ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix merge from upstream

Browse Source
This commit is contained in:
carlushuang
2024-03-26 14:09:54 +00:00
parent c94b545747
commit 04ee01191a
105 changed files with 16558 additions and 2285 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
test/wrapper/CMakeLists.txt
View File

@@ -1,14 +1,21 @@
add_gtest_executable(test_layout test_layout.cpp)
target_link_libraries(test_layout PRIVATE utility)
add_gtest_executable(test_tensor test_tensor.cpp)
target_link_libraries(test_tensor PRIVATE utility)
add_gtest_executable(test_copy test_copy.cpp)
target_link_libraries(test_copy PRIVATE utility)
add_gtest_executable(test_partition test_partition.cpp)
target_link_libraries(test_partition PRIVATE utility)
add_custom_target(test_wrapper)
add_gtest_executable(test_wrapper_layout test_wrapper_layout.cpp)
target_link_libraries(test_wrapper_layout PRIVATE utility)
add_dependencies(test_wrapper test_wrapper_layout)
add_gtest_executable(test_wrapper_tensor test_wrapper_tensor.cpp)
target_link_libraries(test_wrapper_tensor PRIVATE utility)
add_dependencies(test_wrapper test_wrapper_tensor)
add_gtest_executable(test_wrapper_copy test_wrapper_copy.cpp)
target_link_libraries(test_wrapper_copy PRIVATE utility)
add_dependencies(test_wrapper test_wrapper_copy)
add_gtest_executable(test_wrapper_partition test_wrapper_partition.cpp)
target_link_libraries(test_wrapper_partition PRIVATE utility)
add_dependencies(test_wrapper test_wrapper_partition)
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR
GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR
GPU_TARGETS MATCHES "gfx942" OR GPU_TARGETS MATCHES "gfx950")
add_gtest_executable(test_gemm test_gemm.cpp)
target_link_libraries(test_gemm PRIVATE utility)
GPU_TARGETS MATCHES "gfx942")
add_gtest_executable(test_wrapper_gemm test_wrapper_gemm.cpp)
target_link_libraries(test_wrapper_gemm PRIVATE utility)
add_dependencies(test_wrapper test_wrapper_gemm)
endif()

135
test/wrapper/test_wrapper_copy.cpp Normal file
View File

@@ -0,0 +1,135 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <numeric>
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/common_header.hpp"
#include "ck/wrapper/layout.hpp"
#include "ck/wrapper/tensor.hpp"
#include "ck/wrapper/operations/copy.hpp"
// Test copy from Global to Global through LDS and VGPR
template <typename InputTensor,
typename OutputTensor,
typename BlockShape,
typename ThreadLayout,
bool UseOptimizedCopy>
__global__ void TestCopyDevice(const InputTensor input_tensor,
OutputTensor output_tensor,
const BlockShape tile_shape,
const ThreadLayout thread_layout)
{
__shared__ ck::index_t p_shared[ck::wrapper::size(tile_shape)];
const auto tensor_lds = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
p_shared, ck::wrapper::make_layout(tile_shape));
const auto block_idxs =
ck::make_tuple(static_cast<ck::index_t>(blockIdx.x), static_cast<ck::index_t>(blockIdx.y));
// Get local tiles for global memory
const auto input_local_tile =
ck::wrapper::make_local_tile(input_tensor, tile_shape, block_idxs);
const auto output_local_tile =
ck::wrapper::make_local_tile(output_tensor, tile_shape, block_idxs);
// Get partition per thread
const auto input_local_partition =
ck::wrapper::make_local_partition(input_local_tile, thread_layout, threadIdx.x);
auto lds_local_partition =
ck::wrapper::make_local_partition(tensor_lds, thread_layout, threadIdx.x);
auto output_local_partition =
ck::wrapper::make_local_partition(output_local_tile, thread_layout, threadIdx.x);
// Allocate VGPR
auto tensor_vgpr =
ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, ck::index_t>(
ck::wrapper::make_layout(shape(lds_local_partition)));
// Perform copy
if constexpr(UseOptimizedCopy)
{
using DimAccessOrder = ck::Tuple<ck::Number<1>, ck::Number<0>>;
constexpr ck::index_t vector_dim = 0;
constexpr ck::index_t scalar_per_vector = 2;
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(input_local_partition,
lds_local_partition);
// TODO: Enable optimized copy for static buffers
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(lds_local_partition,
tensor_vgpr);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(tensor_vgpr,
output_local_partition);
}
else
{
ck::wrapper::copy(input_local_partition, lds_local_partition);
ck::wrapper::copy(lds_local_partition, tensor_vgpr);
ck::wrapper::copy(tensor_vgpr, output_local_partition);
}
}
template <bool UseOptimizedCopy>
void PerformCopyGlobalToGlobalViaLDS()
{
const auto shape =
ck::make_tuple(ck::make_tuple(ck::Number<2>{}, ck::Number<2>{}), ck::Number<256>{});
const auto strides =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<2>{}), ck::Number<4>{});
const auto layout = ck::wrapper::make_layout(shape, strides);
// 0, 1, 2, ..., size(shape) - 1
std::vector<ck::index_t> input_data(ck::wrapper::size(shape));
std::iota(input_data.begin(), input_data.end(), 0);
// Global memory buffers
DeviceMem in_buf(ck::wrapper::size(layout) * sizeof(ck::index_t));
DeviceMem out_buf(ck::wrapper::size(layout) * sizeof(ck::index_t));
in_buf.ToDevice(input_data.data());
out_buf.SetZero();
// Create tensors for global memory
const auto input_tensor_global = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<const ck::index_t*>(in_buf.GetDeviceBuffer()), layout);
auto output_tensor_global = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<ck::index_t*>(out_buf.GetDeviceBuffer()), layout);
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<1>{}, ck::Number<32>{}));
const auto tile_shape = ck::make_tuple(ck::Number<4>{}, ck::Number<64>{});
