src1_idx(uint idx) {
const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
const uint i02_offset = i02*p.ne01*p.ne00;
const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
uint src1_idx(uint i00, uint i01, uint i02, uint i03) {
if (norepeat) {
return i03*p.nb13 + i02*p.nb12 + i01*p.nb11 + i00*p.nb10;
} else {
return fastmod(i03, p.ne13)*p.nb13 + fastmod(i02, p.ne12)*p.nb12 + fastmod(i01, p.ne11)*p.nb11 + fastmod(i00, p.ne10)*p.nb10;
}
}
uint dst_idx(uint idx) {
const uint i23 = idx / (p.ne22*p.ne21*p.ne20);
const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20;
const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20);
const uint i22_offset = i22*p.ne21*p.ne20;
const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
return i03*p.nb23 + i02*p.nb22 + i01*p.nb21 + i00*p.nb20;
}
ggml/src/vulkan-shaders/generic_unary_head.comp
+1 -2
View File
@@ -1,4 +1,5 @@
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_control_flow_attributes : require
layout (push_constant) uniform parameter
{
@@ -9,8 +10,6 @@ layout (push_constant) uniform parameter
float param1; float param2;
} p;
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
ggml/src/vulkan-shaders/get_rows.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_binary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint i00 = gl_GlobalInvocationID.x;
const uint i10 = gl_GlobalInvocationID.y;
ggml/src/vulkan-shaders/get_rows_quant.comp
+2
View File
@@ -4,6 +4,8 @@
#include "generic_binary_head.comp"
#include "dequant_funcs.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint i00 = (gl_GlobalInvocationID.x)*2;
const uint i10 = gl_GlobalInvocationID.y;
ggml/src/vulkan-shaders/mul.comp
+18 -5
View File
@@ -3,12 +3,25 @@
#include "types.comp"
#include "generic_binary_head.comp"
const uint num_threads = 256;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
uint idx = get_idx();
if (idx >= p.ne) {
return;
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 2;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
idx += num_threads;
}
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
}
ggml/src/vulkan-shaders/pad.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
ggml/src/vulkan-shaders/repeat.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
uint src0_idx_mod(uint idx) {
const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
ggml/src/vulkan-shaders/scale.comp
+15 -5
View File
@@ -3,12 +3,22 @@
#include "types.comp"
#include "generic_unary_head.comp"
const uint num_threads = 128;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
uint idx = get_idx();
if (idx >= p.ne) {
return;
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 4;
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
data_d[p.d_offset + idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(p.param1));
idx += num_threads;
}
data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1));
}
ggml/src/vulkan-shaders/sin.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
ggml/src/vulkan-shaders/square.comp
+2
View File
@@ -3,6 +3,8 @@
#include "types.comp"
#include "generic_unary_head.comp"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+99 -184
View File
@@ -16,6 +16,7 @@
#include <cstdio>
#include <cstring>
#include <cstdlib>
#include <cassert>
#include <sys/stat.h>
#include <sys/types.h>
#include <algorithm>
@@ -93,11 +94,11 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
std::array<char, 128> buffer;
DWORD bytes_read;
while (ReadFile(stdout_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
while (ReadFile(stdout_read, buffer.data(), (DWORD)buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
stdout_str.append(buffer.data(), bytes_read);
}
while (ReadFile(stderr_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
while (ReadFile(stderr_read, buffer.data(), (DWORD)buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
stderr_str.append(buffer.data(), bytes_read);
}
@@ -191,7 +192,12 @@ std::string basename(const std::string &path) {
return path.substr(path.find_last_of("/\\") + 1);
}
void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
// variables to track number of compiles in progress
static uint32_t compile_count = 0;
static std::mutex compile_count_mutex;
static std::condition_variable compile_count_cond;
void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
std::string name = _name + (fp16 ? "" : "_fp32");
std::string out_fname = join_paths(output_dir, name + ".spv");
std::string in_path = join_paths(input_dir, in_fname);
@@ -234,6 +240,12 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
} catch (const std::exception& e) {
std::cerr << "Error executing command for " << name << ": " << e.what() << std::endl;
}
{
std::lock_guard<std::mutex> guard(compile_count_mutex);
assert(compile_count > 0);
compile_count--;
}
compile_count_cond.notify_all();
}
std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b) {
@@ -242,7 +254,22 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
return result;
}
void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id) {
static std::vector<std::future<void>> compiles;
void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
{
// wait until fewer than N compiles are in progress.
// 16 is an arbitrary limit, the goal is to avoid "failed to create pipe" errors.
uint32_t N = 16;
std::unique_lock<std::mutex> guard(compile_count_mutex);
while (compile_count >= N) {
compile_count_cond.wait(guard);
}
compile_count++;
}
compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16));
}
void matmul_shaders(bool fp16, bool matmul_id) {
std::string load_vec = fp16 ? "8" : "4";
std::string aligned_b_type_f32 = fp16 ? "mat2x4" : "vec4";
std::string aligned_b_type_f16 = fp16 ? "f16mat2x4" : "f16vec4";
@@ -260,19 +287,11 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu
}
// Shaders with f16 B_TYPE
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_f32_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_f32_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
}));
string_to_spv(shader_name + "_f32_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
string_to_spv(shader_name + "_f32_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
}));
string_to_spv(shader_name + "_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
string_to_spv(shader_name + "_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
for (const auto& tname : type_names) {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
@@ -280,22 +299,18 @@ void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmu
std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
// For aligned matmul loads
std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
}));
string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
}
}
void process_shaders(std::vector<std::future<void>>& tasks) {
void process_shaders() {
std::cout << "ggml_vulkan: Generating and compiling shaders to SPIR-V" << std::endl;
std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
for (const auto& fp16 : {false, true}) {
matmul_shaders(tasks, fp16, false);
matmul_shaders(tasks, fp16, true);
matmul_shaders(fp16, false);
matmul_shaders(fp16, true);
}
for (const auto& tname : type_names) {
@@ -303,201 +318,106 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
}));
string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
// Dequant shaders
if (tname != "f16") {
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
}));
string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
}
if (!string_ends_with(tname, "_k")) {
shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
if (tname == "f16") {
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
}));
string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
} else {
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}});
}));
string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}});
}
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
}));
string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
}
}
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
// Norms
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
}));
string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
}));
string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}));
string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
}));
string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
}));
string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
}));
string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
}));
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
}));
string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
tasks.push_back(std::async(std::launch::async, [] {
string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
}));
string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
}));
string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
tasks.push_back(std::async(std::launch::async, [=] {
string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
}));
string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
for (auto &c : compiles) {
c.wait();
}
}
void write_output_files() {
@@ -592,12 +512,7 @@ int main(int argc, char** argv) {
}
}
std::vector<std::future<void>> tasks;
process_shaders(tasks);
for (auto& task : tasks) {
task.get();
}
process_shaders();
write_output_files();