Skip to content

GitLab

Commit 4100d1d8 authored by Alan Turner's avatar Alan Turner
Browse files

Merge remote-tracking branch 'origin/develop' into migx-flash-attn

parents 48717006 c8a8385f
Hide whitespace changes
Inline Side-by-side
Showing with 812 additions and 134 deletions
test/contraction/test_contraction_interface.cpp
View file @ 4100d1d8
......@@ -38,7 +38,7 @@ class ContractionInstanceWrapper
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
// clang-format on
bool isSupported(std::vector<ck::index_t>& ADims,
......
test/data_type/CMakeLists.txt
View file @ 4100d1d8
......@@ -2,3 +2,6 @@ if (USE_BITINT_EXTENSION_INT4)
add_gtest_executable(test_int4 int4.cpp)
target_link_libraries(test_int4 PRIVATE utility)
endif()
add_gtest_executable(test_fp8 fp8.cpp)
target_link_libraries(test_fp8 PRIVATE utility)
test/data_type/fp8.cpp 0 → 100644
View file @ 4100d1d8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/utility/data_type.hpp"
#include "ck/utility/type_convert.hpp"
using ck::f8_convert_sr;
using ck::f8_t;
using ck::half_t;
using ck::type_convert;
TEST(FP8, NumericLimits)
{
EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), 0x08);
EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), 0x77);
EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), 0xF7);
EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), 0x80);
}
TEST(FP8, ConvertFP32Nearest)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(type_convert<f8_t>(0.0f)), abs_tol);
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal f8_t to float and check if equal to 240.0
ASSERT_NEAR(240.0f, type_convert<float>(type_convert<f8_t>(240.0f)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(240.0f,
type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(0x80, type_convert<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
// positive float value to fp8 and back, check if holds
float pos_float = 0.0078125f;
ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
// negative float value to fp8 and back, check if holds
float neg_float = -0.0156250f;
ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
}
TEST(FP8, ConvertFP32Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-6;
// convert 0 float to fp8 and back, check if holds
ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<f8_t>(0.0f)), abs_tol);
// convert minimal float to fp8 and back, check if holds
ASSERT_NEAR(std::numeric_limits<float>::min(),
type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::min())),
abs_tol);
// convert maximal f8_t to float and check if equal to 240.0
ASSERT_NEAR(240.0f, type_convert<float>(f8_convert_sr<f8_t>(240.0f)), abs_tol);
// convert maximal float to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(240.0f,
type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::max())),
abs_tol);
// convert inf float to f8_t and check if it is qNan
ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
// positive float value to fp8 and back, check if holds
float pos_float = 0.0078125f;
ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
// negative float value to fp8 and back, check if holds
float neg_float = -0.0156250f;
ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
}
TEST(FP8, ConvertFP16Nearest)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to fp8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(type_convert<f8_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal f8_t to fp16 and check if equal to 240.0
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(type_convert<f8_t>(half_t{240.0})), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(half_t{240.0},
type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(0x80, type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
// positive fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.0078125};
ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
// negative fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.0156250};
ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
}
TEST(FP8, ConvertFP16Stochastic)
{
// fix the tolerance value
float abs_tol = 1e-3;
// convert 0 fp16 to fp8 and back, check if holds
ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{0.0})), abs_tol);
// convert minimal fp16 to fp8 and back, check if holds
ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Min())),
abs_tol);
// convert maximal f8_t to fp16 and check if equal to 240.0
ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{240.0})), abs_tol);
// convert maximal fp16 to fp8 and back, check if clipped to 240.0
ASSERT_NEAR(half_t{240.0},
type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Max())),
abs_tol);
// convert QuietNaN fp16 to f8_t and check if it is QuietNaN
ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