const ck::index_t grid_size_x = ck::math::integer_divide_ceil(
ck::wrapper::size<0>(input_tensor_global), ck::wrapper::size<0>(tile_shape));
const ck::index_t grid_size_y = ck::math::integer_divide_ceil(
ck::wrapper::size<1>(input_tensor_global), ck::wrapper::size<1>(tile_shape));
const auto kernel = TestCopyDevice<decltype(input_tensor_global),
decltype(output_tensor_global),
decltype(tile_shape),
decltype(thread_layout),
UseOptimizedCopy>;
launch_and_time_kernel(StreamConfig{},
kernel,
dim3(grid_size_x, grid_size_y, 1),
dim3(ck::wrapper::size(thread_layout)),
0,
input_tensor_global,
output_tensor_global,
tile_shape,
thread_layout);
// Verify results
std::vector<ck::index_t> output_data(ck::wrapper::size(shape));
out_buf.FromDevice(output_data.data());
EXPECT_TRUE(ck::utils::check_err(output_data, input_data));
}
TEST(TestCopyGlobalToGlobalViaLDS, GenericCopy) { PerformCopyGlobalToGlobalViaLDS<false>(); }
TEST(TestCopyGlobalToGlobalViaLDS, OptimizedCopy) { PerformCopyGlobalToGlobalViaLDS<true>(); }

376
test/wrapper/test_wrapper_gemm.cpp Normal file
View File

@@ -0,0 +1,376 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <numeric>
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "ck/library/utility/host_tensor.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/common_header.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/wrapper/layout.hpp"
#include "ck/wrapper/tensor.hpp"
#include "ck/wrapper/operations/copy.hpp"
#include "ck/wrapper/operations/gemm.hpp"
#include "ck/wrapper/utils/kernel_utils.hpp"
template <typename DataType>
void CheckResult(const std::vector<DataType>& a_data,
const std::vector<DataType>& b_data,
std::vector<DataType>& c_m_n_device_result,
const ck::index_t M,
const ck::index_t N,
const ck::index_t K)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<DataType, DataType, DataType, float, PassThrough, PassThrough, PassThrough>;
Tensor<DataType> a_m_k(HostTensorDescriptor({M, K}));
Tensor<DataType> b_k_n(HostTensorDescriptor({K, N}, {1, K}));
Tensor<DataType> c_m_n_host_result(HostTensorDescriptor({M, N}));
a_m_k.mData = a_data;
b_k_n.mData = b_data;
auto ref_op = ReferenceGemmInstance{};
auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument = ref_op.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
EXPECT_TRUE(ck::utils::check_err(c_m_n_device_result, c_m_n_host_result.mData));
}
template <bool DoPad, typename Layout, typename PaddingDims>
__device__ auto ApplyPadding(const Layout& layout, const PaddingDims& padding_dims)
{
if constexpr(DoPad)
{
return ck::wrapper::pad(layout, padding_dims);
}
else
{
return layout;
}
}
template <typename DataType,
typename GemmTraits,
ck::index_t scalar_per_vector,
typename BlockShape,
typename ThreadLayout,
bool DoPadding>
__global__ void __CK_WRAPPER_LAUNCH_BOUNDS__ DeviceGemm(const void* p_a,
const void* p_b,
void* p_c,
const ck::index_t M,
const ck::index_t N,
const ck::index_t K,
const BlockShape tile_shape,
const ThreadLayout thread_layout)
{
constexpr auto MPerBlock = ck::wrapper::size<0>(tile_shape);
constexpr auto NPerBlock = ck::wrapper::size<1>(tile_shape);
constexpr auto KPerBlock = ck::wrapper::size<2>(tile_shape);
constexpr auto K1 = GemmTraits::K1;
constexpr auto K0PerBlock = KPerBlock / K1;
const auto K0 = ck::math::integer_divide_ceil(K, K1);
const auto tile_shape_k0_m_n_k1 = ck::make_tuple(K0PerBlock, MPerBlock, NPerBlock, K1);
const auto a_global_layout =
ck::wrapper::make_layout(ck::make_tuple(M, K), ck::make_tuple(K, 1));
const auto b_global_layout =
ck::wrapper::make_layout(ck::make_tuple(N, K), ck::make_tuple(K, 1));
const auto c_global_layout =
ck::wrapper::make_layout(ck::make_tuple(M, N), ck::make_tuple(N, 1));
auto a_padded_global_layout =
ApplyPadding<DoPadding>(a_global_layout, ck::make_tuple(MPerBlock, KPerBlock));
auto b_padded_global_layout =
ApplyPadding<DoPadding>(b_global_layout, ck::make_tuple(NPerBlock, KPerBlock));
auto c_padded_global_layout =
ApplyPadding<DoPadding>(c_global_layout, ck::make_tuple(MPerBlock, NPerBlock));
// Reshape from M,K to K0,M,K1
const auto reshaped_dims_idxs =
ck::make_tuple(ck::Number<1>{}, ck::make_tuple(ck::Number<0>{}, ck::Number<2>{}));
auto a_padded_unmerged_global_layout =
ck::wrapper::unmerge<1>(a_padded_global_layout, ck::make_tuple(K0, K1), reshaped_dims_idxs);
auto b_padded_unmerged_global_layout =
ck::wrapper::unmerge<1>(b_padded_global_layout, ck::make_tuple(K0, K1), reshaped_dims_idxs);
auto a_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<const DataType*>(p_a), a_padded_unmerged_global_layout);
auto b_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<const DataType*>(p_b), b_padded_unmerged_global_layout);
auto c_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
static_cast<DataType*>(p_c), c_padded_global_layout);
// Add extra M and N
constexpr auto a_tile_layout = ck::wrapper::make_layout(
ck::make_tuple(K0PerBlock, MPerBlock, K1),
ck::make_tuple((MPerBlock + ck::Number<1>{}) * K1, K1, ck::Number<1>{}));
constexpr auto b_tile_layout = ck::wrapper::make_layout(
ck::make_tuple(K0PerBlock, NPerBlock, K1),
ck::make_tuple((NPerBlock + ck::Number<1>{}) * K1, K1, ck::Number<1>{}));
__shared__ DataType lds_a[ck::wrapper::size(a_tile_layout) + NPerBlock];
__shared__ DataType lds_b[ck::wrapper::size(b_tile_layout) + NPerBlock];
auto a_lds_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
static_cast<DataType*>(lds_a), a_tile_layout);
auto b_lds_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
static_cast<DataType*>(lds_b), b_tile_layout);
const auto block_idxs = ck::make_tuple(ck::wrapper::slice(),
