Skip to content

GitLab

Unverified Commit 24af0144 authored by Po Yen Chen's avatar Po Yen Chen Committed by GitHub
Browse files

Merge branch 'develop' into gemm_layernorm_welford

parents 961f5e9e b79bbbc2
Show whitespace changes
Inline Side-by-side
Showing with 1110 additions and 923 deletions
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp 0 → 100644
View file @ 24af0144
// 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/device_grouped_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 F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
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::MaskOutUpperTriangle;
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::DeviceGroupedGemmSoftmaxGemmPermute_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: 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_grouped_gemm_scale_softmax_gemm_permute.inc"
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 @ 24af0144
......@@ -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"
......@@ -33,9 +34,6 @@ using S = ck::Sequence<Is...>;
using F16 = ck::half_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 = F16;
......@@ -44,13 +42,14 @@ using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<1, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_M_O
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;
......@@ -59,17 +58,27 @@ 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::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
......@@ -78,6 +87,10 @@ using DeviceGemmInstance =
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
......@@ -118,7 +131,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>;
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -142,303 +155,6 @@ 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;
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
{
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");
exit(0);
}
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 13;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
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 Batch = rand() % 8 + 1;
const int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int StrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int StrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int BatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int BatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int BatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
std::vector<ck::index_t> c_gs_ms_os_lengths{Batch, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{O, Batch * O, 1};
problem_descs.push_back({M,
N,
K,
O,
Batch,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
c_gs_ms_os_lengths,
c_gs_ms_os_strides});
}
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}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::size_t flop = 0, num_byte = 0;
std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& K = problem_descs[i].K;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
const auto& StrideA = problem_descs[i].StrideA;
const auto& StrideB0 = problem_descs[i].StrideB0;
const auto& StrideB1 = problem_descs[i].StrideB1;
const auto& BatchStrideA = problem_descs[i].BatchStrideA;
const auto& BatchStrideB0 = problem_descs[i].BatchStrideB0;
const auto& BatchStrideB1 = problem_descs[i].BatchStrideB1;
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;
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(Batch, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(Batch, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(Batch, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_gs_ms_os_device_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
Batch;
if(i < 4)
{
std::cout << "a_g_m_k[" << i << "]: " << a_g_m_k.mDesc << ", "
<< "b0_g_k_n[" << i << "]: " << b0_g_k_n.mDesc << ", "
<< "b1_g_n_o[" << i << "]: " << b1_g_n_o.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
}
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
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>{});
}
a_tensors.push_back(a_g_m_k);
b0_tensors.push_back(b0_g_k_n);
b1_tensors.push_back(b1_g_n_o);
c_tensors.push_back(c_gs_ms_os_device_result);
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_g_m_k.mData.data());
b0_tensors_device[i]->ToDevice(b0_g_k_n.mData.data());
b1_tensors_device[i]->ToDevice(b1_g_n_o.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
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(p_a,
p_b0,
p_b1,
p_c,
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
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});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
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;
const auto& a_g_m_k = a_tensors[i];
const auto& b0_g_k_n = b0_tensors[i];
const auto& b1_g_n_o = b1_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
Tensor<CDataType> c_gs_ms_os_host_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{Batch, M, O},
std::vector<int>{M * O, O, 1});
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_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_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);
// Note: in this example, we merely permute the dimensions by changing underlying
// strides so we simply access data as-is
c_gs_ms_os_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = c_g_m_o_host_result(idx); });
bool pass_ =
ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
pass &= pass_;
}
}
#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
return pass ? 0 : 1;
}
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc 0 → 100644
View file @ 24af0144
// 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 = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
float alpha = 1;
bool input_permute = false;
bool output_permute = true;
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 == 13)
{
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]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
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 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.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
// TODO ANT: replace array with vector?