// positive fp16 value to fp8 and back, check if holds
half_t pos_half = half_t{0.0078125};
ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
// negative fp16 value to fp8 and back, check if holds
half_t neg_half = half_t{-0.0156250};
ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
}
test/elementwise_normalization/CMakeLists.txt
View file @ 4100d1d8
add_custom_target(test_elementwise_normalization)
add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_elementwise_normalization)
add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
endif()
\ No newline at end of file
test/gemm/CMakeLists.txt
View file @ 4100d1d8
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_fp32 gemm_fp32.cpp)
target_link_libraries(test_gemm_fp32 PRIVATE utility)
target_link_libraries(test_gemm_fp32 PRIVATE device_gemm_instance)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_fp16 gemm_fp16.cpp)
target_link_libraries(test_gemm_fp16 PRIVATE utility)
target_link_libraries(test_gemm_fp16 PRIVATE device_gemm_instance)
add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
target_link_libraries(test_gemm_bf16 PRIVATE utility)
target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
add_test_executable(test_gemm_int8 gemm_int8.cpp)
target_link_libraries(test_gemm_int8 PRIVATE utility)
target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
add_library(gemm_standalone_xdl_fp16_instances STATIC
instance/gemm_f16_nn_instance.cpp
instance/gemm_f16_nt_instance.cpp
......@@ -24,3 +17,14 @@ add_library(gemm_standalone_xdl_fp16_instances STATIC
add_test_executable(test_gemm_standalone_xdl_fp16 gemm_standalone_xdl_fp16.cpp)
target_link_libraries(test_gemm_standalone_xdl_fp16 PRIVATE gemm_standalone_xdl_fp16_instances utility)
target_include_directories(test_gemm_standalone_xdl_fp16 PRIVATE instance/)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
target_link_libraries(test_gemm_bf16 PRIVATE utility)
target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_int8 gemm_int8.cpp)
target_link_libraries(test_gemm_int8 PRIVATE utility)
target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
endif()
\ No newline at end of file
test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
View file @ 4100d1d8
......@@ -5,7 +5,7 @@
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......
test/gemm_layernorm/CMakeLists.txt
View file @ 4100d1d8
......@@ -2,10 +2,12 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(test_gemm_layernorm)
add_gtest_executable(test_gemm_add_relu_add_layernorm_fp16 test_gemm_add_relu_add_layernorm_fp16.cpp)
target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance)
add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16)
set(target 1)
endif()
endif()
endforeach()
test/gemm_reduce/CMakeLists.txt
View file @ 4100d1d8
add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
endif()
\ No newline at end of file
test/grouped_convnd_bwd_data/CMakeLists.txt 0 → 100644
View file @ 4100d1d8
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data.cpp)
target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
add_gtest_executable(test_grouped_convnd_bwd_data_interface test_grouped_convnd_bwd_data_interface.cpp)
target_link_libraries(test_grouped_convnd_bwd_data_interface PRIVATE utility device_grouped_conv2d_bwd_data_instance)
endif()
\ No newline at end of file
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp 0 → 100644
View file @ 4100d1d8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdData : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using OutLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using InLayout = std::tuple_element_t<3, Tuple>;
std::vector<ck::utils::conv::ConvParam> conv_params;
template <ck::index_t NDimSpatial>
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
OutLayout,
WeiLayout,
InLayout,
DataType,
DataType,
DataType>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param);
}
EXPECT_TRUE(pass);
}
};
using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
std::tuple<float, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<float, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdData2d : public TestGroupedConvndBwdData<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdData3d : public TestGroupedConvndBwdData<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdData2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdData3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->template Run<3>();
}
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface.cpp 0 → 100644
View file @ 4100d1d8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include <gtest/gtest.h>
using DataType = ck::half_t;
using AccDataType = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ConvBackwardDataSpecialization =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization;
