More updates for 3.1 (#958)
* Updates for 3.1 * Minor change * doc link fix * Minor updates
This commit is contained in:
ANIKET SHIVAM
GitHub
@ -41,7 +41,6 @@ add_custom_target(
|
||||
cutlass_test_unit_gemm_device_tensorop_planar_complex
|
||||
cutlass_test_unit_gemm_device_sparse_tensorop_sm80
|
||||
cutlass_test_unit_gemv_device
|
||||
cutlass_test_unit_gemv_device_strided_batched
|
||||
cutlass_test_unit_gemm_device_tensorop_sm90
|
||||
cutlass_test_unit_gemm_device_tensorop_cluster_multicast_sm90
|
||||
)
|
||||
@ -61,7 +60,6 @@ add_custom_target(
|
||||
test_unit_gemm_device_tensorop_planar_complex
|
||||
test_unit_gemm_device_sparse_tensorop_sm80
|
||||
test_unit_gemv_device
|
||||
test_unit_gemv_device_strided_batched
|
||||
test_unit_gemm_device_tensorop_sm90
|
||||
)
|
||||
|
||||
@ -500,15 +498,6 @@ cutlass_test_unit_add_executable(
|
||||
gemv.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemv_device_strided_batched
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemv_strided_batched.cu
|
||||
)
|
||||
|
||||
if (NOT CUDA_COMPILER MATCHES "[Cc]lang")
|
||||
|
||||
add_dependencies(
|
||||
|
||||
@ -98,7 +98,7 @@ public:
|
||||
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = 2080
|
||||
uint64_t seed_ = 2023
|
||||
):
|
||||
init_A(init_A_), init_B(init_B_), init_C(init_C_), seed(seed_) { }
|
||||
|
||||
@ -156,22 +156,29 @@ public:
|
||||
|
||||
/// Initializes data structures
|
||||
void initialize(
|
||||
cutlass::MatrixCoord problem_size
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count
|
||||
) {
|
||||
|
||||
//
|
||||
// Allocate the GEMM workspace
|
||||
// Allocate the GEMV workspace
|
||||
//
|
||||
|
||||
tensor_A.resize(problem_size);
|
||||
tensor_B.resize({problem_size.column(), 1});
|
||||
tensor_C.resize({problem_size.row(), 1});
|
||||
tensor_D.resize({problem_size.row(), 1});
|
||||
reference_D.resize({problem_size.row(), 1}, false);
|
||||
if(std::is_same<LayoutA, cutlass::layout::ColumnMajor>::value) {
|
||||
tensor_A.resize({problem_size.row(), batch_count * problem_size.column()});
|
||||
}
|
||||
else {
|
||||
tensor_A.resize({batch_count * problem_size.row(), problem_size.column()});
|
||||
}
|
||||
|
||||
tensor_B.resize({batch_count * problem_size.column(), 1});
|
||||
tensor_C.resize({batch_count * problem_size.row(), 1});
|
||||
tensor_D.resize({batch_count * problem_size.row(), 1});
|
||||
reference_D.resize({batch_count * problem_size.row(), 1}, false);
|
||||
|
||||
EXPECT_TRUE(initialize_tensor(tensor_A.host_view(), init_A, seed + 2019));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_B.host_view(), init_B, seed + 2018));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_C.host_view(), init_C, seed + 2017));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_A.host_view(), init_A, seed + 1));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_B.host_view(), init_B, seed + 2));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_C.host_view(), init_C, seed + 3));
|
||||
|
||||
// It is possible to randomly initialize to all zeros, so override this with non-zeros
|
||||
// in the upper left corner of each operand.