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases;
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
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;
}
ck::index_t BatchCount = G0 * G1;
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;
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, M, K});
Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
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);
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
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);
// gemm1
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);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
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)
? 0
: 1;
}
return 0;
}
example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc 0 → 100644
View file @ 24af0144
// 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 = 1;
bool time_kernel = false;
bool input_permute = false;
bool output_permute = true;
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 == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
input_permute = std::stoi(argv[4]);
output_permute = std::stoi(argv[5]);
}
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 5: input / output permute\n");
exit(0);
}
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 7;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
std::vector<std::vector<int>> g0_g1_m_n_k_o;
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::size_t flop = 0, num_byte = 0;
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 G0 = rand() % 3 + 1;
int G1 = rand() % 5 + 1;
g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
problem_descs.push_back({a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
{}, // acc0_biases_gs_ms_ns_lengths
{}, // acc0_biases_gs_ms_ns_strides
{}, // acc1_biases_gs_ms_os_lengths
{}}); // acc1_biases_gs_ms_os_strides
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
int Batch = G0 * G1;
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
Batch;
if(i < 4)
{
std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
<< "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
<< "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
}
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
a_tensors.push_back(a_gs_ms_ks);
b0_tensors.push_back(b0_gs_ns_ks);
b1_tensors.push_back(b1_gs_os_ns);
c_tensors.push_back(c_gs_ms_os_device_result);
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
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(p_a,
p_b0,
p_b1,
p_c,
{}, // p_acc0_biases
{}, // p_acc1_biases
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
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});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
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 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;
const auto& a_gs_ms_ks = a_tensors[i];
const auto& b0_gs_ns_ks = b0_tensors[i];
const auto& b1_gs_os_ns = b1_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({G0 * G1, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
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);
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
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);
// gemm 1
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);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
bool pass_ =
ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
pass &= pass_;
}
}
return pass ? 0 : 1;
}
example/33_multiple_reduce/dual_reduce_common.hpp
View file @ 24af0144
......@@ -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/33_multiple_reduce/dual_reduce_multiblock.cpp
View file @ 24af0144
......@@ -13,7 +13,7 @@
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_multiple_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "dual_reduce_common.hpp"
......
example/33_multiple_reduce/dual_reduce_threadwise.cpp
View file @ 24af0144
......@@ -13,7 +13,7 @@
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_multiple_reduce_threadwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "dual_reduce_common.hpp"
......
example/34_batchnorm/batchnorm_common.hpp
View file @ 24af0144
......@@ -10,131 +10,17 @@
#include "ck/utility/data_type.hpp"
// binary operation used to calculate invVariance from mean and meansquare
struct InvVariance
{
InvVariance(double epsilon) : epsilon_(epsilon){};
template <typename T>
__host__ __device__ constexpr void operator()(T& y, const T& mean, const T& meansquare) const
{
static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
using ck::math::sqrt;
T tmp_epsilon = type_convert<T>(epsilon_);
y = meansquare - mean * mean;
y = 1.0f / sqrt(tmp_epsilon + y);
};
double epsilon_;
};
// (4-in, 2-out) element-wise operation used to update the moving average of mean and variance
struct MovingAverage
{
MovingAverage(double factor) : factor_(factor){};
template <typename T>
__host__ __device__ constexpr void operator()(T& y0,
T& y1,