static constexpr auto ConvBwdDataDefault = ConvBackwardDataSpecialization::Default;
static constexpr auto Filter1x1Stride1Pad0 = ConvBackwardDataSpecialization::Filter1x1Stride1Pad0;
template <typename Tuple, ConvBackwardDataSpecialization ConvSpec>
class TestGroupedConvndBwdData : public ::testing::Test
{
protected:
static constexpr ck::index_t NDimSpatial = 2;
using OutLayout = std::tuple_element_t<0, Tuple>;
using WeiLayout = std::tuple_element_t<1, Tuple>;
using InLayout = std::tuple_element_t<2, Tuple>;
// clang-format off
using GroupedConvBwdDataDeviceInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
// ######| NDimSpatial| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ######| | | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, OutLayout, WeiLayout, ck::Tuple<>, InLayout, DataType, DataType, AccDataType, DataType, ck::Tuple<>, DataType, Pass, Pass, Pass, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
ck::utils::conv::ConvParam conv_param;
template <ck::index_t NDimSpatial>
bool Run()
{
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
std::array<ck::index_t, NDimSpatial + 3> out_strides{};
std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};
std::array<ck::index_t, NDimSpatial + 3> wei_strides{};
std::array<ck::index_t, NDimSpatial + 3> in_lengths{};
std::array<ck::index_t, NDimSpatial + 3> in_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), out_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), out_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), wei_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), wei_strides);
copy(in_g_n_c_wis_desc.GetLengths(), in_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), in_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
auto conv = GroupedConvBwdDataDeviceInstance{};
auto argument = conv.MakeArgument(nullptr,
nullptr,
std::array<const void*, 0>{},
nullptr,
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
Pass{},
Pass{},
Pass{});
return conv.IsSupportedArgument(argument);
}
};
using GNHWC = ck::tensor_layout::convolution::GNHWC;
using NHWGC = ck::tensor_layout::convolution::NHWGC;
using GKYXC = ck::tensor_layout::convolution::GKYXC;
using GNHWK = ck::tensor_layout::convolution::GNHWK;
using NHWGK = ck::tensor_layout::convolution::NHWGK;
using KernelTypes =
::testing::Types<std::tuple<GNHWK, GKYXC, GNHWC>, std::tuple<NHWGK, GKYXC, NHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdDataDefault : public TestGroupedConvndBwdData<Tuple, ConvBwdDataDefault>
{
};
template <typename Tuple>
class TestGroupedConvndBwdDataFilter1x1
: public TestGroupedConvndBwdData<Tuple, Filter1x1Stride1Pad0>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdDataDefault, KernelTypes);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataFilter1x1, KernelTypes);
TYPED_TEST(TestGroupedConvndBwdDataFilter1x1, SpecializationCheck)
{
// Check filter 3,3 instead of 1,1
this->conv_param = {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
bool is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Check strides 2,2 instead of 1,1
this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Check with pad
this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Supported version
this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_TRUE(is_supported);
}
TYPED_TEST(TestGroupedConvndBwdDataDefault, VectorLoadCheck)
{
// vector load for A
this->conv_param = {2, 2, 128, 129, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
bool is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// vector load for B, E, Ds
this->conv_param = {2, 2, 128, 128, 257, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
}
test/grouped_convnd_bwd_weight/CMakeLists.txt
View file @ 4100d1d8
......@@ -2,8 +2,10 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_gtest_executable(test_grouped_convnd_bwd_weight grouped_convnd_bwd_weight.cpp)
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_interface test_grouped_convnd_bwd_weight_interface.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_interface PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
set(target 1)
endif()
endforeach()
\ No newline at end of file
test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdWeight : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
std::vector<ck::utils::conv::ConvParam> conv_params;
ck::index_t split_k{2};
template <ck::index_t NDimSpatial>
void Run()
{
for(auto& param : conv_params)
{
bool pass;
EXPECT_FALSE(conv_params.empty());
pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
DataType,
DataType,
DataType>(true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param,
split_k);
EXPECT_TRUE(pass);
}
}
};
using KernelTypes =
::testing::Types<std::tuple<float>, std::tuple<ck::half_t>, std::tuple<ck::bhalf_t>>;