|
||||
@ -225,9 +232,14 @@ public:
|
||||
return passed;
|
||||
}
|
||||
|
||||
/// Verifies the result is a GEMM
|
||||
/// Verifies the result
|
||||
bool verify(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count,
|
||||
int64_t batch_stride_A,
|
||||
int64_t batch_stride_B,
|
||||
int64_t batch_stride_C,
|
||||
int64_t batch_stride_D,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
@ -242,7 +254,7 @@ public:
|
||||
ElementCompute, ElementAccumulator
|
||||
>(
|
||||
{problem_size.row(), 1, problem_size.column()},
|
||||
alpha,
|
||||
alpha,
|
||||
tensor_A.host_ref(),
|
||||
Gemv::kTransformA,
|
||||
tensor_B.host_ref(),
|
||||
@ -250,7 +262,12 @@ public:
|
||||
beta,
|
||||
tensor_C.host_ref(),
|
||||
reference_D.host_ref(),
|
||||
ElementAccumulator(0)
|
||||
ElementAccumulator(0),
|
||||
batch_count,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D
|
||||
);
|
||||
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
@ -259,39 +276,50 @@ public:
|
||||
/// Runs one problem size
|
||||
bool run(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count,
|
||||
int64_t batch_stride_A,
|
||||
int64_t batch_stride_B,
|
||||
int64_t batch_stride_C,
|
||||
int64_t batch_stride_D,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
this->initialize(problem_size);
|
||||
this->initialize(problem_size, batch_count);
|
||||
|
||||
//
|
||||
// Initialize the GEMM operator
|
||||
// Initialize the GEMV operator
|
||||
//
|
||||
|
||||
typename Gemv::Arguments arguments{
|
||||
problem_size,
|
||||
batch_count,
|
||||
{alpha, beta},
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_data(),
|
||||
tensor_C.device_data(),
|
||||
tensor_D.device_data(),
|
||||
tensor_B.layout().stride(0),
|
||||
tensor_C.layout().stride(0),
|
||||
tensor_D.layout().stride(0)
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D
|
||||
};
|
||||
|
||||
Gemv gemm_op;
|
||||
|
||||
cutlass::Status status = gemm_op.can_implement(arguments);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
size_t workspace_size = Gemv::get_workspace_size(arguments);
|
||||
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
cutlass::Status status = gemm_op.initialize(arguments, workspace.get());
|
||||
status = gemm_op.initialize(arguments, workspace.get());
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
//
|
||||
// Run the GEMM
|
||||
// Run the GEMV
|
||||
//
|
||||
|
||||
status = gemm_op();
|
||||
@ -302,8 +330,15 @@ public:
|
||||
// Verify
|
||||
//
|
||||
|
||||
bool passed = this->verify(problem_size, alpha, beta);
|
||||
|
||||
bool passed = this->verify(
|
||||
problem_size,
|
||||
batch_count,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D,
|
||||
alpha,
|
||||
beta);
|
||||
return passed;
|
||||
}
|
||||
};
|
||||
@ -315,12 +350,16 @@ bool TestAllGemv() {
|
||||
|
||||
using ElementCompute = typename Gemv::EpilogueOutputOp::ElementCompute;
|
||||
|
||||
int Batch[] = {
|
||||
1, 520, 1314
|
||||
};
|
||||
|
||||
int M[] = {
|
||||
8, 48, 192, 520
|
||||
1, 5, 16
|
||||
};
|
||||
|
||||
int K[] = {
|
||||
8, 192, 528
|
||||
8, 128, 256
|
||||
};
|
||||
|
||||
double Alpha[] = {
|
||||
@ -331,15 +370,25 @@ bool TestAllGemv() {
|
||||
0, 1, 1.25
|
||||
};
|
||||
|
||||
for (int m : M) {
|
||||
for (int k : K) {
|
||||
for (double alpha : Alpha) {
|
||||
for (double beta : Beta) {
|
||||
for (int b : Batch) {
|
||||
for (int m : M) {
|
||||
for (int k : K) {
|
||||
for (double alpha : Alpha) {
|
||||
for (double beta : Beta) {
|
||||
|
||||
TestbedGemv<Gemv> testbed;
|
||||
TestbedGemv<Gemv> testbed;
|
||||
|
||||
if (!testbed.run({m, k}, ElementCompute(alpha), ElementCompute(beta))) {
|
||||
return false;
|
||||
if (!testbed.run(
|
||||
{m, k},
|
||||
b,
|
||||
m * k,
|
||||
k,
|
||||
m,
|
||||
m,
|
||||
ElementCompute(alpha),
|
||||
ElementCompute(beta))) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -354,66 +403,100 @@ bool TestAllGemv() {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f32n_f32_f32_simt_f32, Simple) {
|
||||
|
||||
using ElementOutput = float;
|
||||
using LayoutA = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementOutput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementOutput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element Accumulator
|
||||
EpilogueOp // Output operator
|
||||
>
|
||||
>;
|
||||
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f16n_f16_f32_simt_f32, Simple) {
|
||||
TEST(SM50_Device_Gemv_f16n_f16_f16_simt_f32, RowMajorA) {
|
||||
|
||||
using ElementInput = cutlass::half_t;
|
||||
using ElementOutput = float;
|
||||
using LayoutA = cutlass::layout::ColumnMajor;