static_cast<ck::index_t>(blockIdx.x),
static_cast<ck::index_t>(blockIdx.y),
ck::wrapper::slice());
using DimAccessOrder = ck::Tuple<ck::Number<1>, ck::Number<0>, ck::Number<2>>;
constexpr ck::index_t vector_dim = 2;
auto c_global_local_tile =
ck::wrapper::make_local_tile(c_global_tensor,
tile_shape_k0_m_n_k1,
block_idxs,
make_tuple(ck::wrapper::slice(K0PerBlock),
ck::Number<1>{},
ck::Number<1>{},
ck::wrapper::slice(K1)));
auto c_global_local_partition =
ck::wrapper::make_blockwise_gemm_xdl_c_local_partition<DataType,
decltype(a_tile_layout),
decltype(b_tile_layout),
ck::wrapper::size(thread_layout),
GemmTraits>(c_global_local_tile);
auto c_vgpr_reg = ck::wrapper::make_blockwise_gemm_xdl_c_vgpr<DataType,
decltype(a_tile_layout),
decltype(b_tile_layout),
ck::wrapper::size(thread_layout),
GemmTraits>();
ck::wrapper::clear(c_vgpr_reg);
auto a_lds_tensor_local_partition =
ck::wrapper::make_local_partition(a_lds_tensor, thread_layout, threadIdx.x);
auto b_lds_tensor_local_partition =
ck::wrapper::make_local_partition(b_lds_tensor, thread_layout, threadIdx.x);
auto make_global_partition = [&](auto tensor, auto projection, ck::index_t i) {
const auto k_slice =
ck::make_tuple(ck::wrapper::slice(i * K0PerBlock, (i + 1) * K0PerBlock),
ck::wrapper::slice(),
ck::wrapper::slice());
auto local_tile = ck::wrapper::make_local_tile(
tensor(k_slice), tile_shape_k0_m_n_k1, block_idxs, projection);
return ck::wrapper::make_local_partition(local_tile, thread_layout, threadIdx.x);
};
auto a_global_local_partition = make_global_partition(
a_global_tensor,
make_tuple(ck::Number<1>{}, ck::Number<1>{}, ck::wrapper::slice(N), ck::Number<1>{}),
0);
auto b_global_local_partition = make_global_partition(
b_global_tensor,
make_tuple(ck::Number<1>{}, ck::wrapper::slice(M), ck::Number<1>{}, ck::Number<1>{}),
0);
// (row-major vgpr layout)
auto a_vgpr_tensor =
ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, DataType>(
ck::wrapper::make_layout(
shape(a_global_local_partition),
ck::make_tuple(ck::wrapper::size<1>(a_global_local_partition) *
ck::wrapper::size<2>(a_global_local_partition),
ck::wrapper::size<2>(a_global_local_partition),
ck::Number<1>{})));
auto b_vgpr_tensor =
ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, DataType>(
ck::wrapper::make_layout(
shape(b_global_local_partition),
ck::make_tuple(ck::wrapper::size<1>(a_global_local_partition) *
ck::wrapper::size<2>(a_global_local_partition),
ck::wrapper::size<2>(a_global_local_partition),
ck::Number<1>{})));
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(a_global_local_partition,
a_vgpr_tensor);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(b_global_local_partition,
b_vgpr_tensor);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(a_vgpr_tensor,
a_lds_tensor_local_partition);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(b_vgpr_tensor,
b_lds_tensor_local_partition);
const ck::index_t num_loop =
__builtin_amdgcn_readfirstlane(ck::math::integer_divide_ceil(K, KPerBlock));
if(num_loop > 1)
{
ck::index_t i = 0;
do
{
auto a_global_local_partition_i = make_global_partition(
a_global_tensor,
make_tuple(
ck::Number<1>{}, ck::Number<1>{}, ck::wrapper::slice(N), ck::Number<1>{}),
i + 1);
auto b_global_local_partition_i = make_global_partition(
b_global_tensor,
make_tuple(
ck::Number<1>{}, ck::wrapper::slice(M), ck::Number<1>{}, ck::Number<1>{}),
i + 1);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(
a_global_local_partition_i, a_vgpr_tensor);
ck::block_sync_lds();
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(
b_global_local_partition_i, b_vgpr_tensor);
ck::wrapper::blockwise_gemm_xdl<DataType, ck::wrapper::size(thread_layout), GemmTraits>(
a_lds_tensor, b_lds_tensor, c_vgpr_reg);
ck::block_sync_lds();
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(
a_vgpr_tensor, a_lds_tensor_local_partition);
ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(
b_vgpr_tensor, b_lds_tensor_local_partition);
++i;
} while(i < (num_loop - 1));
}
ck::block_sync_lds();
ck::wrapper::blockwise_gemm_xdl<DataType, ck::wrapper::size(thread_layout), GemmTraits>(
a_lds_tensor, b_lds_tensor, c_vgpr_reg);
ck::wrapper::copy(c_vgpr_reg, c_global_local_partition);
}
template <typename DataType,
typename GemmTraits,
ck::index_t scalar_per_vector,
bool DoPadding,
typename BlockShape,
typename ThreadLayout>
void PerformGemm(const ck::index_t M,
const ck::index_t N,
const ck::index_t K,
const BlockShape& tile_shape,
const ThreadLayout& thread_layout)
{
// Global memory buffers
DeviceMem a_mem(M * K * sizeof(DataType));
DeviceMem b_mem(K * N * sizeof(DataType));
DeviceMem c_mem(M * N * sizeof(DataType));
std::vector<DataType> a_data(M * K);
std::vector<DataType> b_data(K * N);
ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(a_data);
ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(b_data);
a_mem.ToDevice(a_data.data());
b_mem.ToDevice(b_data.data());
c_mem.SetZero();
const ck::index_t grid_size_x =
ck::math::integer_divide_ceil(M, ck::wrapper::size<0>(tile_shape));
const ck::index_t grid_size_y =
ck::math::integer_divide_ceil(N, ck::wrapper::size<1>(tile_shape));
const auto kernel =
DeviceGemm<DataType, GemmTraits, scalar_per_vector, BlockShape, ThreadLayout, DoPadding>;
const float avg_time = launch_and_time_kernel(StreamConfig{nullptr, true},
kernel,
dim3(grid_size_x, grid_size_y, 1),
dim3(ck::wrapper::size(thread_layout)),
0,
a_mem.GetDeviceBuffer(),
b_mem.GetDeviceBuffer(),
c_mem.GetDeviceBuffer(),
M,
N,
K,
tile_shape,
