leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

update 3.8 v2 (#2112)

Browse Source 
* update 3.8 v2

* update 3.8

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
This commit is contained in:
Yujia Zhai
2025-02-19 22:03:14 -05:00
committed by GitHub
co-authored by yuzhai
parent e9627ce55b
commit b84e9802d8
166 changed files with 3986 additions and 4037 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/CMakeLists.txt
+10 -10
View File
@@ -30,7 +30,7 @@
#
if(NOT CUTLASS_NVCC_ARCHS STREQUAL "100")
if (CUTLASS_NVCC_ARCHS MATCHES 100a)
add_custom_target(
cutlass_test_unit_gemm_device_sm100_blockscaled
DEPENDS
@@ -57,7 +57,7 @@ cutlass_test_unit_gemm_device_add_executable(
nvf4_nvf4_f16_nvfp4_epilogue.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf4xmxf4
BATCH_SOURCES ON
@@ -67,7 +67,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf4_mxf4_void_f16_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf6xmxf6
BATCH_SOURCES ON
@@ -77,7 +77,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf6_mxf6_void_bf16_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf8xmxf8
BATCH_SOURCES ON
@@ -87,7 +87,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf8_mxf8_void_f8_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf6xmxf8
BATCH_SOURCES ON
@@ -97,7 +97,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf6_mxf8_void_f32_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf8xmxf6
BATCH_SOURCES ON
@@ -107,7 +107,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf8_mxf6_f16_f8_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf4xmxf8
BATCH_SOURCES ON
@@ -117,7 +117,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf4_mxf8_bf16_bf16_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf8xmxf4
BATCH_SOURCES ON
@@ -127,7 +127,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf8_mxf4_f16_bf16_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf6xmxf4
BATCH_SOURCES ON
@@ -137,7 +137,7 @@ cutlass_test_unit_gemm_device_add_executable_split_file(
mxf6_mxf4_f16_f16_nt_layout.cu
)
cutlass_test_unit_gemm_device_add_executable_split_file(
cutlass_test_unit_gemm_device_add_executable(
cutlass_test_unit_gemm_device_bstensorop_sm100_mxf4xmxf6
BATCH_SOURCES ON
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf4_void_f16_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -117,7 +115,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -181,8 +179,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -190,12 +186,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -254,8 +250,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -263,12 +257,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m1t_void_f16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf4_void_f16_tn_layout.cu
+10 -22
View File
@@ -109,8 +109,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -118,7 +116,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_192,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_256>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -328,8 +322,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -337,12 +329,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -401,8 +393,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_192,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_256>;
//
// Construct CollectiveEpilogue
@@ -410,7 +400,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -474,8 +464,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -483,12 +471,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_void_f16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf6_f32_f16_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f32_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f32_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,7 +258,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe2m3t_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf6_f32_f16_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f32_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f32_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,7 +259,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe3m2n_f32_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,7 +472,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m3n_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf8_bf16_bf16_nt_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,7 +187,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,12 +259,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,7 +330,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe4m3t_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,12 +472,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1n_ue8m0xe5m2t_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf4_mxf8_bf16_bf16_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,7 +115,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,12 +259,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,7 +330,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe4m3n_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,12 +472,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe5m2n_bf16_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf4_f16_f16_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,7 +258,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m1t_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf4_f16_f16_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,7 +259,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m1n_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,7 +472,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m1n_f16_f16t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf6_void_bf16_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -191,7 +187,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe2m3t_void_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf6_void_bf16_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,7 +187,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 12
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,12 +259,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 12
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,7 +401,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe3m2n_void_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 25
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,12 +472,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe2m3n_void_bf16t_bstensorop_f32, 25
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf8_void_f32_nt_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,7 +187,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -328,8 +322,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -337,7 +329,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -401,8 +393,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -410,12 +400,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3n_ue8m0xe4m3t_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -474,8 +464,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -483,12 +471,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2n_ue8m0xe5m2t_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf8_void_f32_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,7 +259,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe3m2t_ue8m0xe4m3n_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,7 +472,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m3t_ue8m0xe5m2n_void_f32t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf4_f16_bf16_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,7 +115,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m1t_f16_bf16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf4_f16_bf16_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,7 +187,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 128
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -328,8 +322,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -337,7 +329,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -401,8 +393,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -410,12 +400,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -474,8 +464,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -483,12 +471,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m1n_f16_bf16t_bstensorop_f32, 256
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf6_f16_f8_nt_layout.cu
+5 -11
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 128x1
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,7 +115,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 128x1
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 128x2
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 128x2
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 256x2
