Skip to content

GitLab

Commit 05ee41c3 authored by Rosty Geyyer's avatar Rosty Geyyer
Browse files

Merge branch 'develop' into lwpck-471

parents 37116c98 ad541ad6
Hide whitespace changes
Inline Side-by-side
Showing with 1177 additions and 302 deletions
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
View file @ 05ee41c3
......@@ -23,6 +23,7 @@ Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
View file @ 05ee41c3
......@@ -23,6 +23,7 @@ Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
View file @ 05ee41c3
......@@ -27,6 +27,7 @@ Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
View file @ 05ee41c3
......@@ -23,6 +23,7 @@ Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......
example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
View file @ 05ee41c3
......@@ -106,15 +106,15 @@ bool run_batched_gemm_gemm_example(int argc, char* argv[])
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
}
};
......@@ -270,7 +270,7 @@ bool run_batched_gemm_gemm_example(int argc, char* argv[])
c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
#endif
return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData);
return ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
}
return true;
......
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ 05ee41c3
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_custom_target(example_gemm_scale_softmax_gemm)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 05ee41c3
......@@ -24,6 +24,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
......
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp 0 → 100644
View file @ 05ee41c3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using B0DataType = BF16;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = BF16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: bf16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, bf16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: bf16 in, bf16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 05ee41c3
......@@ -24,6 +24,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
......
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp 0 → 100644
View file @ 05ee41c3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using B0DataType = BF16;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = BF16;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>;
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_batched_gemm_scale_softmax_gemm.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
View file @ 05ee41c3
......@@ -23,6 +23,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
......@@ -138,261 +139,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
#include "run_batched_gemm_scale_softmax_gemm.inc"
// GEMM shape
ck::index_t M = 1020;
ck::index_t N = 1020;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t BatchCount = 4;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t StrideC = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
ck::index_t BatchStrideC = -1;
float alpha = 1;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
}
else if(argc == 18)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
StrideA = std::stoi(argv[9]);
StrideB0 = std::stoi(argv[10]);
StrideB1 = std::stoi(argv[11]);
StrideC = std::stoi(argv[12]);
BatchStrideA = std::stoi(argv[13]);
BatchStrideB0 = std::stoi(argv[14]);
BatchStrideB1 = std::stoi(argv[15]);
BatchStrideC = std::stoi(argv[16]);
alpha = std::stof(argv[17]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
"BatchStrideB0, BatchStrideB1, BatchStrideC\n");
printf("arg17: scale (alpha)\n");
exit(0);
}
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? O : M;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
StrideC = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
BatchStrideC = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_g_m_o_host_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_o_device_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
c_g_m_o_device_result.mDesc.GetElementSpaceSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
StrideA,
StrideB0,
StrideB1,
StrideC,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
BatchStrideC,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
if(do_verification)
{
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
}
return 0;
}
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ 05ee41c3
......@@ -24,6 +24,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
......
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc 0 → 100644
View file @ 05ee41c3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 2;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1020;
ck::index_t N = 1020;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t BatchCount = 4;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t StrideC = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
ck::index_t BatchStrideC = -1;
float alpha = 1;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
}
else if(argc == 18)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
StrideA = std::stoi(argv[9]);
StrideB0 = std::stoi(argv[10]);
StrideB1 = std::stoi(argv[11]);
StrideC = std::stoi(argv[12]);
BatchStrideA = std::stoi(argv[13]);
BatchStrideB0 = std::stoi(argv[14]);
BatchStrideB1 = std::stoi(argv[15]);
BatchStrideC = std::stoi(argv[16]);
alpha = std::stof(argv[17]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
"BatchStrideB0, BatchStrideB1, BatchStrideC\n");
printf("arg17: scale (alpha)\n");
exit(0);
}
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? O : M;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
StrideC = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
BatchStrideC = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_g_m_o_host_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_o_device_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
c_g_m_o_device_result.mDesc.GetElementSpaceSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
StrideA,
StrideB0,
StrideB1,
StrideC,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
BatchStrideC,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
if(do_verification)
{
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
}
return 0;
}
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
View file @ 05ee41c3
......@@ -22,7 +22,7 @@ int run(int argc, char* argv[])
float alpha = 1;
bool input_permute = false;
bool input_permute = false;
bool output_permute = true;
if(argc == 1)
......@@ -50,7 +50,7 @@ int run(int argc, char* argv[])
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
else
......@@ -253,7 +253,23 @@ int run(int argc, char* argv[])
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
return ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol)
? 0
: 1;
}
......