const T& mean,
const T& runningMean,
const T& meansquare,
const T& runningVariance) const
{
static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
T tmp_factor = type_convert<T>(factor_);
T variance = meansquare - mean * mean;
y0 = runningMean * (type_convert<T>(1.0f) - tmp_factor) + mean * tmp_factor;
y1 = runningVariance * (type_convert<T>(1.0f) - tmp_factor) + variance * tmp_factor;
};
double factor_;
};
struct MovingAverageAndInvVariance
{
MovingAverageAndInvVariance(double epsilon, double factor)
: epsilon_(epsilon), factor_(factor){};
template <typename T>
__host__ __device__ constexpr void operator()(T& y0, // resultRunningMean
T& y1, // resultRunningVariance
T& y2, // saveInvVariance
const T& mean,
const T& runningMean,
const T& meansquare,
const T& runningVariance) const
{
static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
using ck::math::sqrt;
T tmp_epsilon = type_convert<T>(epsilon_);
T tmp_factor = type_convert<T>(factor_);
T variance = meansquare - mean * mean;
y0 = runningMean * (type_convert<T>(1.0f) - tmp_factor) + mean * tmp_factor;
y1 = runningVariance * (type_convert<T>(1.0f) - tmp_factor) + variance * tmp_factor;
y2 = 1.0f / sqrt(tmp_epsilon + variance);
};
double epsilon_;
double factor_;
};
struct NormalizeInInfer
{
NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
template <typename T1, typename T2>
template <typename T1, typename T2, typename T3, typename T4>
__host__ __device__ constexpr void operator()(T1& y,
const T1& x,
const T2& mean,
const T2& variance,
const T2& gamma,
const T2& beta) const
{
static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
using ck::math::sqrt;
T2 tmp_x, tmp_y;
tmp_x = type_convert<T2>(x);
tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) * gamma + beta;
y = type_convert<T1>(tmp_y);
};
double epsilon_;
};
struct NormalizeInForward
{
NormalizeInForward(double epsilon = 1e-4) : epsilon_(epsilon) {}
template <typename T1, typename T2>
__host__ __device__ constexpr void operator()(T1& y,
const T1& x,
const T2& mean,
const T2& meansquare,
const T2& gamma,
const T2& beta) const
const T3& gamma,
const T4& beta) const
{
static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
"Data type is not supported by this operation!");
......@@ -143,11 +29,12 @@ struct NormalizeInForward
using ck::math::sqrt;
T2 tmp_x, tmp_y;
T2 variance = meansquare - mean * mean;
tmp_x = type_convert<T2>(x);
tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) * gamma + beta;
tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
type_convert<T2>(gamma) +
type_convert<T2>(beta);
y = type_convert<T1>(tmp_y);
};
......
example/34_batchnorm/batchnorm_forward_impl.hpp deleted 100644 → 0
View file @ 961f5e9e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cassert>
#include <vector>
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_multiple_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "batchnorm_common.hpp"
template <typename InOutDataType,
typename AccDataType,
ck::index_t Rank,
ck::index_t NumBatchNormReduceDim,
bool fastest_dim_is_reduced = false>
int bnorm_fwd(bool time_kernel,
bool updateMovingAverage,
bool saveMeanAndInvVariance,
const std::array<int, NumBatchNormReduceDim> reduceDims,
const std::array<ck::index_t, Rank> xyLengths,
const std::array<ck::index_t, Rank> xStrides,
const std::array<ck::index_t, Rank> yStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides,
const void* p_x,
const void* p_scale,
const void* p_bias,
void* p_y,
double exponentialAverageFactor,
void* p_runningMean,
void* p_runningVariance,
double epsilon,
void* p_saveMean,
void* p_saveInvVariance,
void* p_tmp_mean,
void* p_tmp_meansquare)
{
static_assert(NumBatchNormReduceDim < Rank,
"Invalid number of reduced dimensions for batchnorm!");
constexpr ck::index_t NumScaleBiasMeanVarDim = Rank - NumBatchNormReduceDim;
using InElementwiseOperation_Mean = ck::tensor_operation::element_wise::PassThrough;
using AccElementwiseOperation_Mean = ck::tensor_operation::element_wise::UnaryDivide;
using InElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnarySquare;
using AccElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnaryDivide;
using DeviceMeanAndMeansquareInstance =
ck::tensor_operation::device::DeviceMultipleReduceMultiBlock<
2,
InOutDataType,
AccDataType,
ck::Tuple<AccDataType, AccDataType>,
Rank,
NumBatchNormReduceDim,
ck::reduce::Add,
ck::Tuple<InElementwiseOperation_Mean, InElementwiseOperation_Meansquare>,
ck::Tuple<AccElementwiseOperation_Mean, AccElementwiseOperation_Meansquare>,
ck::InMemoryDataOperationEnum::Set,
false, // PropagateNan
256,
16,
16,
1,
1,
fastest_dim_is_reduced ? 1 : 0,
1,
ck::Sequence<1, 1>>;
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<InOutDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
// meansquare,
// scale, bias
ck::Tuple<InOutDataType>, // y
NormalizeInForward,
Rank,
2, // MPerthread
ck::Sequence<1, 1, 1, 1, 1>, // scalarPerVector: x, mean, meansquare, scale, bias