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
int const kElementsPerAccess = 8;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element Accumulator
|
||||
EpilogueOp // Output operator
|
||||
>
|
||||
>;
|
||||
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
EpilogueOp, // Output operator
|
||||
kElementsPerAccess // Element access granularity
|
||||
>
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f16n_f16_f16_simt_f32, Simple) {
|
||||
TEST(SM50_Device_Gemv_f32n_f32_f32_simt_f32, RowMajorA) {
|
||||
|
||||
using ElementInput = float;
|
||||
using ElementOutput = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementAccumulator = float;
|
||||
int const kElementsPerAccess = 4;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
EpilogueOp, // Output operator
|
||||
kElementsPerAccess // Element access granularity
|
||||
>
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f64n_f64_f64_simt_f64, RowMajorA) {
|
||||
|
||||
using ElementInput = double;
|
||||
using ElementOutput = double;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementAccumulator = double;
|
||||
int const kElementsPerAccess = 2;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
EpilogueOp, // Output operator
|
||||
kElementsPerAccess // Element access granularity
|
||||
>
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f16n_f16_f16_simt_f32, ColumnMajorA) {
|
||||
|
||||
using ElementInput = cutlass::half_t;
|
||||
using ElementOutput = cutlass::half_t;
|
||||
@ -442,3 +525,63 @@ TEST(SM50_Device_Gemv_f16n_f16_f16_simt_f32, Simple) {
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f32n_f32_f32_simt_f32, ColumnMajorA) {
|
||||
|
||||
using ElementInput = float;
|
||||
using ElementOutput = float;
|
||||
using LayoutA = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element Accumulator
|
||||
EpilogueOp // Output operator
|
||||
>
|
||||
>;
|
||||
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Gemv_f64n_f64_f64_simt_f64, ColumnMajorA) {
|
||||
|
||||
using ElementInput = double;
|
||||
using ElementOutput = double;
|
||||
using LayoutA = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using Gemv = cutlass::gemm::device::Gemv<
|
||||
cutlass::gemm::kernel::Gemv<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element Accumulator
|
||||
EpilogueOp // Output operator
|
||||
>
|
||||
>;
|
||||
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<Gemv>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@ -1,490 +0,0 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide strided batched GEMV interface
|
||||
*/
|
||||
|
||||
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
#include <sstream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/kernel/gemv_strided_batched.h"
|
||||
#include "cutlass/gemm/device/gemv_strided_batched.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/distribution.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_norm.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/gemm_complex.h"
|
||||
|
||||
#include "testbed_utils.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace test {
|
||||
namespace gemm {
|
||||
|
||||
template <typename GemvStridedBatched>
|
||||
class TestbedStridedBatchedGemv
|
||||
{
|
||||
public:
|
||||
|
||||
using ElementA = typename GemvStridedBatched::ElementA;
|
||||
using LayoutA = typename GemvStridedBatched::LayoutA;
|
||||
using ElementB = typename GemvStridedBatched::ElementB;
|
||||
using ElementC = typename GemvStridedBatched::ElementC;
|
||||
|
||||
using ElementAccumulator = typename GemvStridedBatched::ElementAccumulator;
|
||||
using ElementCompute = typename GemvStridedBatched::EpilogueOutputOp::ElementCompute;
|
||||
|
||||
using LayoutV = cutlass::layout::RowMajor;
|
||||
|
||||
private:
|
||||
|
||||
/// Initialization
|
||||
cutlass::Distribution::Kind init_A;
|
||||
cutlass::Distribution::Kind init_B;
|
||||
cutlass::Distribution::Kind init_C;
|
||||
uint64_t seed;
|
||||
|
||||
cutlass::HostTensor<ElementA, LayoutA> tensor_A;
|
||||
cutlass::HostTensor<ElementB, LayoutV> tensor_B;
|
||||
cutlass::HostTensor<ElementC, LayoutV> tensor_C;
|
||||
cutlass::HostTensor<ElementC, LayoutV> tensor_D;
|
||||
cutlass::HostTensor<ElementC, LayoutV> reference_D;
|
||||
|
||||
public:
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
TestbedStridedBatchedGemv(
|
||||
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = 2023):
|
||||
init_A(init_A_), init_B(init_B_), init_C(init_C_), seed(seed_) {}
|
||||
|
||||
/// Helper to initialize a tensor view
|
||||
template <typename Element, typename Layout>
|
||||
bool initialize_tensor(
|
||||
cutlass::TensorView<Element, Layout> view,
