youngkingdom/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v4.2 tag release. (#2638)

Browse Source
This commit is contained in:
Junkai-Wu
2025-09-16 00:21:53 +08:00
committed by GitHub
parent 56f0718a97
commit 6a35b4d22f
161 changed files with 14056 additions and 3793 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/cutlass_library/emit_kernel_listing.py
View File

@ -348,11 +348,15 @@ def emit_gemm_kernel_testlist(manifest, curr_build_dir, arch, mode
'gemm.*ue8m0xf6_ue8m0xf6_f32_f16_ue8m0xe3m2',
]
block_scaled_tile_k = ['x128_', 'x256_']
sm103_block_scaled_data_type = [
'gemm.*ue8m0xf4_ue8m0xf4_f32_f16_e5m2',
'gemm.*ue8m0xf4_ue8m0xf4_f32_f16_ue8m0xe2m1',
]
sm103_block_scaled_tile_k = ['x768_']
block_scaled_cluster_size = [
'4x4x1', '2x1x1',
'0x0x1' # dynamic cluster
@ -360,11 +364,12 @@ def emit_gemm_kernel_testlist(manifest, curr_build_dir, arch, mode
block_scaled_layouts = ['tnt']
# regex list must be in kernel procedural name order
block_scaled_filter_regex_1sm = "cutlass3x_sm100_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [block_scaled_data_type, block_scaled_cluster_size, block_scaled_layouts]]) + ").*1sm.*"
block_scaled_filter_regex_2sm = "cutlass3x_sm100_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [block_scaled_data_type, block_scaled_cluster_size, block_scaled_layouts]]) + ").*2sm.*"
block_scaled_filter_regex_1sm = "cutlass3x_sm100_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [block_scaled_data_type, block_scaled_tile_k, block_scaled_cluster_size, block_scaled_layouts]]) + ").*1sm.*"
block_scaled_filter_regex_2sm = "cutlass3x_sm100_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [block_scaled_data_type, block_scaled_tile_k, block_scaled_cluster_size, block_scaled_layouts]]) + ").*2sm.*"
sm103_block_scaled_filter_regex_1sm = "cutlass3x_sm103_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [sm103_block_scaled_data_type, block_scaled_cluster_size, block_scaled_layouts]]) + ").*1sm.*"
sm103_block_scaled_filter_regex_2sm = "cutlass3x_sm103_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [sm103_block_scaled_data_type, block_scaled_cluster_size, block_scaled_layouts]]) + ").*2sm.*"
sm103_block_scaled_prefetch_policy = ['tmapf']
sm103_block_scaled_filter_regex_1sm = "cutlass3x_sm103_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [sm103_block_scaled_data_type, sm103_block_scaled_tile_k, block_scaled_cluster_size, block_scaled_layouts]]) + ").*1sm.*(" + "|".join(sm103_block_scaled_prefetch_policy) + ").*"
sm103_block_scaled_filter_regex_2sm = "cutlass3x_sm103_bstensorop.*(" + ").*(".join([ "|".join(x) for x in [sm103_block_scaled_data_type, sm103_block_scaled_tile_k, block_scaled_cluster_size, block_scaled_layouts]]) + ").*2sm.*(" + "|".join(sm103_block_scaled_prefetch_policy) + ").*"
if arch in ["100a", "100f"]:
kernel_filter = f"({sm100_mma_filter_regex_1sm})|" \