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

v4.2 release. (#2587)

Browse Source 
* Fix default cluster callback values to 1 to avoid profiler failure when these values are not set in command line.

* v4.2 release.
This commit is contained in:
Junkai-Wu
2025-08-22 18:11:24 -04:00
committed by GitHub
parent 11cad1f67b
commit a49a78ffef
351 changed files with 28182 additions and 2032 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/79_blackwell_geforce_gemm/79a_blackwell_geforce_nvfp4_bf16_gemm.cu
+20 -11
View File
@@ -88,7 +88,7 @@
using namespace cute;
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
@@ -189,7 +189,7 @@ cutlass::HostTensor<ElementC, cutlass::layout::PackedVectorLayout> block_C;
cutlass::HostTensor<ElementD, cutlass::layout::PackedVectorLayout> block_D;
// Reference Output Tensor
cutlass::HostTensor<ElementD, cutlass::layout::PackedVectorLayout> block_reference_D;
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
@@ -283,7 +283,7 @@ struct Result
};
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
/// GEMM setup and evaluation
@@ -489,19 +489,28 @@ int run(Options &options)
return 0;
}
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
///////////////////////////////////////////////////////////////////////////////////////////////////
int main(int argc, char const **args) {
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit to run this example
// and must have compute capability at least 100.
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit for SM120 support,
// or CUDA 12.9 or higher for SM121 support.
// Must have compute capability at least 120.
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 8)) {
std::cerr << "This example requires CUDA 12.8 or newer." << std::endl;
std::cerr << "This example requires CUDA 12.8 or newer for SM120 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#elif defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 9)) {
std::cerr << "This example requires CUDA 12.9 or newer for SM121 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#endif
cudaDeviceProp props;
int current_device_id;
@@ -509,8 +518,8 @@ int main(int argc, char const **args) {
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
if (!(props.major == 12 && props.minor == 0)) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120)." << std::endl;
if (!(props.major == 12 && (props.minor == 0 || props.minor == 1))) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120 or 121)." << std::endl;
return 0;
}
@@ -530,9 +539,9 @@ int main(int argc, char const **args) {
//
// Evaluate CUTLASS kernels
//
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
run<Gemm>(options);
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
return 0;
}
examples/79_blackwell_geforce_gemm/79b_blackwell_geforce_nvfp4_nvfp4_gemm.cu
+20 -11
View File
@@ -86,7 +86,7 @@
using namespace cute;
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
@@ -217,7 +217,7 @@ cutlass::HostTensor<ElementSFD, cutlass::layout::PackedVectorLayout> block_refer
// Matrix-wide normalization constant
cutlass::HostTensor<ElementCompute, cutlass::layout::PackedVectorLayout> block_Normconst;
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
@@ -311,7 +311,7 @@ struct Result
};
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
/// GEMM setup and evaluation
@@ -536,19 +536,28 @@ int run(Options &options)
return 0;
}
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
///////////////////////////////////////////////////////////////////////////////////////////////////
int main(int argc, char const **args) {
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit to run this example
// and must have compute capability at least 100.
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit for SM120 support,
// or CUDA 12.9 or higher for SM121 support.
// Must have compute capability at least 120.
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 8)) {
std::cerr << "This example requires CUDA 12.8 or newer." << std::endl;
std::cerr << "This example requires CUDA 12.8 or newer for SM120 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#elif defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 9)) {
std::cerr << "This example requires CUDA 12.9 or newer for SM121 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#endif
cudaDeviceProp props;
int current_device_id;
@@ -556,8 +565,8 @@ int main(int argc, char const **args) {
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
if (!(props.major == 12 && props.minor == 0)) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120)." << std::endl;
if (!(props.major == 12 && (props.minor == 0 || props.minor == 1))) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120 or 121)." << std::endl;
return 0;
}
@@ -577,9 +586,9 @@ int main(int argc, char const **args) {
//
// Evaluate CUTLASS kernels
//
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
run<Gemm>(options);
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
return 0;
}
examples/79_blackwell_geforce_gemm/79c_blackwell_geforce_mixed_mxfp8_mxfp6_bf16_gemm.cu
+20 -11
View File
@@ -88,7 +88,7 @@
using namespace cute;
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
@@ -189,7 +189,7 @@ cutlass::HostTensor<ElementC, cutlass::layout::PackedVectorLayout> block_C;
cutlass::HostTensor<ElementD, cutlass::layout::PackedVectorLayout> block_D;
// Reference Output Tensor
cutlass::HostTensor<ElementD, cutlass::layout::PackedVectorLayout> block_reference_D;