|
||||
cutlass::Distribution::Kind dist_kind,
|
||||
uint64_t seed) {
|
||||
|
||||
if (dist_kind == cutlass::Distribution::Uniform) {
|
||||
|
||||
double scope_max, scope_min;
|
||||
int bits_input = cutlass::sizeof_bits<Element>::value;
|
||||
int bits_output = cutlass::sizeof_bits<typename GemvStridedBatched::ElementC>::value;
|
||||
|
||||
if (bits_input == 1) {
|
||||
scope_max = 2;
|
||||
scope_min = 0;
|
||||
} else if (bits_input <= 8) {
|
||||
scope_max = 2;
|
||||
scope_min = -2;
|
||||
} else if (bits_output == 16) {
|
||||
scope_max = 5;
|
||||
scope_min = -5;
|
||||
} else {
|
||||
scope_max = 8;
|
||||
scope_min = -8;
|
||||
}
|
||||
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
view, seed, scope_max, scope_min, 0);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Identity) {
|
||||
|
||||
cutlass::reference::host::TensorFillIdentity(view);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Gaussian) {
|
||||
|
||||
cutlass::reference::host::TensorFillRandomGaussian(view, seed, 0, 0.5);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Sequential) {
|
||||
|
||||
cutlass::reference::host::BlockFillSequential(
|
||||
view.data(), view.capacity());
|
||||
}
|
||||
else {
|
||||
// TODO: Implement the rest
|
||||
EXPECT_TRUE(false) << "Not implemented";
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Initializes data structures
|
||||
void initialize(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count
|
||||
) {
|
||||
|
||||
//
|
||||
// Allocate the GEMV workspace
|
||||
//
|
||||
|
||||
tensor_A.resize({batch_count * problem_size.row(), problem_size.column()});
|
||||
tensor_B.resize({batch_count * problem_size.column(), 1});
|
||||
tensor_C.resize({batch_count * problem_size.row(), 1});
|
||||
tensor_D.resize({batch_count * problem_size.row(), 1});
|
||||
reference_D.resize({batch_count * problem_size.row(), 1}, false);
|
||||
|
||||
EXPECT_TRUE(initialize_tensor(tensor_A.host_view(), init_A, seed + 1));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_B.host_view(), init_B, seed + 2));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_C.host_view(), init_C, seed + 3));
|
||||
|
||||
// It is possible to randomly initialize to all zeros, so override this with non-zeros
|
||||
// in the upper left corner of each operand.
|
||||
tensor_A.host_view().at({0, 0}) = typename GemvStridedBatched::ElementA(1);
|
||||
tensor_B.host_view().at({0, 0}) = typename GemvStridedBatched::ElementB(1);
|
||||
tensor_C.host_view().at({0, 0}) = typename GemvStridedBatched::ElementC(1);
|
||||
|
||||
cutlass::reference::host::TensorCopy(reference_D.host_view(), tensor_C.host_view());
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_D.sync_device();
|
||||
}
|
||||
|
||||
/// Compares computed reference with device reference and outputs to a file if incorrect
|
||||
bool compare_reference(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
tensor_D.sync_host();
|
||||
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_A.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_B.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_C.host_view()), 0);
|
||||
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_D.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(reference_D.host_view()), 0);
|
||||
|
||||
bool passed = cutlass::reference::host::TensorEquals(reference_D.host_view(), tensor_D.host_view());
|
||||
|
||||
EXPECT_TRUE(passed) << " mismatched reference";
|
||||
|
||||
if (!passed) {
|
||||
|
||||
std::ofstream file("testbed_universal_errors.txt");
|
||||
|
||||
file
|
||||
<< "problem: " << problem_size
|
||||
<< ", alpha: " << alpha << ", beta: " << beta << "\n\n";
|
||||
|
||||
file
|
||||
<< "A =\n" << tensor_A.host_view()
|
||||
<< "\nB =\n" << tensor_B.host_view()
|
||||
<< "\nC =\n" << tensor_C.host_view()
|
||||
<< "\n\nReference =\n" << reference_D.host_view()
|
||||
<< "\nComputed =\n" << tensor_D.host_view();
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
/// Verifies the result
|
||||
bool verify(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count,
|
||||
int64_t batch_stride_A,
|
||||
int64_t batch_stride_B,
|
||||
int64_t batch_stride_C,
|
||||
int64_t batch_stride_D,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
//
|
||||
// Verify
|
||||
//
|
||||
|
||||
cutlass::reference::host::GemmComplex<
|
||||
typename GemvStridedBatched::ElementA, typename GemvStridedBatched::LayoutA,
|
||||
typename GemvStridedBatched::ElementB, LayoutV,
|
||||
typename GemvStridedBatched::ElementC, LayoutV,
|
||||
ElementCompute, ElementAccumulator
|
||||
>(
|
||||
{problem_size.row(), 1, problem_size.column()},
