Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 01:04:55 +02:00
Code Issues Packages Projects Releases Wiki Activity

add faster RDNA int->float conversion

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 10:45:00 +02:00
parent 6ca9b4548a
commit 44d1dc488d
1 changed files with 23 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
+23 -7
View File
@@ -96,6 +96,9 @@ shared float buf_b_d[BN * BK_STEP];
#define NUM_WARPS (BLOCK_SIZE / WARP)
const uint CM_ELEMS = (TM * TN) / WARP;
#define ACC_BIAS_BITS 0x4B400000
#define ACC_BIAS_F 12582912.0f
const bool USE_MAGIC_BIAS = WARP != 32;
shared int32_t cm_layout_probe[TM * TN];
@@ -343,7 +346,8 @@ void main() {
// Compute from shmem
[[unroll]] for (uint ks = 0; ks < BK_STEP; ks++) {
[[unroll]] for (uint idx = 0; idx < cms_per_row * cms_per_col; idx++) {
cm_result[idx] = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0);
cm_result[idx] = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(
USE_MAGIC_BIAS ? ACC_BIAS_BITS : 0);
}
const uint K_SUB = BK / TK;
@@ -370,16 +374,20 @@ void main() {
}
// Pre-load scales into registers
ACC_TYPE scale_a[cms_per_row * CM_ELEMS];
ACC_TYPE scale_b[cms_per_col * CM_ELEMS];
float scale_a[cms_per_row * CM_ELEMS];
float nbias_a[cms_per_row * CM_ELEMS];
float scale_b[cms_per_col * CM_ELEMS];
[[unroll]] for (uint r = 0; r < cms_per_row; r++) {
[[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
scale_a[r * CM_ELEMS + e] = ACC_TYPE(buf_a_d[ks * BM + warp_r * WM + r * TM + elem_row[e]]);
scale_a[r * CM_ELEMS + e] = buf_a_d[ks * BM + warp_r * WM + r * TM + elem_row[e]];
if (USE_MAGIC_BIAS) {
nbias_a[r * CM_ELEMS + e] = -ACC_BIAS_F * scale_a[r * CM_ELEMS + e];
}
}
}
[[unroll]] for (uint c = 0; c < cms_per_col; c++) {
[[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
scale_b[c * CM_ELEMS + e] = ACC_TYPE(buf_b_d[ks * BN + warp_c * WN + c * TN + elem_col[e]]);
scale_b[c * CM_ELEMS + e] = buf_b_d[ks * BN + warp_c * WN + c * TN + elem_col[e]];
}
}
@@ -388,8 +396,16 @@ void main() {
[[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
const uint tile_idx = cm_col * cms_per_row + cm_row;
[[unroll]] for (uint e = 0; e < CM_ELEMS; e++) {
sums[tile_idx * CM_ELEMS + e] += ACC_TYPE(cm_result[tile_idx][e])
* scale_a[cm_row * CM_ELEMS + e] * scale_b[cm_col * CM_ELEMS + e];
if (USE_MAGIC_BIAS) {
const float t = fma(intBitsToFloat(int(cm_result[tile_idx][e])),
scale_a[cm_row * CM_ELEMS + e],
nbias_a[cm_row * CM_ELEMS + e]);
sums[tile_idx * CM_ELEMS + e] = ACC_TYPE(fma(t, scale_b[cm_col * CM_ELEMS + e],
float(sums[tile_idx * CM_ELEMS + e])));
} else {
sums[tile_idx * CM_ELEMS + e] += ACC_TYPE(float(cm_result[tile_idx][e])
* scale_a[cm_row * CM_ELEMS + e] * scale_b[cm_col * CM_ELEMS + e]);
}
}
}
}