ck::Sequence<1>>; // scalarPerVector: y
using DeviceInvVarianceInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<AccDataType, AccDataType>, // mean, meansquare
ck::Tuple<AccDataType>, // invVariance
InvVariance,
NumScaleBiasMeanVarDim,
2, // MPerthread
ck::Sequence<1, 1>, // scalarPerVector: mean, meansquare
ck::Sequence<1>>; // scalarPerVector: invVariance
using DeviceMovingAverageInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<AccDataType, AccDataType, AccDataType, AccDataType>, // old moving mean, new mean,
// old moving variance, new
// meansquare
ck::Tuple<AccDataType, AccDataType>, // updated moving mean, updated moving variance
MovingAverage,
NumScaleBiasMeanVarDim,
4, // MPerthread
ck::Sequence<1, 1, 1, 1>, // scalarPerVector: old moving mean, new mean, old moving
// variance, new meansquare
ck::Sequence<1, 1>>; // scalarPerVector: updated moving mean, updated moving variance
using DeviceMovingAverageAndInvVarianceInstance =
ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<AccDataType, AccDataType, AccDataType, AccDataType>, // old moving mean, new
// mean, old moving
// variance, new
// meansquare
ck::Tuple<AccDataType, AccDataType, AccDataType>, // updated moving mean, updated moving
// variancem, invVariance
MovingAverageAndInvVariance,
NumScaleBiasMeanVarDim,
4, // MPerthread
ck::Sequence<1, 1, 1, 1>, // scalarPerVector: old moving mean, new mean, old moving
// variance, new meansquare
ck::Sequence<1, 1, 1>>; // scalarPerVector: updated moving mean, updated moving variance
auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
int i = 0;
for(auto dim : invariantDims)
{
assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
aligned_scaleBiasMeanVarStrides[dim] = bnScaleBiasMeanVarStrides[i];
i++;
};
int32_t reduceLength = 1;
for(auto dim : reduceDims)
reduceLength *= xyLengths[dim];
int32_t invariantLength = 1;
for(auto dim : invariantDims)
invariantLength *= xyLengths[dim];
size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
float avg_time = 0.0f;
std::size_t num_bytes = 0;
auto dev_mean_and_meansquare = DeviceMeanAndMeansquareInstance{};
void* p_mean = saveMeanAndInvVariance ? p_saveMean : p_tmp_mean;
const AccDataType alpha = ck::type_convert<AccDataType>(1.0f);
const AccDataType beta = ck::type_convert<AccDataType>(0.0f);
auto argument_ptr1 = dev_mean_and_meansquare.MakeArgumentPointer(
xyLengths,
xStrides,
bnScaleBiasMeanVarLengths,
{bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
reduceDims,
{&alpha, &alpha},
{&beta, &beta},
p_x,
{p_mean, p_tmp_meansquare},
ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),
ck::make_tuple(AccElementwiseOperation_Mean{reduceLength},
AccElementwiseOperation_Meansquare{reduceLength}));
auto dev_normalize = DeviceNormalizeInstance{};
auto argument_ptr2 =
dev_normalize.MakeArgumentPointer(xyLengths,
{xStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides},
{yStrides},
{p_x, p_mean, p_tmp_meansquare, p_scale, p_bias},
{p_y},
NormalizeInForward{epsilon});
if(!dev_mean_and_meansquare.IsSupportedArgument(argument_ptr1.get()) ||
!dev_normalize.IsSupportedArgument(argument_ptr2.get()))
{
std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
<< std::endl;
return (-1);
};
auto invoker_ptr1 = dev_mean_and_meansquare.MakeInvokerPointer();
auto invoker_ptr2 = dev_normalize.MakeInvokerPointer();
avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
avg_time += invoker_ptr2->Run(argument_ptr2.get(), StreamConfig{nullptr, time_kernel});
num_bytes +=
(total_length * sizeof(InOutDataType) + invariantLength * 2 * sizeof(AccDataType)) + // No.1
(total_length * (1 * sizeof(InOutDataType) + 4 * sizeof(AccDataType)) +
total_length * sizeof(InOutDataType)); // No.2
if(saveMeanAndInvVariance && updateMovingAverage)
{
auto dev_moving_average_inv_variance = DeviceMovingAverageAndInvVarianceInstance{};
auto argument_ptr3 = dev_moving_average_inv_variance.MakeArgumentPointer(
bnScaleBiasMeanVarLengths,
{bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides},
{bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
{p_mean, p_runningMean, p_tmp_meansquare, p_runningVariance},
{p_runningMean, p_runningVariance, p_saveInvVariance},
MovingAverageAndInvVariance{epsilon, exponentialAverageFactor});
if(!dev_moving_average_inv_variance.IsSupportedArgument(argument_ptr3.get()))
{
std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
return (-1);
};
auto invoker_ptr3 = dev_moving_average_inv_variance.MakeInvokerPointer();
avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
num_bytes += invariantLength * (4 + 3) * sizeof(AccDataType) * 2; // No.5
}
else if(saveMeanAndInvVariance)
{
auto dev_inv_variance = DeviceInvVarianceInstance{};
auto argument_ptr3 = dev_inv_variance.MakeArgumentPointer(
bnScaleBiasMeanVarLengths,
{bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
{bnScaleBiasMeanVarStrides},
{p_mean, p_tmp_meansquare},
{p_saveInvVariance},
InvVariance{epsilon});
if(!dev_inv_variance.IsSupportedArgument(argument_ptr3.get()))