|
||||
alpha,
|
||||
tensor_A.host_ref(),
|
||||
GemvStridedBatched::kTransformA,
|
||||
tensor_B.host_ref(),
|
||||
GemvStridedBatched::kTransformB,
|
||||
beta,
|
||||
tensor_C.host_ref(),
|
||||
reference_D.host_ref(),
|
||||
ElementAccumulator(0),
|
||||
batch_count,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D
|
||||
);
|
||||
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
/// Runs one problem size
|
||||
bool run(
|
||||
cutlass::MatrixCoord problem_size,
|
||||
int32_t batch_count,
|
||||
int64_t batch_stride_A,
|
||||
int64_t batch_stride_B,
|
||||
int64_t batch_stride_C,
|
||||
int64_t batch_stride_D,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
this->initialize(problem_size, batch_count);
|
||||
|
||||
//
|
||||
// Initialize the GEMV operator
|
||||
//
|
||||
|
||||
typename GemvStridedBatched::Arguments arguments{
|
||||
problem_size,
|
||||
batch_count,
|
||||
{alpha, beta},
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_data(),
|
||||
tensor_C.device_data(),
|
||||
tensor_D.device_data(),
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D
|
||||
};
|
||||
|
||||
GemvStridedBatched gemm_op;
|
||||
|
||||
cutlass::Status status = gemm_op.can_implement(arguments);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
size_t workspace_size = GemvStridedBatched::get_workspace_size(arguments);
|
||||
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
status = gemm_op.initialize(arguments, workspace.get());
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
//
|
||||
// Run the GEMV
|
||||
//
|
||||
|
||||
status = gemm_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
//
|
||||
// Verify
|
||||
//
|
||||
|
||||
bool passed = this->verify(
|
||||
problem_size,
|
||||
batch_count,
|
||||
batch_stride_A,
|
||||
batch_stride_B,
|
||||
batch_stride_C,
|
||||
batch_stride_D,
|
||||
alpha,
|
||||
beta);
|
||||
return passed;
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename GemvStridedBatched>
|
||||
bool TestAllGemv() {
|
||||
|
||||
using ElementCompute = typename GemvStridedBatched::EpilogueOutputOp::ElementCompute;
|
||||
|
||||
int Batch[] = {
|
||||
1, 520, 1314
|
||||
};
|
||||
|
||||
int M[] = {
|
||||
1, 5, 16
|
||||
};
|
||||
|
||||
int K[] = {
|
||||
8, 128, 256
|
||||
};
|
||||
|
||||
double Alpha[] = {
|
||||
1, 1.25
|
||||
};
|
||||
|
||||
double Beta[] = {
|
||||
0, 1, 1.25
|
||||
};
|
||||
|
||||
for (int b : Batch) {
|
||||
for (int m : M) {
|
||||
for (int k : K) {
|
||||
for (double alpha : Alpha) {
|
||||
for (double beta : Beta) {
|
||||
|
||||
TestbedStridedBatchedGemv<GemvStridedBatched> testbed;
|
||||
|
||||
if (!testbed.run(
|
||||
{m, k},
|
||||
b,
|
||||
m * k,
|
||||
k,
|
||||
m,
|
||||
m,
|
||||
ElementCompute(alpha),
|
||||
ElementCompute(beta))) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
} // namespace gemm
|
||||
} // namespace test
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_StridedBatchedGemv_f16n_f16_f16_simt_f32, Simple) {
|
||||
|
||||
using ElementInput = cutlass::half_t;
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
|
||||
using ElementAccumulator = float;
|
||||
int const kElementsPerAccess = 8;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using GemvStridedBatched = cutlass::gemm::device::GemvStridedBatched<
|
||||
cutlass::gemm::kernel::GemvStridedBatched<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
kElementsPerAccess, // Element access granularity
|
||||
EpilogueOp // Output operator
|
||||
>>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<GemvStridedBatched>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_StridedBatchedGemv_f32n_f32_f32_simt_f32, Simple) {
|
||||
|
||||
using ElementInput = float;
|
||||
using ElementOutput = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
|
||||
using ElementAccumulator = float;
|
||||
int const kElementsPerAccess = 4;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using GemvStridedBatched = cutlass::gemm::device::GemvStridedBatched<
|
||||
cutlass::gemm::kernel::GemvStridedBatched<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
kElementsPerAccess, // Element access granularity
|
||||
EpilogueOp // Output operator
|
||||
>>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<GemvStridedBatched>());}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_StridedBatchedGemv_f64n_f64_f64_simt_f64, Simple) {
|
||||
|
||||
using ElementInput = double;
|
||||
using ElementOutput = double;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
|
||||
using ElementAccumulator = double;
|
||||
int const kElementsPerAccess = 2;
|
||||
|
||||
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator>;