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe2m3t_f16_f8t_bstensorop_f32, 256x2
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf6_f16_f8_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 128x1
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 128x1
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 128x1
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 128x1
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 128x2
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,7 +259,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 128x2
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 256x1
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 256x1
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 256x1
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe3m2n_f16_f8t_bstensorop_f32, 256x1
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 256x2
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,7 +472,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe2m3n_f16_f8t_bstensorop_f32, 256x2
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_nt_layout.cu
+10 -22
View File
@@ -107,8 +107,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -116,7 +114,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -181,8 +179,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -190,12 +186,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -255,8 +251,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -264,12 +258,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -328,8 +322,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -337,7 +329,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -401,8 +393,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -410,12 +400,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2n_ue8m0xe4m3t_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -474,8 +464,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -483,12 +471,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3n_ue8m0xe5m2t_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_tn_layout.cu
+10 -22
View File
@@ -108,8 +108,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -117,12 +115,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -182,8 +180,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -191,12 +187,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -256,8 +252,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 128x
using MmaTileShape_MNK = Shape<_128,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -265,7 +259,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 128x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -329,8 +323,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_128,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_128>;
//
// Construct CollectiveEpilogue
@@ -338,12 +330,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -402,8 +394,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_192,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_128>;
//
// Construct CollectiveEpilogue
@@ -411,12 +401,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe5m2t_ue8m0xe4m3n_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -475,8 +465,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 256x
using MmaTileShape_MNK = Shape<_256,_256,_128>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_128>;
//
// Construct CollectiveEpilogue
@@ -484,7 +472,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe4m3t_ue8m0xe5m2n_void_f8t_bstensorop_f32, 256x
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/nvf4_nvf4_bf16_bf16.cu
+12 -28
View File
@@ -107,8 +107,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -116,7 +114,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -180,8 +178,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -189,7 +185,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -259,8 +255,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -268,7 +262,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -334,8 +328,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -343,7 +335,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -413,8 +405,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_192,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_256>;
//
// Construct CollectiveEpilogue
@@ -422,12 +412,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -486,8 +476,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_1,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -495,12 +483,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -561,8 +549,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_192,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_192,_256>;
//
// Construct CollectiveEpilogue
@@ -570,12 +556,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -634,8 +620,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_256,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_256,_256>;
//
// Construct CollectiveEpilogue
@@ -643,12 +627,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/nvf4_nvf4_bf16_bf16_features.cu
+6 -14
View File
@@ -92,8 +92,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -101,12 +99,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -165,8 +163,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct CollectiveEpilogue
@@ -174,12 +170,12 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -244,8 +240,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_1,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
// Tile Scheduler
using TileScheduler = cutlass::gemm::StreamKScheduler;
@@ -256,7 +250,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -321,8 +315,6 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_2,_2,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
// Tile Scheduler
using TileScheduler = cutlass::gemm::StreamKScheduler;
@@ -333,7 +325,7 @@ TEST(SM100Only_Device_Gemm_NVue4m3xe2m1t_NVue4m3xe2m1n_bf16t_bf16t_bstensorop_f3
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/nvf4_nvf4_f16_nvfp4_epilogue.cu
+7 -15
View File
@@ -104,8 +104,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1t_outputVs16_bstens
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
// Mma's accumulator type
using ElementAccumulator = float;
@@ -130,12 +128,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1t_outputVs16_bstens
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized1Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -195,8 +193,6 @@ TEST(SM100Only_Device_Gemm_ue4m3xe2m1t_ue4m3xe2m1n_ue4m3xe2m1t_outputVs16_bstens
using MmaTileShape_MNK = Shape<_256,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
// Mma's accumulator type
using ElementAccumulator = float;
@@ -220,12 +216,12 @@ TEST(SM100Only_Device_Gemm_ue4m3xe2m1t_ue4m3xe2m1n_ue4m3xe2m1t_outputVs16_bstens
//
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
ElementD, GmemLayoutD, AlignD, // D tensor description
cutlass::epilogue::collective::EpilogueScheduleAuto // Epilogue schedule policy
cutlass::epilogue::TmaWarpSpecialized2Sm // Epilogue schedule policy
>::CollectiveOp;
//
@@ -289,8 +285,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1t_outputVs32_bstens
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
//
// Construct FusionOperation
@@ -310,7 +304,7 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1t_outputVs32_bstens
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp, // Arch and Tensorop spec
PerSmTileShape_MNK, ClusterShape_MNK, // Epilogue tile shape, and cluster shape
MmaTileShape_MNK, ClusterShape_MNK, // Mma instruction tile shape, cluster shape
cutlass::epilogue::collective::EpilogueTileAuto, // Epilogue subtile shape. Auto will find a suitable tile shape
ElementAccumulator, ElementCompute, // Mma instr's accumulator type and compute precision for epilogue
ElementC, GmemLayoutC, AlignC, // C tensor description
@@ -382,8 +376,6 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1n_outputVs16_bstens
using MmaTileShape_MNK = Shape<_128,_128,_256>;
// Cluster size for multicast
using ClusterShape_MNK = Shape<_4,_4,_1>;
// Collective Epilogue takes the output tile shape for 1 CTA
using PerSmTileShape_MNK = Shape<_128,_128,_256>;
// Mma's accumulator type
using ElementAccumulator = float;
@@ -399,12 +391,12 @@ TEST(SM100Only_Device_Gemm_ue8m0xe2m1t_ue8m0xe2m1n_ue8m0xe2m1n_outputVs16_bstens
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassBlockScaledTensorOp,
PerSmTileShape_MNK, ClusterShape_MNK,
MmaTileShape_MNK, ClusterShape_MNK,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAccumulator, ElementCompute,
ElementC, GmemLayoutC, AlignC,
ElementD, GmemLayoutC, AlignD,
cutlass::epilogue::collective::EpilogueScheduleAuto,
cutlass::epilogue::TmaWarpSpecialized1Sm,
FusionOperation
>::CollectiveOp;