thread_layout);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(DataType) * M * K + sizeof(DataType) * K * N + sizeof(DataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << std::endl;
std::vector<DataType> c_data(M * N);
c_mem.FromDevice(c_data.data());
CheckResult<DataType>(a_data, b_data, c_data, M, N, K);
}
TEST(TestGemm, Float)
{
using DataType = float;
// (dim1, dim2, dim0 thread layout)
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<4>{}, ck::Number<64>{}, ck::Number<1>{}),
ck::make_tuple(ck::Number<1>{}, ck::Number<4>{}, ck::Number<1>{}));
const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<16>{});
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_4K1, 4, false>(
512, 512, 128, tile_shape, thread_layout);
// Irregular case
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_4K1, 1, true>(
129, 129, 67, tile_shape, thread_layout);
}
TEST(TestGemm, Int8)
{
using DataType = int8_t;
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<4>{}, ck::Number<64>{}, ck::Number<1>{}),
ck::make_tuple(ck::Number<1>{}, ck::Number<4>{}, ck::Number<1>{}));
const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<64>{});
PerformGemm<DataType,
ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_16K1,
16,
false>(512, 512, 128, tile_shape, thread_layout);
// Irregular case
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_16K1, 1, true>(
129, 129, 67, tile_shape, thread_layout);
}
TEST(TestGemm, Half)
{
using DataType = ck::half_t;
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<4>{}, ck::Number<64>{}, ck::Number<1>{}),
ck::make_tuple(ck::Number<1>{}, ck::Number<4>{}, ck::Number<1>{}));
const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<32>{});
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_8K1, 8, false>(
512, 512, 128, tile_shape, thread_layout);
// Irregular case
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_8K1, 1, true>(
129, 129, 67, tile_shape, thread_layout);
}
TEST(TestGemm, Float_2x4_4x2_XdlPerWave)
{
using DataType = float;
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<4>{}, ck::Number<64>{}, ck::Number<1>{}),
ck::make_tuple(ck::Number<1>{}, ck::Number<4>{}, ck::Number<1>{}));
const auto tile_shape = ck::make_tuple(ck::Number<256>{}, ck::Number<128>{}, ck::Number<16>{});
PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_4x2XdlPerWave_4K1, 4, false>(
512, 512, 128, tile_shape, thread_layout);
}

474
test/wrapper/test_wrapper_layout.cpp Normal file
View File

@@ -0,0 +1,474 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "ck/utility/common_header.hpp"
#include "ck/wrapper/layout.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
class TestWrapperLayout : public ::testing::Test
{
protected:
static constexpr auto I0 = ck::Number<0>{};
static constexpr auto I1 = ck::Number<1>{};
template <typename Desc,
typename Desc1d,
typename LayoutRuntime,
typename LayoutCompiletime,
typename Idxs>
void Run(Desc& desc,
Desc1d& desc_1d,
LayoutRuntime& layout_runtime,
LayoutCompiletime& layout_compiletime,
const std::vector<Idxs>& idxs)
{
// 1d check
EXPECT_EQ(desc_1d.GetLength(I0), ck::wrapper::size(layout_runtime));
// Check layout compiletime and runtime result consistency
EXPECT_EQ(ck::wrapper::size(layout_runtime), ck::wrapper::size(layout_compiletime));
for(ck::index_t i = 0; i < desc_1d.GetLength(I0); i++)
{
const ck::index_t layout_runtime_offset_1d = layout_runtime(ck::make_tuple(i));
const ck::index_t layout_compiletime_offset_1d = layout_compiletime(ck::make_tuple(i));
const ck::index_t desc_offset_1d = desc_1d.CalculateOffset(ck::make_tuple(i));
EXPECT_EQ(layout_runtime_offset_1d, desc_offset_1d);
EXPECT_EQ(layout_compiletime_offset_1d, layout_runtime_offset_1d);
}
// size(layout)-d check, don't check if access is hierarchical
if constexpr(!IsNestedTuple(Idxs{}))
{
ck::static_for<0, Idxs::Size(), 1>{}([&](auto d) {
EXPECT_EQ(desc.GetLength(ck::Number<d>{}), ck::wrapper::size<d>(layout_runtime));
EXPECT_EQ(ck::wrapper::size<d>(layout_runtime),
ck::wrapper::size<d>(layout_compiletime));
});
}
for(const auto idx : idxs)
{
const ck::index_t layout_runtime_offset = layout_runtime(idx);
const ck::index_t layout_compiletime_offset = layout_compiletime(idx);
const ck::index_t desc_offset =
desc.CalculateOffset(UnrollNestedTuple(idx)); // Unroll if nested
EXPECT_EQ(layout_runtime_offset, desc_offset);
EXPECT_EQ(layout_runtime_offset, layout_compiletime_offset);
}
}
};
TEST_F(TestWrapperLayout, 2d)
{
// dims:(4, 3) strides:(1, 4)
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
constexpr ck::index_t s1 = 1;
constexpr ck::index_t s0 = 4;
const auto desc =
ck::make_naive_tensor_descriptor(ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}),
ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
// Reverse due to column major
const auto desc_1d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1))),
ck::make_tuple(ck::Sequence<1, 0>{}),
ck::make_tuple(ck::Sequence<0>{}));
const auto layout_runtime = ck::wrapper::make_layout(ck::make_tuple(d1, d0));
const auto layout_compiletime =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}),
ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs;
for(ck::index_t h = 0; h < d1; h++)
{
for(ck::index_t w = 0; w < d0; w++)
{
idxs.emplace_back(h, w);
}
}
this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs);
}
TEST_F(TestWrapperLayout, 3d_nested)
{
// dims:((2, 3), 4, 3) strides:((2, 4), 12, 48)
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
constexpr ck::index_t s3 = 2;
constexpr ck::index_t s2 = 4;
constexpr ck::index_t s1 = 12;
constexpr ck::index_t s0 = 48;
const auto desc = ck::make_naive_tensor_descriptor(
ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}, ck::Number<d1>{}, ck::Number<d0>{}),
ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}, ck::Number<s1>{}, ck::Number<s0>{}));
// Reverse due to column major
const auto desc_1d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3))),
ck::make_tuple(ck::Sequence<3, 2, 1, 0>{}),
ck::make_tuple(ck::Sequence<0>{}));
const auto desc_3d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3)),
ck::make_pass_through_transform(d1),
ck::make_pass_through_transform(d2)),
ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}, ck::Sequence<3>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}, ck::Sequence<2>{}));
const auto layout_runtime =
ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(d3, d2), d1, d0),
ck::make_tuple(ck::make_tuple(s3, s2), s1, s0));
const auto layout_compiletime = ck::wrapper::make_layout(
ck::make_tuple(
ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}), ck::Number<d1>{}, ck::Number<d0>{}),
ck::make_tuple(ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}),
ck::Number<s1>{},
ck::Number<s0>{}));
std::vector<ck::Tuple<ck::index_t, ck::index_t, ck::index_t>> idxs_3d;
for(ck::index_t d = 0; d < d2 * d3; d++)
{
for(ck::index_t h = 0; h < d1; h++)
{
for(ck::index_t w = 0; w < d0; w++)
{
idxs_3d.emplace_back(d, h, w);
}
}
}
this->Run(desc_3d, desc_1d, layout_runtime, layout_compiletime, idxs_3d);
// Check also 4d iteration
std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t, ck::index_t>> idxs_4d;
for(ck::index_t e = 0; e < d3; e++)
{
for(ck::index_t d = 0; d < d2; d++)
{
for(ck::index_t h = 0; h < d1; h++)
{
for(ck::index_t w = 0; w < d0; w++)
{
idxs_4d.emplace_back(ck::make_tuple(e, d), h, w);
}
}
}
}
this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_4d);
}
TEST_F(TestWrapperLayout, 2d_nested)
{
// dims:((2, 3), (4, 3)) strides:((2, 4), (48, 12))
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
constexpr ck::index_t s3 = 2;
constexpr ck::index_t s2 = 4;
constexpr ck::index_t s1 = 48;
constexpr ck::index_t s0 = 12;
const auto desc = ck::make_naive_tensor_descriptor(
ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}, ck::Number<d1>{}, ck::Number<d0>{}),
ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}, ck::Number<s1>{}, ck::Number<s0>{}));
// Reverse due to column major
const auto desc_1d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3))),
ck::make_tuple(ck::Sequence<3, 2, 1, 0>{}),
ck::make_tuple(ck::Sequence<0>{}));
const auto desc_2d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3)),
ck::make_merge_transform(ck::make_tuple(d0, d1))),
ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<3, 2>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
const auto layout_runtime =
ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(d3, d2), ck::make_tuple(d1, d0)),
ck::make_tuple(ck::make_tuple(s3, s2), ck::make_tuple(s1, s0)));
const auto layout_compiletime = ck::wrapper::make_layout(
ck::make_tuple(ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})),
ck::make_tuple(ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}),
ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{})));
std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs_2d;
for(ck::index_t h = 0; h < d2 * d3; h++)
{
for(ck::index_t w = 0; w < d0 * d1; w++)
{
idxs_2d.emplace_back(h, w);
}
}
this->Run(desc_2d, desc_1d, layout_runtime, layout_compiletime, idxs_2d);
// Check also 4d iteration
std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::Tuple<ck::index_t, ck::index_t>>>
idxs_4d;
for(ck::index_t e = 0; e < d3; e++)
{
for(ck::index_t d = 0; d < d2; d++)
{
for(ck::index_t h = 0; h < d1; h++)
{
for(ck::index_t w = 0; w < d0; w++)
{
idxs_4d.emplace_back(ck::make_tuple(e, d), ck::make_tuple(h, w));
}
}
}
}
this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_4d);
}
TEST_F(TestWrapperLayout, 3d_double_nested)
{
// dims:(((2, 2), 3), (4, 3)) strides:(((2, 4), 8), (96, 24))
constexpr ck::index_t d4 = 2;
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
constexpr ck::index_t s4 = 2;
constexpr ck::index_t s3 = 4;
constexpr ck::index_t s2 = 8;
constexpr ck::index_t s1 = 96;
constexpr ck::index_t s0 = 24;
const auto desc = ck::make_naive_tensor_descriptor(ck::make_tuple(ck::Number<d4>{},
ck::Number<d3>{},
ck::Number<d2>{},
ck::Number<d1>{},
ck::Number<d0>{}),
ck::make_tuple(ck::Number<s4>{},
ck::Number<s3>{},
ck::Number<s2>{},
ck::Number<s1>{},
ck::Number<s0>{}));
// Reverse due to column major
const auto desc_1d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3, d4))),
ck::make_tuple(ck::Sequence<4, 3, 2, 1, 0>{}),
ck::make_tuple(ck::Sequence<0>{}));
const auto desc_3d = transform_tensor_descriptor(
desc,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d3, d4)),
ck::make_pass_through_transform(d2),
ck::make_merge_transform(ck::make_tuple(d0, d1))),
ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}, ck::Sequence<4, 3>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}, ck::Sequence<2>{}));
const auto desc_2d = transform_tensor_descriptor(
desc_3d,
ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3 * d4)),
ck::make_pass_through_transform(d1 * d0)),
ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}),
ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
const auto layout_runtime = ck::wrapper::make_layout(
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)),
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, s3), s2), ck::make_tuple(s1, s0)));
const auto layout_compiletime = ck::wrapper::make_layout(
ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})),
ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<s3>{}), ck::Number<s2>{}),
ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{})));
std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs_2d;
for(ck::index_t h = 0; h < d2 * d3 * d4; h++)
{
for(ck::index_t w = 0; w < d0 * d1; w++)
{
idxs_2d.emplace_back(h, w);
}
}
this->Run(desc_2d, desc_1d, layout_runtime, layout_compiletime, idxs_2d);
// Check also 3d iteration
std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t>> idxs_3d;
for(ck::index_t d = 0; d < d3 * d4; d++)
{
for(ck::index_t h = 0; h < d2; h++)
{
for(ck::index_t w = 0; w < d1 * d0; w++)
{
idxs_3d.emplace_back(ck::make_tuple(d, h), w);
}
}
}
this->Run(desc_3d, desc_1d, layout_runtime, layout_compiletime, idxs_3d);
// Check also 5d iteration
std::vector<ck::Tuple<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t>,
ck::Tuple<ck::index_t, ck::index_t>>>
idxs_5d;
for(ck::index_t f = 0; f < d4; f++)
{
for(ck::index_t e = 0; e < d3; e++)
{
for(ck::index_t d = 0; d < d2; d++)
{
for(ck::index_t h = 0; h < d1; h++)
{
for(ck::index_t w = 0; w < d0; w++)
{
idxs_5d.emplace_back(ck::make_tuple(ck::make_tuple(f, e), d),
ck::make_tuple(h, w));
}
}
}
}
}
this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_5d);
}
TEST(TestLayoutHelpers, SizeAndGet)
{
// dims:(((2, 2), 3), (4, 3))
constexpr ck::index_t d4 = 2;
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
const auto layout_runtime = ck::wrapper::make_layout(
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)));
const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
// Size of layout
EXPECT_EQ(ck::wrapper::size(layout_runtime), d4 * d3 * d2 * d1 * d0);
EXPECT_EQ(ck::wrapper::size(layout_compiletime), d4 * d3 * d2 * d1 * d0);
// Size of dims
EXPECT_EQ(ck::wrapper::size<0>(layout_runtime), d4 * d3 * d2);
EXPECT_EQ(ck::wrapper::size<0>(layout_compiletime), d4 * d3 * d2);
EXPECT_EQ(ck::wrapper::size<1>(layout_runtime), d1 * d0);
EXPECT_EQ(ck::wrapper::size<1>(layout_compiletime), d1 * d0);
// Access through new layout (using get with layout object)
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(layout_runtime)), d4 * d3);
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(layout_compiletime)), d4 * d3);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_runtime)), d2);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_compiletime)), d2);
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_runtime))), d4);
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_compiletime))),
d4);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_runtime))), d3);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_compiletime))),
d3);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_runtime)), d2);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_compiletime)), d2);
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<1>(layout_runtime)), d1);
EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<1>(layout_compiletime)), d1);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<1>(layout_runtime)), d0);
EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<1>(layout_compiletime)), d0);
}
TEST(TestLayoutHelpers, DepthAndRank)
{
// dims:(((2, 2), 3), (4, 3))
constexpr ck::index_t d4 = 2;
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
const auto layout_runtime = ck::wrapper::make_layout(
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)));
const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
EXPECT_EQ(ck::wrapper::depth(layout_runtime), 3);
EXPECT_EQ(ck::wrapper::depth(layout_compiletime), 3);
EXPECT_EQ(ck::wrapper::depth(ck::make_tuple(ck::make_tuple(d4, d3), d2)), 2);
// Check for integer
EXPECT_EQ(ck::wrapper::depth(d0), 0);
EXPECT_EQ(ck::wrapper::rank(layout_runtime), 2);
EXPECT_EQ(ck::wrapper::rank(layout_compiletime), 2);
EXPECT_EQ(ck::wrapper::rank(ck::make_tuple(ck::make_tuple(d4, d3), d2)), 2);
// Check for integer
EXPECT_EQ(ck::wrapper::rank(d0), 1);
}
TEST(TestLayoutHelpers, ShapeAndStrides)
{
// dims:(((2, 2), 3), (4, 3))
constexpr ck::index_t d4 = 2;
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
constexpr ck::index_t s4 = 2;
constexpr ck::index_t s3 = 4;
constexpr ck::index_t s2 = 8;
constexpr ck::index_t s1 = 96;
constexpr ck::index_t s0 = 24;
const auto shape_compiletime = ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}));
const auto strides_compiletime = ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<s4>{}, ck::Number<s3>{}), ck::Number<s2>{}),
ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
const auto shape_runtime =
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0));
const auto strides_runtime =
ck::make_tuple(ck::make_tuple(ck::make_tuple(s4, s3), s2), ck::make_tuple(s1, s0));
const auto layout_runtime = ck::wrapper::make_layout(shape_runtime, strides_runtime);
const auto layout_compiletime =
ck::wrapper::make_layout(shape_compiletime, strides_compiletime);
constexpr bool check_compiletime_shape =
std::is_same_v<decltype(shape_compiletime),
std::remove_reference_t<decltype(shape(layout_compiletime))>>;
constexpr bool check_runtime_shape =
std::is_same_v<decltype(shape_runtime),
std::remove_reference_t<decltype(shape(layout_runtime))>>;
EXPECT_TRUE(check_compiletime_shape);
EXPECT_TRUE(check_runtime_shape);
}
TEST(TestLayoutHelpers, Hierarchical)
{
// dims:(((2, 2), 3), (4, 3))
constexpr ck::index_t d4 = 2;
constexpr ck::index_t d3 = 2;
constexpr ck::index_t d2 = 3;
constexpr ck::index_t d1 = 4;
constexpr ck::index_t d0 = 3;
const auto runtime_shape =
ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0));
const auto layout_runtime = ck::wrapper::make_layout(runtime_shape);
const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
EXPECT_EQ((ck::wrapper::rank<0, 0>(runtime_shape)), 2);
EXPECT_EQ((ck::wrapper::rank<0, 0>(layout_runtime)), 2);
EXPECT_EQ((ck::wrapper::rank<0, 0>(layout_compiletime)), 2);
EXPECT_EQ((ck::wrapper::depth<0, 0>(runtime_shape)), 1);
EXPECT_EQ((ck::wrapper::depth<0, 0>(layout_runtime)), 1);
EXPECT_EQ((ck::wrapper::depth<0, 0>(layout_compiletime)), 1);
EXPECT_EQ((ck::wrapper::size<0, 0>(runtime_shape)), d4 * d3);
EXPECT_EQ((ck::wrapper::size<0, 0>(layout_runtime)), d4 * d3);
EXPECT_EQ((ck::wrapper::size<0, 0>(layout_compiletime)), d4 * d3);
EXPECT_EQ((ck::wrapper::get<0, 0, 0>(runtime_shape)), d4);
}

115
test/wrapper/test_wrapper_partition.cpp Normal file
View File

@@ -0,0 +1,115 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <numeric>
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/common_header.hpp"
#include "ck/wrapper/layout.hpp"
#include "ck/wrapper/tensor.hpp"
TEST(TestPartition, LocalPartition)
{
const auto shape =
ck::make_tuple(ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}), ck::Number<4>{});
const auto strides =
ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}), ck::Number<64>{});
const auto layout = ck::wrapper::make_layout(shape, strides);
std::vector<ck::index_t> data(ck::wrapper::size(layout));
std::iota(data.begin(), data.end(), 0);
const auto tensor =
ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
const auto thread_steps = ck::make_tuple(ck::Number<1>{}, ck::Number<8>{}, ck::Number<1>{});
// row-major thread layout
const auto thread_layout =
ck::wrapper::make_layout(ck::make_tuple(ck::Number<4>{}, ck::Number<8>{}, ck::Number<1>{}),
ck::make_tuple(ck::Number<8>{}, ck::Number<1>{}, ck::Number<1>{}));
// 3d partition on 2d shape (calculate partition on 3d thread layout, and then skip first dim)
const auto thread_projection =
ck::make_tuple(ck::wrapper::slice(4), ck::Number<1>{}, ck::Number<1>{});
constexpr ck::index_t projection_thread_length = ck::Number<4>{};
for(ck::index_t thread_id = 0;
thread_id < ck::wrapper::size(thread_layout) / projection_thread_length;
thread_id++)
{
const auto packed_partition =
ck::wrapper::make_local_partition(tensor, thread_layout, thread_id, thread_projection);
const auto expected_partition_size =
ck::wrapper::size(tensor) /
(ck::wrapper::size(thread_layout) / projection_thread_length);
const auto expected_partition_first_val = thread_id * ck::wrapper::size<1>(thread_steps);
const auto expected_partition_second_val = expected_partition_first_val + 1;
EXPECT_EQ(ck::wrapper::size(packed_partition), expected_partition_size);
EXPECT_EQ(packed_partition(0), expected_partition_first_val);
EXPECT_EQ(packed_partition(1), expected_partition_second_val);
}
}
TEST(TestPartition, LocalTile)
{
const auto shape = ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}, ck::Number<4>{});
const auto strides = ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}, ck::Number<64>{});
const auto layout = ck::wrapper::make_layout(shape, strides);
std::vector<ck::index_t> data(ck::wrapper::size(layout));
std::iota(data.begin(), data.end(), 0);
const auto tensor =
ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
// 4d tile partitioning on 3d shape (calculate tile on 4d tile layout, and then skip last dim)
const auto block_shape =
ck::make_tuple(ck::Number<2>{}, ck::Number<4>{}, ck::Number<2>{}, ck::Number<2>{});
const auto block_projection =
ck::make_tuple(ck::Number<1>{}, ck::Number<1>{}, ck::Number<1>{}, ck::wrapper::slice(2));
const auto grid_shape =
ck::make_tuple(ck::wrapper::size<0>(shape) / ck::wrapper::size<0>(block_shape),
ck::wrapper::size<1>(shape) / ck::wrapper::size<1>(block_shape),
ck::wrapper::size<2>(shape) / ck::wrapper::size<2>(block_shape));
std::vector<ck::Tuple<ck::index_t, ck::index_t, ck::index_t, ck::index_t>> block_idxs;
for(int i = 0; i < ck::wrapper::size<0>(grid_shape); i++)
{
for(int j = 0; j < ck::wrapper::size<1>(grid_shape); j++)
{
for(int k = 0; k < ck::wrapper::size<2>(grid_shape); k++)
{
block_idxs.emplace_back(i, j, k, 0);
}
}
}
for(auto block_idx : block_idxs)
{
constexpr ck::index_t projection_block_dim = ck::Number<2>{};
const auto packed_tile =
ck::wrapper::make_local_tile(tensor, block_shape, block_idx, block_projection);
const auto expected_tile_size = ck::wrapper::size(block_shape) / projection_block_dim;
auto expected_tile_first_val = ck::wrapper::size<2>(block_idx) *
ck::wrapper::size<2>(block_shape) *
ck::wrapper::size<2>(strides);
expected_tile_first_val += ck::wrapper::size<1>(block_idx) *
ck::wrapper::size<1>(block_shape) *
ck::wrapper::size<1>(strides);
expected_tile_first_val += ck::wrapper::size<0>(block_idx) *
ck::wrapper::size<0>(block_shape) *
ck::wrapper::size<0>(strides);
const auto expected_tile_second_val = expected_tile_first_val + 1;
EXPECT_EQ(ck::wrapper::size(packed_tile), expected_tile_size);
EXPECT_EQ(packed_tile(0), expected_tile_first_val);
EXPECT_EQ(packed_tile(1), expected_tile_second_val);
}
}

209
test/wrapper/test_wrapper_tensor.cpp Normal file
View File

@@ -0,0 +1,209 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "ck/library/utility/device_memory.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/utility/common_header.hpp"
#include "ck/wrapper/layout.hpp"
#include "ck/wrapper/tensor.hpp"
// Compare data in tensor with offset from layout.
// Data and offset should match if physical memory has been initialized with
// sequentially increasing values from 0.
template <typename TensorType>
__host__ __device__ bool TestTensorCheck3d(TensorType& tensor)
{
const auto& layout = ck::wrapper::layout(tensor);
for(ck::index_t d = 0; d < ck::wrapper::size<0>(ck::wrapper::get<0>(layout)); d++)
{
for(ck::index_t h = 0; h < ck::wrapper::size<1>(ck::wrapper::get<0>(layout)); h++)
{
for(ck::index_t w = 0; w < ck::wrapper::size<1>(layout); w++)
{
const auto idx = ck::make_tuple(ck::make_tuple(d, h), w);
if(tensor(idx) != layout(idx))
{
return false;
}
}
}
}
return true;
}
template <typename TensorType>
__host__ __device__ bool TestTensorCheck1d(TensorType& tensor, ck::index_t start_offset = 0)
{
const auto& layout = ck::wrapper::layout(tensor);
for(ck::index_t w = 0; w < ck::wrapper::size<0>(layout); w++)
{
if(tensor(w) - start_offset != layout(ck::make_tuple(w)))
{
return false;
}
}
return true;
}
template <ck::index_t nelems, typename TensorType>
__host__ __device__ bool StaticTestTensorCheck1d(TensorType& tensor)
{
const auto& layout = ck::wrapper::layout(tensor);
bool success = true;
ck::static_for<0, nelems, 1>{}([&](auto w) {
if(tensor(ck::Number<w.value>{}) != layout(ck::make_tuple(w.value)))
{
success = false;
}
});
return success;
}
template <typename TensorType>
__host__ __device__ void InitTensor(TensorType& tensor)
{
for(ck::index_t i = 0; i < ck::wrapper::size(ck::wrapper::layout(tensor)); i++)
{
tensor(i) = i;
}
}
template <ck::index_t nelems, typename TensorType>
__host__ __device__ void StaticInitTensor(TensorType& tensor)
{
ck::static_for<0, nelems, 1>{}([&](auto i) { tensor(ck::Number<i.value>{}) = i.value; });
}
// Tests
TEST(TestTensor, ReadWriteHostMemory)
{
constexpr ck::index_t nelems = 8;
std::array<ck::index_t, nelems> data;
const auto layout = ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(2, 2), 2));
auto tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(&data[0], layout);
InitTensor(tensor);
EXPECT_TRUE(TestTensorCheck1d(tensor));
EXPECT_TRUE(TestTensorCheck3d(tensor));
}
__global__ void TestTensorReadWriteDevice(void* data, void* success)
{
constexpr ck::index_t nelems = 8;
__shared__ ck::index_t p_shared[nelems];
ck::index_t* casted_data_ptr = static_cast<ck::index_t*>(data);
bool* casted_success_ptr = static_cast<bool*>(success);
const auto layout = ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(2, 2), 2));
constexpr auto vgpr_layout =
ck::wrapper::make_layout(make_tuple(ck::Number<nelems>{}), make_tuple(ck::Number<1>{}));
auto tensor_global =
ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(casted_data_ptr, layout);
auto tensor_lds = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(p_shared, layout);
auto tensor_vgpr =
ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, ck::index_t>(
vgpr_layout);
InitTensor(tensor_global);
InitTensor(tensor_lds);
StaticInitTensor<nelems>(tensor_vgpr);
*casted_success_ptr = TestTensorCheck1d(tensor_global);
*casted_success_ptr &= TestTensorCheck3d(tensor_global);
*casted_success_ptr &= TestTensorCheck1d(tensor_lds);
*casted_success_ptr &= TestTensorCheck3d(tensor_lds);
*casted_success_ptr &= StaticTestTensorCheck1d<nelems>(tensor_vgpr);
}
TEST(TestTensor, ReadWriteGlobalLdsRegistersMemory)
{
constexpr ck::index_t nelems = 8;
std::array<ck::index_t, nelems> host_data;
DeviceMem data_buf(nelems * sizeof(ck::index_t));
data_buf.ToDevice(&host_data[0]);
DeviceMem success_buf(sizeof(bool));
launch_and_time_kernel(StreamConfig{},
TestTensorReadWriteDevice,
dim3(1),
dim3(1),
0,
data_buf.GetDeviceBuffer(),
success_buf.GetDeviceBuffer());
bool success;
success_buf.FromDevice(&success);
EXPECT_TRUE(success);
}
TEST(TestTensor, Slicing)
{
constexpr ck::index_t nelems = 8;
std::array<ck::index_t, nelems> data;
const auto shape = ck::make_tuple(ck::make_tuple(2, 2), 2);
const auto strides = ck::make_tuple(ck::make_tuple(1, 2), 4);
const auto layout = ck::wrapper::make_layout(shape, strides);
auto tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(&data[0], layout);
InitTensor(tensor);
auto tensor2x2x2 =
tensor(ck::make_tuple(ck::wrapper::slice(2), ck::wrapper::slice(2)), ck::wrapper::slice(2));
EXPECT_EQ(tensor2x2x2(0), layout(ck::make_tuple(ck::make_tuple(0, 0), 0)));
EXPECT_EQ(ck::wrapper::rank(tensor2x2x2), 2);
EXPECT_EQ(ck::wrapper::depth(tensor2x2x2), 2);
EXPECT_EQ(ck::wrapper::size(tensor2x2x2), 8);
EXPECT_TRUE(TestTensorCheck1d(tensor2x2x2));
auto tensor2x2 = tensor(ck::make_tuple(1, ck::wrapper::slice(2)), ck::wrapper::slice(2));
EXPECT_EQ(tensor2x2(0), layout(ck::make_tuple(ck::make_tuple(1, 0), 0)));
EXPECT_EQ(ck::wrapper::rank(tensor2x2), 2);
EXPECT_EQ(ck::wrapper::depth(tensor2x2), 2);
EXPECT_EQ(ck::wrapper::size(tensor2x2), 4);
EXPECT_TRUE(TestTensorCheck1d(tensor2x2));
auto tensor1x1 = tensor(ck::make_tuple(1, ck::wrapper::slice(1, 2)), ck::wrapper::slice(1, 2));
EXPECT_EQ(tensor1x1(0), layout(ck::make_tuple(ck::make_tuple(1, 1), 1)));
EXPECT_EQ(rank(tensor1x1), 2);
EXPECT_EQ(depth(tensor1x1), 2);
EXPECT_EQ(size(tensor1x1), 1);
EXPECT_TRUE(TestTensorCheck1d(tensor1x1));
auto tensor2 = tensor(ck::make_tuple(1, 1), ck::wrapper::slice(0, 2));
EXPECT_EQ(tensor2(0), layout(ck::make_tuple(ck::make_tuple(1, 1), 0)));
EXPECT_EQ(ck::wrapper::rank(tensor2), 1);
EXPECT_EQ(ck::wrapper::depth(tensor2), 1);
EXPECT_EQ(ck::wrapper::size(tensor2), 2);
EXPECT_TRUE(TestTensorCheck1d(tensor2));
auto tensor2_v2 = tensor(2, ck::wrapper::slice(0, 2));
EXPECT_EQ(tensor2_v2(0), layout(ck::make_tuple(2, 0)));
EXPECT_EQ(ck::wrapper::rank(tensor2_v2), 1);
EXPECT_EQ(ck::wrapper::depth(tensor2_v2), 1);
EXPECT_EQ(ck::wrapper::size(tensor2_v2), 2);
EXPECT_TRUE(TestTensorCheck1d(tensor2_v2));
// negative indexing
auto tensor1x2 = tensor(ck::make_tuple(1, ck::wrapper::slice(0, -2)), ck::wrapper::slice());
EXPECT_EQ(tensor1x2(0), layout(ck::make_tuple(ck::make_tuple(1, 0), 0)));
EXPECT_EQ(rank(tensor1x2), 2);
EXPECT_EQ(depth(tensor1x2), 2);
EXPECT_EQ(size(tensor1x2), 2);
EXPECT_TRUE(TestTensorCheck1d(tensor1x2));
}