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight, KernelTypes);
TYPED_TEST(TestGroupedConvndBwdWeight, Test1D)
{
this->conv_params.clear();
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
this->template Run<1>();
}
TYPED_TEST(TestGroupedConvndBwdWeight, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdWeight, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->template Run<3>();
}
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp 0 → 100644
View file @ 4100d1d8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdWeight : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, Tuple>;
using WeiDataType = std::tuple_element_t<1, Tuple>;
using OutDataType = std::tuple_element_t<2, Tuple>;
using InLayout = std::tuple_element_t<3, Tuple>;
using WeiLayout = std::tuple_element_t<4, Tuple>;
using OutLayout = std::tuple_element_t<5, Tuple>;
using NDimSpatial = std::tuple_element_t<6, Tuple>;
std::vector<ck::utils::conv::ConvParam> conv_params;
ck::index_t split_k{2};
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial{},
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param,
split_k);
}
EXPECT_TRUE(pass);
}
};
template <typename Tuple>
class TestGroupedConvndBwdWeight1d : public TestGroupedConvndBwdWeight<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdWeight2d : public TestGroupedConvndBwdWeight<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdWeight3d : public TestGroupedConvndBwdWeight<Tuple>
{
};
using namespace ck::tensor_layout::convolution;
using KernelTypes1d = ::testing::Types<
std::tuple<float, float, float, GNWC, GKXC, GNWK, ck::Number<1>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, GNWC, GKXC, GNWK, ck::Number<1>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNWC, GKXC, GNWK, ck::Number<1>>>;
using KernelTypes2d = ::testing::Types<
std::tuple<float, float, float, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
std::tuple<float, float, float, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>>;
using KernelTypes3d = ::testing::Types<
std::tuple<float, float, float, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
std::tuple<float, float, float, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>>;
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight1d, KernelTypes1d);
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdWeight1d, Test1D)
{
this->conv_params.clear();
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
this->conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 1, 1, 1, {3}, {32}, {1}, {1}, {1}, {1}});
this->Run();
}
TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run();
}
TYPED_TEST(TestGroupedConvndBwdWeight3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->Run();
}
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface.cpp 0 → 100644
View file @ 4100d1d8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include <gtest/gtest.h>
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ConvolutionBackwardWeightSpecialization =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization;
static constexpr auto ConvBwdWeightDefault = ConvolutionBackwardWeightSpecialization::Default;
static constexpr auto Filter1x1Stride1Pad0 =
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
template <typename Tuple, ConvolutionBackwardWeightSpecialization ConvSpec>
class TestGroupedConvndBwdWeight : public ::testing::Test
{
protected:
static constexpr ck::index_t NDimSpatial = 2;
using InLayout = std::tuple_element_t<2, Tuple>;
using WeiLayout = std::tuple_element_t<1, Tuple>;
using OutLayout = std::tuple_element_t<0, Tuple>;
// clang-format off
using GroupedConvBwdWeightDeviceInstance = ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle
//##########| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer|
//##########| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector|
//##########| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl|
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| |
< NDimSpatial, InLayout, WeiLayout,OutLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>;
// clang-format on
ck::utils::conv::ConvParam conv_param;
ck::index_t split_k{2};
template <ck::index_t NDimSpatial>
bool Run()
{
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
std::array<ck::index_t, NDimSpatial + 3> input_lengths{};
std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
std::array<ck::index_t, NDimSpatial + 3> input_strides{};
std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
std::array<ck::index_t, NDimSpatial + 3> output_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
range_copy(conv_param.input_left_pads_, begin(input_left_pads));
range_copy(conv_param.input_right_pads_, begin(input_right_pads));
auto conv = GroupedConvBwdWeightDeviceInstance{};
