leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

v4.1 release

Browse Source
This commit is contained in:
Junkai-Wu
2025-07-03 08:07:53 -04:00
committed by GitHub
parent b995f93317
commit a1aaf2300a
155 changed files with 18407 additions and 6068 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/conv/device_3x/fprop/sm100_conv3d_fprop_implicit_gemm_f16_f16_f32_tensorop_f32.cu
+47
View File
@@ -100,6 +100,53 @@ TEST(SM100_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
}
TEST(SM100_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f16ndhwc_tensor_op_f32, 64x64x64_1x1x1_alpha_beta_scaled_bias_relu_residual) {
using ElementAct = cutlass::half_t;
using ElementFlt = cutlass::half_t;
using ElementOut = cutlass::half_t;
using ElementAcc = float;
using ElementCompute = float;
using ElementSrc = cutlass::half_t;
using ElementBias = float;
using MmaTileShape = Shape<_64, _64, Shape<_64>>;
using ClusterShape = Shape<_1,_1,_1>;
using FusionOperation = cutlass::epilogue::fusion::PerColResAddPerColBiasEltAct<
cutlass::epilogue::thread::ReLu, ElementOut, ElementCompute, ElementBias, ElementSrc>;
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassTensorOp,
MmaTileShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAcc, ElementCompute,
ElementSrc, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementSrc>::value,
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
cutlass::epilogue::collective::EpilogueScheduleAuto,
FusionOperation
>::CollectiveOp;
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassTensorOp,
cutlass::conv::Operator::kFprop,
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
ElementAcc,
MmaTileShape, ClusterShape,
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
cutlass::conv::collective::KernelScheduleAuto
>::CollectiveOp;
using ProblemShape=cutlass::conv::ConvProblemShape<CollectiveMainloop::DispatchPolicy::ConvOp, CollectiveMainloop::DispatchPolicy::NumSpatialDimensions>;
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
ProblemShape,
CollectiveMainloop,
CollectiveEpilogue
>;
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
}
//
// Cluster tile shape 128x64x64
// Cluster shape 1x1x1
test/unit/conv/device_3x/fprop/sm100_conv3d_fprop_implicit_gemm_s8_s8_s32_tensorop_s32_with_fusion.cu
+47
View File
@@ -234,6 +234,53 @@ TEST(SM100_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
}
TEST(SM100_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_1x1x1_alpha_beta_scaled_bias_relu_residual) {
using ElementAct = int8_t;
using ElementFlt = int8_t;
using ElementOut = int32_t;
using ElementSrc = int8_t;
using ElementAcc = int32_t;
using ElementCompute = float;
using ElementBias = float;
using MmaTileShape = Shape<_64, _64, Shape<_64>>;
using ClusterShape = Shape<_1,_1,_1>;
using FusionOperation = cutlass::epilogue::fusion::PerColResAddPerColBiasEltAct<
cutlass::epilogue::thread::ReLu, ElementOut, ElementCompute, ElementBias, ElementSrc>;
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassTensorOp,
MmaTileShape, ClusterShape,
cutlass::epilogue::collective::EpilogueTileAuto,
ElementAcc, ElementCompute,
ElementSrc, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementSrc>::value,
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
cutlass::epilogue::collective::EpilogueScheduleAuto,
FusionOperation
>::CollectiveOp;
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
cutlass::arch::Sm100, cutlass::arch::OpClassTensorOp,
cutlass::conv::Operator::kFprop,
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
ElementAcc,
MmaTileShape, ClusterShape,
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
cutlass::conv::collective::KernelScheduleAuto
>::CollectiveOp;
using ProblemShape=cutlass::conv::ConvProblemShape<CollectiveMainloop::DispatchPolicy::ConvOp, CollectiveMainloop::DispatchPolicy::NumSpatialDimensions>;
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
ProblemShape,
CollectiveMainloop,
CollectiveEpilogue
>;
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
}
// alpha != 1 && beta != 0 && bias && gelu
TEST(SM100_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_1x1x1_alpha_beta_bias_gelu) {
using ElementAct = int8_t;
test/unit/conv/device_3x/testbed_conv.hpp
+3
View File
@@ -176,6 +176,8 @@ struct ConvTestbed {
static constexpr bool DisableSource = cute::is_void_v<typename FusionOp::ElementSource>;
static constexpr bool IsResidualEnabled = cutlass::epilogue::collective::detail::IsThreadEpilogueOpWithResidualAdd<FusionOp>::value;
using StrideC = typename Conv::ConvKernel::StrideC;
using StrideD = typename Conv::ConvKernel::StrideD;
using ThreadEpilogueOp = typename Conv::ConvKernel::CollectiveEpilogue::ThreadEpilogueOp;
@@ -494,6 +496,7 @@ struct ConvTestbed {
ElementCompute,
ElementC,
ElementD,
IsResidualEnabled,
decltype(mAlpha),
decltype(mBeta),
decltype(mBias),