example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
View file @ 05ee41c3
......@@ -7,7 +7,7 @@ int run(int argc, char* argv[])
int init_method = 1;
bool time_kernel = false;
bool input_permute = false;
bool input_permute = false;
bool output_permute = true;
if(argc == 1)
......@@ -26,7 +26,7 @@ int run(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
input_permute = std::stoi(argv[4]);
input_permute = std::stoi(argv[4]);
output_permute = std::stoi(argv[5]);
}
else
......@@ -66,10 +66,10 @@ int run(int argc, char* argv[])
std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
int M = 128 * (rand() % 8 + 1);
int N = 128 * (rand() % 8 + 1);
int K = 40;
int O = 40 * (rand() % 2 + 1);
int M = 128 * (rand() % 8 + 1);
int N = 128 * (rand() % 8 + 1);
int K = 40;
int O = 40 * (rand() % 2 + 1);
int G0 = rand() % 3 + 1;
int G1 = rand() % 5 + 1;
......@@ -228,12 +228,12 @@ int run(int argc, char* argv[])
{
for(std::size_t i = 0; i < group_count; i++)
{
const int& G0 = g0_g1_m_n_k_o[i][0];
const int& G1 = g0_g1_m_n_k_o[i][1];
const int& M = g0_g1_m_n_k_o[i][2];
const int& N = g0_g1_m_n_k_o[i][3];
const int& K = g0_g1_m_n_k_o[i][4];
const int& O = g0_g1_m_n_k_o[i][5];
const int& G0 = g0_g1_m_n_k_o[i][0];
const int& G1 = g0_g1_m_n_k_o[i][1];
const int& M = g0_g1_m_n_k_o[i][2];
const int& N = g0_g1_m_n_k_o[i][3];
const int& K = g0_g1_m_n_k_o[i][4];
const int& O = g0_g1_m_n_k_o[i][5];
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
......
example/33_multiple_reduce/dual_reduce_common.hpp
View file @ 05ee41c3
......@@ -12,6 +12,7 @@
#include "ck/utility/reduction_enums.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -253,10 +254,10 @@ int mean_meansquare_dual_reduce_test(size_t n,
std::array<ck::index_t, NumOutputDim> i_outLengths;
std::array<ck::index_t, NumOutputDim> i_outStrides;
std::copy(inLengths.begin(), inLengths.end(), i_inLengths.begin());
std::copy(inStrides.begin(), inStrides.end(), i_inStrides.begin());
std::copy(outLengths.begin(), outLengths.end(), i_outLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), i_outStrides.begin());
ck::ranges::copy(inLengths, i_inLengths.begin());
ck::ranges::copy(inStrides, i_inStrides.begin());
ck::ranges::copy(outLengths, i_outLengths.begin());
ck::ranges::copy(outStrides, i_outStrides.begin());
auto dual_reduce_op = DeviceDualReduce{};
......@@ -305,8 +306,8 @@ int mean_meansquare_dual_reduce_test(size_t n,
{
mean_dev.FromDevice(mean.mData.data());
meansquare_dev.FromDevice(meansquare.mData.data());
pass = pass && ck::utils::check_err(mean.mData, mean_ref.mData);
pass = pass && ck::utils::check_err(meansquare.mData, meansquare_ref.mData);
pass = pass && ck::utils::check_err(mean, mean_ref);
pass = pass && ck::utils::check_err(meansquare, meansquare_ref);
};
return (pass ? 0 : 1);
......
example/34_batchnorm/CMakeLists.txt
View file @ 05ee41c3
add_example_executable(example_batchnorm_forward batchnorm_forward_nhwc.cpp)
add_example_executable(example_batchnorm_infer batchnorm_infer_nhwc.cpp)
add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)
example/34_batchnorm/README.md
View file @ 05ee41c3
......@@ -53,4 +53,29 @@ Start running 10 times...