auto argument = conv.MakeArgument(nullptr,
nullptr,
nullptr,
input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{},
split_k);
return conv.IsSupportedArgument(argument);
}
};
using GNHWC = ck::tensor_layout::convolution::GNHWC;
using NHWGC = ck::tensor_layout::convolution::NHWGC;
using GKYXC = ck::tensor_layout::convolution::GKYXC;
using GNHWK = ck::tensor_layout::convolution::GNHWK;
using NHWGK = ck::tensor_layout::convolution::NHWGK;
using KernelTypes =
::testing::Types<std::tuple<GNHWK, GKYXC, GNHWC>, std::tuple<NHWGK, GKYXC, NHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdWeightDefault
: public TestGroupedConvndBwdWeight<Tuple, ConvBwdWeightDefault>
{
};
template <typename Tuple>
class TestGroupedConvndBwdWeightFilter1x1
: public TestGroupedConvndBwdWeight<Tuple, Filter1x1Stride1Pad0>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdWeightDefault, KernelTypes);
TYPED_TEST_SUITE(TestGroupedConvndBwdWeightFilter1x1, KernelTypes);
TYPED_TEST(TestGroupedConvndBwdWeightFilter1x1, SpecializationCheck)
{
// Check filter 3,3 instead of 1,1
this->conv_param = {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
bool is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Check strides 2,2 instead of 1,1
this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Check with pad
this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// Supported version
this->conv_param = {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_TRUE(is_supported);
}
TYPED_TEST(TestGroupedConvndBwdWeightDefault, VectorLoadCheck)
{
// vector load for A
this->conv_param = {2, 2, 128, 129, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
bool is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
// vector load for B, E, Ds
this->conv_param = {2, 2, 128, 128, 257, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
is_supported = this->template Run<2>();
EXPECT_FALSE(is_supported);
}
test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
View file @ 4100d1d8
......@@ -22,6 +22,8 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC)
conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
for(auto& param : conv_params)
{
......@@ -96,6 +98,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC)
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
for(auto& param : conv_params)
{
......@@ -173,6 +178,12 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC)
{3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
for(auto& param : conv_params)
{
......@@ -247,6 +258,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
for(auto& param : conv_params)
{
......@@ -255,7 +269,7 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
// fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWGC,
ck::tensor_layout::convolution::KYXGC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::NHWGK,
ck::half_t,
ck::half_t,
......
test/grouped_gemm/CMakeLists.txt
View file @ 4100d1d8
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......@@ -12,3 +13,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()
test/grouped_gemm/test_grouped_gemm_interface.cpp
View file @ 4100d1d8
......@@ -108,7 +108,7 @@ TEST_F(TestGGemmSplitKInterface_MKNKMN, KLoops)
// kloops % 2
Ks = std::vector<int>{256, 512, 320, 768};
EXPECT_FALSE(
EXPECT_TRUE(
DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, kbatch));
// Not all gemms have same value for main_k0_block_loop!
......
test/grouped_gemm/test_grouped_gemm_util.hpp
View file @ 4100d1d8
......@@ -147,14 +147,14 @@ struct DeviceGroupedGemmSplitkInstanceWrapper
32,
4,
2,
S<1, 4, 32, 1>,
S<1, 4, 16, 1>,
ABlockTransferThreadClusterArrageOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim::value,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1::value,
ABlockLdsAddExtraM::value,
S<1, 4, 32, 1>,
S<1, 4, 16, 1>,
BBlockTransferThreadClusterArrageOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim::value,
......
test/normalization/CMakeLists.txt
View file @ 4100d1d8
add_custom_target(test_normalization)
add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
add_dependencies(test_normalization test_layernorm2d_fp32)
add_dependencies(test_normalization test_layernorm2d_fp16)
add_dependencies(test_normalization test_groupnorm_fp16)
add_dependencies(test_normalization test_groupnorm_fp32)
if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_custom_target(test_normalization)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
add_dependencies(test_normalization test_layernorm2d_fp32)
add_dependencies(test_normalization test_groupnorm_fp32)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
add_dependencies(test_normalization test_layernorm2d_fp16)
add_dependencies(test_normalization test_groupnorm_fp16)
endif()
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment