From 06fe4c7f20f4668cb19913d24e56822ef3a9001e Mon Sep 17 00:00:00 2001 From: Disty0 Date: Sun, 29 Mar 2026 00:44:58 +0300 Subject: [PATCH] SDNQ cleanup triton_mm --- modules/sdnq/triton_mm.py | 68 ++++++--------------------------------- 1 file changed, 9 insertions(+), 59 deletions(-) diff --git a/modules/sdnq/triton_mm.py b/modules/sdnq/triton_mm.py index c29b6234d..a633a395b 100644 --- a/modules/sdnq/triton_mm.py +++ b/modules/sdnq/triton_mm.py @@ -22,14 +22,15 @@ matmul_configs = [ ] -@triton.autotune(configs=matmul_configs, key=["M", "N", "K", "stride_bk"]) +@triton.autotune(configs=matmul_configs, key=["M", "N", "K", "stride_bk", "ACCUMULATOR_DTYPE"]) @triton.jit -def int_mm_kernel( +def triton_mm_kernel( a_ptr, b_ptr, c_ptr, M: int, N: int, K: int, stride_am: int, stride_ak: int, stride_bk: int, stride_bn: int, stride_cm: int, stride_cn: int, + ACCUMULATOR_DTYPE: tl.constexpr, BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr, @@ -60,11 +61,11 @@ def int_mm_kernel( a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) - accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.int32) + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=ACCUMULATOR_DTYPE) for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) - accumulator = tl.dot(a, b, accumulator, out_dtype=tl.int32) + accumulator = tl.dot(a, b, accumulator, out_dtype=ACCUMULATOR_DTYPE) a_ptrs += BLOCK_SIZE_K * stride_ak b_ptrs += BLOCK_SIZE_K * stride_bk @@ -83,69 +84,17 @@ def int_mm(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor: c = torch.empty((M, N), device=a.device, dtype=torch.int32) def grid(META): return (triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), ) - int_mm_kernel[grid]( + triton_mm_kernel[grid]( a, b, c, M, N, K, a.stride(0), a.stride(1), b.stride(0), b.stride(1), c.stride(0), c.stride(1), + tl.int32, ) return c -@triton.autotune(configs=matmul_configs, key=["M", "N", "K", "stride_bk"]) -@triton.jit -def fp_mm_kernel( - a_ptr, b_ptr, c_ptr, - M: int, N: int, K: int, - stride_am: int, stride_ak: int, - stride_bk: int, stride_bn: int, - stride_cm: int, stride_cn: int, - BLOCK_SIZE_M: tl.constexpr, - BLOCK_SIZE_N: tl.constexpr, - BLOCK_SIZE_K: tl.constexpr, - GROUP_SIZE_M: tl.constexpr, -): - pid = tl.program_id(axis=0) - num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) - num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) - num_pid_in_group = GROUP_SIZE_M * num_pid_n - group_id = pid // num_pid_in_group - first_pid_m = group_id * GROUP_SIZE_M - group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) - pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) - pid_n = (pid % num_pid_in_group) // group_size_m - - tl.assume(pid_m >= 0) - tl.assume(pid_n >= 0) - tl.assume(stride_am > 0) - tl.assume(stride_ak > 0) - tl.assume(stride_bn > 0) - tl.assume(stride_bk > 0) - tl.assume(stride_cm > 0) - tl.assume(stride_cn > 0) - - offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M - offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N - offs_k = tl.arange(0, BLOCK_SIZE_K) - a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) - b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) - - accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) - for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): - a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) - b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) - accumulator = tl.dot(a, b, accumulator, out_dtype=tl.float32) - a_ptrs += BLOCK_SIZE_K * stride_ak - b_ptrs += BLOCK_SIZE_K * stride_bk - - offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) - offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) - c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :] - c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N) - tl.store(c_ptrs, accumulator, mask=c_mask) - - def fp_mm(a: torch.FloatTensor, b: torch.FloatTensor) -> torch.FloatTensor: assert a.shape[1] == b.shape[0], "Incompatible dimensions" assert a.is_contiguous(), "Matrix A must be contiguous" @@ -154,11 +103,12 @@ def fp_mm(a: torch.FloatTensor, b: torch.FloatTensor) -> torch.FloatTensor: c = torch.empty((M, N), device=a.device, dtype=torch.float32) def grid(META): return (triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), ) - fp_mm_kernel[grid]( + triton_mm_kernel[grid]( a, b, c, M, N, K, a.stride(0), a.stride(1), b.stride(0), b.stride(1), c.stride(0), c.stride(1), + tl.float32, ) return c