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
@@ -283,7 +283,7 @@ struct Result
};
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
/// GEMM setup and evaluation
@@ -489,19 +489,28 @@ int run(Options &options)
return 0;
}
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
///////////////////////////////////////////////////////////////////////////////////////////////////
int main(int argc, char const **args) {
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit to run this example
// and must have compute capability at least 100.
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit for SM120 support,
// or CUDA 12.9 or higher for SM121 support.
// Must have compute capability at least 120.
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 8)) {
std::cerr << "This example requires CUDA 12.8 or newer." << std::endl;
std::cerr << "This example requires CUDA 12.8 or newer for SM120 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#elif defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 9)) {
std::cerr << "This example requires CUDA 12.9 or newer for SM121 support." << std::endl;
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#endif
cudaDeviceProp props;
int current_device_id;
@@ -509,8 +518,8 @@ int main(int argc, char const **args) {
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
if (!(props.major == 12 && props.minor == 0)) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120)." << std::endl;
if (!(props.major == 12 && (props.minor == 0 || props.minor == 1))) {
std::cerr << "This example requires a GPU of NVIDIA's Blackwell architecture (compute capability 120 or 121)." << std::endl;
return 0;
}
@@ -530,9 +539,9 @@ int main(int argc, char const **args) {
//
// Evaluate CUTLASS kernels
//
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
run<Gemm>(options);
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
return 0;
}
examples/79_blackwell_geforce_gemm/79d_blackwell_geforce_nvfp4_grouped_gemm.cu
+18 -9
View File
@@ -97,7 +97,7 @@ using namespace cute;
using ProblemShape = cutlass::gemm::GroupProblemShape<Shape<int,int,int>>; // <M,N,K> per group
using ElementInput = cutlass::float_e2m1_t; // Element type for Input matrix operands
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
/// GEMM kernel configurations
/////////////////////////////////////////////////////////////////////////////////////////////////
@@ -263,7 +263,7 @@ cutlass::DeviceAllocation<ElementAccumulator> block_beta;
// NormConst is a single device-side constant value, its not per-batch or per-group
cutlass::DeviceAllocation<ElementAccumulator> norm_constant_device;
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
template <typename T>
auto make_iterator(T* ptr) {
@@ -466,7 +466,7 @@ struct Result
bool passed = false;
};
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
/////////////////////////////////////////////////////////////////////////////////////////////////
/// GEMM setup and evaluation
@@ -861,30 +861,39 @@ int run(Options &options, bool host_problem_shapes_available = true)
return 0;
}
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#endif // defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
///////////////////////////////////////////////////////////////////////////////////////////////////
int main(int argc, char const **args) {
// CUTLASS must be compiled with CUDA 12.8 Toolkit to run this example
// CUTLASS must be compiled with CUDA 12.8 or higher Toolkit for SM120 support,
// or CUDA 12.9 or higher for SM121 support.
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 ||
((__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 8)
)
) {
std::cerr << "This example requires CUDA 12.8 or newer.\n";
std::cerr << "This example requires CUDA 12.8 or newer for SM120 support.\n";
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#elif defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
if (__CUDACC_VER_MAJOR__ < 12 || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ < 9)) {
std::cerr << "This example requires CUDA 12.9 or newer for SM121 support.\n";
// Returning zero so this test passes on older Toolkits. Its actions are no-op.
return 0;
}
#endif
cudaDeviceProp props;
int current_device_id;
CUDA_CHECK(cudaGetDevice(&current_device_id));
CUDA_CHECK(cudaGetDeviceProperties(&props, current_device_id));
cudaError_t error = cudaGetDeviceProperties(&props, 0);
if (!(props.major == 12 && props.minor == 0)) {
if (!(props.major == 12 && (props.minor == 0 || props.minor == 1))) {
std::cerr
<< "This example requires a GPU of NVIDIA's Blackwell Architecture (compute capability 120a).\n";
<< "This example requires a GPU of NVIDIA's Blackwell Architecture (compute capability 120 or 121).\n";
return 0;
}
@@ -901,7 +910,7 @@ int main(int argc, char const **args) {
return 0;
}
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED) || defined(CUTLASS_ARCH_MMA_SM121_SUPPORTED)
allocate(options);
initialize(options);
examples/79_blackwell_geforce_gemm/CMakeLists.txt
+1 -1
View File
@@ -46,7 +46,7 @@ set(TEST_SMALL_LARGE_GROUP --m=128 --n=128 --groups=50 --iterations=0)
set(TEST_RANDOM_PERF --iterations=10) # Random problem sizes
set(TEST_RANDOM_PERF_LARGE_GROUP --groups=50 --iterations=10) # Random problem sizes
if (CUTLASS_NVCC_ARCHS MATCHES 120a)
if (CUTLASS_NVCC_ARCHS MATCHES "120a|121a")
cutlass_example_add_executable(
79a_blackwell_geforce_nvfp4_bf16_gemm
79a_blackwell_geforce_nvfp4_bf16_gemm.cu