Perf: 1.28235 ms, 523.329 GB/s
```
## Run ```batchnorm backward nhwc```
```bash
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
Arg4 -- time kernel (0=no, 1=yes)
Arg5: use multi-block welford (0=n0, 1=yes)
./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
```
Result
```
./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 0.411026 ms, 91.8702 GB/s
```
example/34_batchnorm/batchnorm_backward_nhwc.cpp 0 → 100644
View file @ 05ee41c3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormBwdArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
bool haveSavedMeanInvVar;
int data_type = 0;
int init_method = 3;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
{
// clang-format off
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
// clang-format on
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 5 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
haveSavedMeanInvVar = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
return (0);
};
};
using namespace ck;
template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_bwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
bool haveSavedMeanInvVar,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr index_t Rank = 4;
constexpr index_t NumReduceDim = 3;
using ScaleDataType = XDataType;
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm backward algorithm
Tensor<XDataType> x(inOutLengths);
Tensor<AccDataType> dy(inOutLengths);
Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
// savedVariance is only used for initializing savedInvVar
Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
// output data of the batchnorm backward algorithm
Tensor<AccDataType> dx_ref(inOutLengths);
Tensor<AccDataType> dx(inOutLengths);
Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
auto inOutStrides = dy.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if(haveSavedMeanInvVar)
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.0001f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
// initialize the savedMean to be values with tiny variation to the mean of the x values
savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
num_thread);
// initialize the variance to be values with tiny variation to the variance of the x values
savedVariance.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
auto it_src = savedVariance.mData.begin();
auto it_dst = savedInvVar.mData.begin();
float tmp_epsilon = std::numeric_limits<float>::epsilon();
while(it_src != savedVariance.mData.end())
{
*it_dst = type_convert<AccDataType>(
1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
it_src++;
it_dst++;
};
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
break;
case 2:
dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
}
};
// input data of the batchnorm backward algorithm
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
// output data of the batchnorm backward algorithm
DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
dy_dev.ToDevice(dy.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
if(haveSavedMeanInvVar)
{
savedMean_dev.ToDevice(savedMean.mData.data());
savedInvVar_dev.ToDevice(savedInvVar.mData.data());
};
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
std::copy(scaleBiasMeanVarLengths.begin(),
scaleBiasMeanVarLengths.end(),
i_scaleBiasMeanVarLengths.begin());
std::copy(scaleBiasMeanVarStrides.begin(),
scaleBiasMeanVarStrides.end(),
i_scaleBiasMeanVarStrides.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormBwdInstance =
ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
AccDataType,
AccDataType,
AccDataType,
ScaleDataType, // ScaleDataType
AccDataType, // DscaleDbiasDataType
AccDataType, // MeanVarDataType
PassThroughOp,
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1, // XSrcVectorSize
1, // DySrcVectorSize
1, // DxDstVectorSize
1, // ScaleSrcVectorSize
1, // DscaleDbiasDstVectorSize
1>; // MeanVarSrcVectorSize
auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
dy_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
epsilon,
PassThroughOp{},
dx_dev.GetDeviceBuffer(),
dscale_dev.GetDeviceBuffer(),
dbias_dev.GetDeviceBuffer());
if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
"exiting!"
<< std::endl;
return (false);
};
size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
if(time_kernel)
{
float avg_time = 0.0f;
size_t num_bytes = 0;
size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
size_t invariant_length = inOutLengths[3];
avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// inputing of x, dy, scale, outputing of dx, dscale, dbias
num_bytes +=
total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
// outputing of mean, inv-variance
num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
}
else
(void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
if(do_verification)
{
using ReferenceBatchNormBwdInstance =
ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
AccDataType,
AccDataType,
AccDataType,
ScaleDataType, // ScaleDataType
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
dy.mData.data(),
bnScale.mData.data(),
haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
epsilon,
PassThroughOp{},
dx_ref.mData.data(),
dscale_ref.mData.data(),
dbias_ref.mData.data());
if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout
<< "The runtime parameters seems not supported by the device instance, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
dx_dev.FromDevice(dx.mData.data());
dscale_dev.FromDevice(dscale.data());
dbias_dev.FromDevice(dbias.data());
// clang-format off
pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
// clang-format on
};
return (pass);
};
static const double epsilon = std::numeric_limits<float>::epsilon();
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormBwdArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 1)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 5)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 6)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
}
else
{
pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
3,
false, // don't time kernel
{128, 16, 6, 512},
false,
epsilon);
pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
3,
false, // don't time kernel
{128, 16, 3, 1024},
false,
epsilon);
};
return (pass ? 0 : 1);
}
example/34_batchnorm/batchnorm_infer_nhwc.cpp → example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
View file @ 05ee41c3
......@@ -9,12 +9,14 @@
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer_nhwc_c.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
#include "batchnorm_infer_impl.hpp"
......@@ -123,6 +125,8 @@ bool bnorm_infer_nhwc_test(bool do_verification,
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
// when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
// eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm forward algorithm
......@@ -220,14 +224,10 @@ bool bnorm_infer_nhwc_test(bool do_verification,
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
std::copy(scaleBiasMeanVarLengths.begin(),
scaleBiasMeanVarLengths.end(),
i_scaleBiasMeanVarLengths.begin());
std::copy(scaleBiasMeanVarStrides.begin(),
scaleBiasMeanVarStrides.end(),
i_scaleBiasMeanVarStrides.begin());
ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
int result = 0;
......@@ -263,20 +263,25 @@ bool bnorm_infer_nhwc_test(bool do_verification,
if(do_verification)
{
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<
InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType>;
ck::tensor_operation::host::ReferenceBatchNormInfer<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
auto argument_ptr_ref =
batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
......@@ -285,6 +290,7 @@ bool bnorm_infer_nhwc_test(bool do_verification,
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
estimatedMean.mData.data(),
estimatedVariance.mData.data(),
y_ref.mData.data());
......@@ -302,7 +308,7 @@ bool bnorm_infer_nhwc_test(bool do_verification,
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y.mData, y_ref.mData);
pass = pass && ck::utils::check_err(y, y_ref);
};
return (pass);
......
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