{
std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
return (-1);
};
auto invoker_ptr3 = dev_inv_variance.MakeInvokerPointer();
avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
num_bytes += invariantLength * (2 + 1) * sizeof(AccDataType);
}
else if(updateMovingAverage)
{
auto dev_moving_average = DeviceMovingAverageInstance{};
auto argument_ptr3 = dev_moving_average.MakeArgumentPointer(
bnScaleBiasMeanVarLengths,
{bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides,
bnScaleBiasMeanVarStrides},
{bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
{p_mean, p_runningMean, p_tmp_meansquare, p_runningVariance},
{p_runningMean, p_runningVariance},
MovingAverage{exponentialAverageFactor});
if(!dev_moving_average.IsSupportedArgument(argument_ptr3.get()))
{
std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
return (-1);
};
auto invoker_ptr3 = dev_moving_average.MakeInvokerPointer();
avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
num_bytes += invariantLength * (4 + 2) * sizeof(AccDataType) * 2; // No.5
};
if(time_kernel)
{
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
};
return (0);
};
example/34_batchnorm/batchnorm_forward_nhwc.cpp
View file @ 24af0144
......@@ -9,19 +9,16 @@
#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_forward_nhwc_c.hpp"
#include "batchnorm_forward_impl.hpp"
template <typename InOutDataType, typename AccDataType>
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd_Input_N_H_W_C_Output_C<InOutDataType,
AccDataType>;
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
......@@ -44,6 +41,7 @@ class BatchNormFwdArg
int data_type = 0;
int init_method = 2;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
......@@ -68,6 +66,7 @@ class BatchNormFwdArg
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
......@@ -110,14 +109,15 @@ class BatchNormFwdArg
};
};
if(optind + 5 > argc)
if(optind + 6 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
updateMovingAverage = std::atoi(argv[optind++]);
saveMeanAndInvVariance = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
......@@ -128,7 +128,7 @@ class BatchNormFwdArg
using namespace ck;
template <typename InOutDataType, typename AccDataType>
template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_fwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
......@@ -264,82 +264,145 @@ bool bnorm_fwd_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());
int result = 0;
// used for saving meansquare
DeviceMem workspace(sizeof(AccDataType) * 2 * resultSaveMean_ref.mDesc.GetElementSpaceSize() +
128);
void* p_tmp_mean = workspace.GetDeviceBuffer();
void* p_tmp_meansquare =
static_cast<char*>(p_tmp_mean) +
(sizeof(AccDataType) * resultSaveMean_ref.mDesc.GetElementSpaceSize() + 63) / 64 * 64;
result = bnorm_fwd<InOutDataType, AccDataType, Rank, NumReduceDim, false>(
time_kernel,
updateMovingAverage,
saveMeanAndInvVariance,
{0, 1, 2},
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());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormFwdInstance =
ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
InOutDataType,
AccDataType,
AccDataType, // ScaleDataType
AccDataType, // BiasDataType
AccDataType, // MeanVarDataType
PassThroughOp, // YElementwiseOp
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1,
1,
1,
1,
1>;
auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
averageFactor,
updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr,
epsilon,
PassThroughOp{},
y_dev.GetDeviceBuffer(),
saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
p_tmp_mean,
p_tmp_meansquare);
averageFactor,
updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
if(result < 0)
if(!batchnorm_fwd.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_fwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_fwd.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, scale, bias, outputing of y
num_bytes +=
total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
// outputing of mean, inv-variance
num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
// updating of moving mean, variance
num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 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)
{
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance<InOutDataType, AccDataType>{};
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd_Input_N_H_W_C_Output_C<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp>;
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
y_ref.mData.data(),
0.1, // exponentialAverageFactor
updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr, // resultRunningMean
updateMovingAverage ? resultRunningVariance_ref.mData.data()
: nullptr, // resultRunningVariance
epsilon,
PassThroughOp{},
y_ref.mData.data(),
saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr);
saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout
<< "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
"instance, exiting!"
<< std::endl;
return (-2);
return (false);
};
auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
......@@ -347,7 +410,7 @@ bool bnorm_fwd_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);
if(updateMovingAverage)
{
......@@ -357,23 +420,22 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass =
pass && ck::utils::check_err(resultRunningMean.mData, resultRunningMean_ref.mData);
pass = pass && ck::utils::check_err(resultRunningVariance.mData,
resultRunningVariance_ref.mData);
pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
};
if(saveMeanAndInvVariance)
{
using ck::host_common::dumpBufferToFile;
Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(resultSaveMean.mData, resultSaveMean_ref.mData);
pass = pass && ck::utils::check_err(resultSaveInvVariance.mData,
resultSaveInvVariance_ref.mData);
pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
};
};
......@@ -396,7 +458,17 @@ int main(int argc, char* argv[])
if(arg.data_type == 0)
{
pass = bnorm_fwd_nhwc_test<ck::half_t, float>(arg.do_verification,
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
......@@ -407,7 +479,17 @@ int main(int argc, char* argv[])
}
else if(arg.data_type == 1)
{
pass = bnorm_fwd_nhwc_test<float, float>(arg.do_verification,
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
......@@ -418,7 +500,17 @@ int main(int argc, char* argv[])
}
else if(arg.data_type == 3)
{
pass = bnorm_fwd_nhwc_test<int8_t, float>(arg.do_verification,
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
......@@ -429,7 +521,17 @@ int main(int argc, char* argv[])
}
else if(arg.data_type == 5)
{
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float>(arg.do_verification,
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
......@@ -440,7 +542,17 @@ int main(int argc, char* argv[])
}
else if(arg.data_type == 6)
{
pass = bnorm_fwd_nhwc_test<double, double>(arg.do_verification,
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
......@@ -452,12 +564,21 @@ int main(int argc, char* argv[])
}
else
{
pass = bnorm_fwd_nhwc_test<ck::half_t, float>(true,
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
2,
false, // don't time kernel
{128, 16, 16, 1024},
{128, 16, 6, 512},
true,
true,
averageFactor,
epsilon);
pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
2,
false, // don't time kernel
{128, 16, 3, 1024},
true,
true,
false,
averageFactor,
epsilon);
};
......
example/34_batchnorm/batchnorm_infer_impl.hpp
View file @ 24af0144
......@@ -10,12 +10,16 @@
#include "ck/utility/sequence.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
#include "batchnorm_common.hpp"
template <typename InOutDataType,
template <typename XDataType,
typename YDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename MeanVarDataType,
ck::index_t Rank,
ck::index_t NumBatchNormReduceDim,
bool fastest_dim_is_reduced = false>
......@@ -26,7 +30,9 @@ int bnorm_infer(
const std::array<ck::index_t, Rank> xStrides,
const std::array<ck::index_t, Rank> yStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
const void* p_x,
const void* p_scale,
const void* p_bias,
......@@ -41,11 +47,11 @@ int bnorm_infer(
"Invalid number of reduced dimensions for batchnorm!");
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<InOutDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
// variance,
// scale,
// bias,
ck::Tuple<InOutDataType>, // y
ck::Tuple<YDataType>, // y
NormalizeInInfer,
Rank,
2, // MPerthread
......@@ -53,14 +59,18 @@ int bnorm_infer(
ck::Sequence<1>>; // scalarPerVector: y
auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
int i = 0;
for(auto dim : invariantDims)
{
assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
aligned_scaleBiasMeanVarStrides[dim] = bnScaleBiasMeanVarStrides[i];
aligned_bnScaleStrides[dim] = bnScaleStrides[i];
aligned_bnBiasStrides[dim] = bnBiasStrides[i];
aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
i++;
};
......@@ -84,10 +94,10 @@ int bnorm_infer(
auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
xyLengths,
{xStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides,
aligned_scaleBiasMeanVarStrides},
aligned_bnMeanVarStrides,
aligned_bnMeanVarStrides,
aligned_bnScaleStrides,
aligned_bnBiasStrides},
{yStrides},
{p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
{p_y},
......@@ -105,8 +115,10 @@ int bnorm_infer(
avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
num_bytes += (total_length * (1 * sizeof(InOutDataType) + 4 * sizeof(AccDataType)) +
total_length * sizeof(InOutDataType));
num_bytes += total_length * sizeof(XDataType) +
invariantLength *
(sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
total_length * sizeof(YDataType);
if(time_kernel)
{
......