|
||||
|
||||
using GemvStridedBatched = cutlass::gemm::device::GemvStridedBatched<
|
||||
cutlass::gemm::kernel::GemvStridedBatched<
|
||||
ElementInput, // Element A
|
||||
LayoutA, // Layout A
|
||||
ElementInput, // Element B
|
||||
ElementOutput, // Element C
|
||||
ElementAccumulator, // Element accumulator
|
||||
kElementsPerAccess, // Element access granularity
|
||||
EpilogueOp // Output operator
|
||||
>>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::TestAllGemv<GemvStridedBatched>());}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -773,7 +773,7 @@ TEST(SM90_Device_Gemm_f16n_f16n_f32n_tensor_op_gmma_f32_cooperative, 256x128x64_
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 256x128x64_2x2x1) {
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_cooperative_epilogue, 256x128x64_2x2x1) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
@ -810,7 +810,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 256x128x64_2x2x1) {
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_cooperative_epilogue, 256x128x64_2x2x1) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
@ -847,4 +847,78 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 128x128x64_2x2x1) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
float, LayoutC, 4,
|
||||
float, LayoutC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::half_t, LayoutA, 8,
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 128x128x64_2x2x1) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
float, LayoutC, 4,
|
||||
float, LayoutC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::half_t, LayoutA, 8,
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
@ -363,4 +363,48 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 256x128x64_2x2x1_BiasMul_ReLU_VoidC) {
|
||||
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
using TileShape_MNK = Shape<_256,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
|
||||
static constexpr bool StoreT = true;
|
||||
using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperativeBiasElementwise<
|
||||
cutlass::epilogue::thread::ReLu, cutlass::half_t, cutlass::multiplies, StoreT, float>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
void, LayoutC, 8,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
EpilogueSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::half_t, LayoutA, 8,
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllBiasElementwise<Gemm>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
@ -1104,13 +1104,13 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_persistent_Epilogue, 128
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent, 128x128x64_2x2x1) {
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_persistent_epilogue, 128x128x64_2x2x1) {
|
||||
using ElementA = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementB = cutlass::half_t;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
using ElementC = ElementA;
|
||||
using ElementC = cutlass::half_t;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
@ -1121,20 +1121,20 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent, 128x128x64_2
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
ElementAccumulator, ElementAccumulator,
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
ElementA, LayoutA, 8,
|
||||
ElementB, LayoutB, 8,
|
||||
ElementA, LayoutA, 16 / sizeof(ElementA),
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
@ -1147,13 +1147,13 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent, 128x128x64_2
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent, 128x128x64_2x2x1) {
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_persistent_epilogue, 128x128x64_2x2x1) {
|
||||
using ElementA = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementB = cutlass::half_t;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
using ElementC = ElementA;
|
||||
using ElementC = cutlass::half_t;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
@ -1164,20 +1164,106 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent, 128x128x64_2
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
ElementAccumulator, ElementAccumulator,
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
ElementA, LayoutA, 8,
|
||||
ElementB, LayoutB, 8,
|
||||
ElementA, LayoutA, 16 / sizeof(ElementA),
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent_epilogue, 128x128x64_2x2x1) {
|
||||
using ElementA = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementB = cutlass::half_t;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
using ElementC = float;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
using StageCountType = cutlass::gemm::collective::StageCountAuto;
|
||||
using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAccumulator, ElementAccumulator,
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
ElementA, LayoutA, 16 / sizeof(ElementA),
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128x128x64_2x2x1) {
|
||||
using ElementA = cutlass::half_t;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using ElementB = cutlass::half_t;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using ElementAccumulator = float;
|
||||
using ElementC = float;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
using StageCountType = cutlass::gemm::collective::StageCountAuto;
|
||||
using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAccumulator, ElementAccumulator,
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
ElementC, LayoutC, 16 / sizeof(ElementC),
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
ElementA, LayoutA, 16 / sizeof(ElementA),
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
|
||||
@ -362,4 +362,48 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128x128x64_2x2x1_BiasMul_ReLU_VoidC) {
|
||||
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
using TileShape_MNK = Shape<_128,_128,_64>;
|
||||
using ClusterShape_MNK = Shape<_2,_2,_1>;
|
||||
|
||||
static constexpr bool StoreT = true;
|
||||
using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedBiasElementwise<
|
||||
cutlass::epilogue::thread::ReLu, cutlass::half_t, cutlass::multiplies, StoreT, float>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
float, float,
|
||||
void, LayoutC, 8,
|
||||
cutlass::half_t, LayoutC, 8,
|
||||
EpilogueSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::half_t, LayoutA, 8,
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllBiasElementwise<Gemm>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
@ -269,6 +269,150 @@ TEST(SM90_Device_Gemm_s8t_s8n_s8n_tensor_op_gmma_s32, 128x128x128_2x2x1) {
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_s8t_s8n_s8n_tensor_op_gmma_s32_pingpong_epilogue, 64x128x128) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
Shape<_64,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
int32_t, int32_t,
|
||||
int8_t, LayoutC, 16,
|
||||
int8_t, LayoutC, 16,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveOp = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
int8_t, LayoutA, 16,
|
||||
int8_t, LayoutB, 16,
|
||||
int32_t,
|
||||
Shape<_64,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveOp,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_s8t_s8n_s8t_tensor_op_gmma_s32_pingpong_epilogue, 64x128x128) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
Shape<_64,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
int32_t, int32_t,
|
||||
int8_t, LayoutC, 16,
|
||||
int8_t, LayoutC, 16,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveOp = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
int8_t, LayoutA, 16,
|
||||
int8_t, LayoutB, 16,
|
||||
int32_t,
|
||||
Shape<_64,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveOp,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_s8t_s8n_s8n_tensor_op_gmma_s32_cooperative_epilogue, 128x128x128) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::ColumnMajor;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
Shape<_128,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
int32_t, int32_t,
|
||||
int8_t, LayoutC, 16,
|
||||
int8_t, LayoutC, 16,
|
||||
cutlass::epilogue::TmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveOp = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
int8_t, LayoutA, 16,
|
||||
int8_t, LayoutB, 16,
|
||||
int32_t,
|
||||
Shape<_128,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveOp,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM90_Device_Gemm_s8t_s8n_s8t_tensor_op_gmma_s32_cooperative_epilogue, 128x128x128) {
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
Shape<_128,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
int32_t, int32_t,
|
||||
int8_t, LayoutC, 16,
|
||||
int8_t, LayoutC, 16,
|
||||
cutlass::epilogue::TmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveOp = typename cutlass::gemm::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
int8_t, LayoutA, 16,
|
||||
int8_t, LayoutB, 16,
|
||||
int32_t,
|
||||
Shape<_128,_128,_128>, Shape<_1,_1,_1>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
|
||||
Shape<int,int,int,int>,
|
||||
CollectiveOp,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
EXPECT_TRUE(test::gemm::device::TestAll<Gemm>());
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
Reference in New Issue
Block a user