example/34_batchnorm/batchnorm_infer_nhwc.cpp
View file @ 24af0144
......@@ -9,6 +9,7 @@
#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"
......@@ -18,11 +19,6 @@
#include "batchnorm_infer_impl.hpp"
template <typename InOutDataType, typename AccDataType>
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<InOutDataType,
AccDataType>;
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
......@@ -225,25 +221,30 @@ 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;
result = bnorm_infer<InOutDataType, AccDataType, Rank, NumReduceDim, false>(
time_kernel,
result = bnorm_infer<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
Rank,
NumReduceDim,
false>(time_kernel,
{0, 1, 2},
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
......@@ -259,7 +260,15 @@ bool bnorm_infer_nhwc_test(bool do_verification,
if(do_verification)
{
auto batchNormInfer_ref = ReferenceBatchNormInferInstance<InOutDataType, AccDataType>{};
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<
InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType>;
auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
auto argument_ptr_ref =
batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
......@@ -267,6 +276,8 @@ bool bnorm_infer_nhwc_test(bool do_verification,
i_inOutStrides,
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
......@@ -288,7 +299,7 @@ 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);
......
example/35_splitK_gemm/run_splitK_gemm_example.inc
View file @ 24af0144
......@@ -34,15 +34,15 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -146,15 +146,12 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
if(std::is_same<CDataType, ck::half_t>::value)
{
pass &= ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"fp16 incorrect result",
3e-3,
1e-3);
pass &= ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
}
else
{
pass &= ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
}
......
example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
View file @ 24af0144
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
View file @ 24af0144
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
View file @ 24af0144
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
View file @ 24af0144
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
View file @ 24af0144
......@@ -8,7 +8,7 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......
example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
View file @ 24af0144
......@@ -9,7 +9,7 @@
#include <ctime>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_sparse_embedding3_forward_layernorm.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
......@@ -86,12 +86,10 @@ int main()
constexpr auto index_length = 2048;
constexpr AccDataType epsilon = 1e-4;
auto f_host_tensor_desc_1d = [](std::size_t len_) {
return HostTensorDescriptor(std::vector<std::size_t>({len_}));
};
auto f_host_tensor_desc_1d = [](std::size_t len_) { return HostTensorDescriptor({len_}); };
auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) {
return HostTensorDescriptor(std::vector<std::size_t>({rows_, cols_}));
return HostTensorDescriptor({rows_, cols_});
};
using ReferenceInstance =
......@@ -203,8 +201,7 @@ int main()
ref_invoker.Run(ref_argument);
out_dev.FromDevice(out_from_dev.mData.data());
pass &= ck::utils::check_err(
out_from_dev.mData, out.mData, "Error: Incorrect results", 1e-3, 1e-3);
pass &= ck::utils::check_err(out_from_dev, out, "Error: Incorrect results", 1e-3, 1e-3);
}
double total_read = current_dim * index_length * 3 * sizeof(EmbType) +
......
example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
View file @ 24af0144
......@@ -12,13 +12,14 @@ Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/binary_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/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......@@ -314,15 +315,15 @@ int main(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});
}
};
......@@ -511,8 +512,7 @@ int main(int argc, char* argv[])
cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
});
return ck::utils::check_err(e1_g_m_o_device_result.mData, e1_g_m_o_host_result.mData) ? 0
: 1;
return ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result) ? 0 : 1;
}
return 0;
......
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