Skip to content

GitLab

Unverified Commit 0d2aafb2 authored by Rostyslav Geyyer's avatar Rostyslav Geyyer Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-359_int4

parents bd78cb4b e0d8806c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2023 additions and 317 deletions
example/24_batched_gemm_e_permute/CMakeLists.txt deleted 100644 → 0
View file @ bd78cb4b
add_example_executable(example_batched_gemm_e_permute_xdl_fp16 batched_gemm_e_permute_xdl_fp16.cpp)
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
View file @ 0d2aafb2
......@@ -280,10 +280,11 @@ int main(int argc, char* argv[])
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
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());
......
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ 0d2aafb2
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_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp 0 → 100644
View file @ 0d2aafb2
// 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_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 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;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmDefault,
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
// 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>;
int main(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 = 128;
ck::index_t N = 1024;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
float alpha = 1;
// 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;
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 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 == 11)
{
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]);
}
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");
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;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
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;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
const int BatchCount = G0 * G1;
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_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()));
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()));
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_gs_ms_os: " << c_gs_ms_os_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_gs_ms_os_device_buf(sizeof(CDataType) *
c_gs_ms_os_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_gs_ms_os_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
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;
if(do_verification)
{
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_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{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, 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_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);
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/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
View file @ 0d2aafb2
......@@ -2,11 +2,11 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
|------------|
Gemm0
|---------------------|
Gemm1
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>
......@@ -212,9 +212,9 @@ int main(int argc, char* argv[])
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 17: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
"BatchStrideB0, BatchStrideB1, BatchStrideC\n");
printf("arg18: alpha\n");
printf("arg17: scale (alpha)\n");
exit(0);
}
......@@ -297,10 +297,11 @@ int main(int argc, char* argv[])
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
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());
......
example/35_splitK_gemm/CMakeLists.txt 0 → 100644
View file @ 0d2aafb2
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
example/35_splitK_gemm/run_splitK_gemm_example.inc 0 → 100644
View file @ 0d2aafb2
#pragma once
struct ProblemSize final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
ck::index_t stride_C = N;
ck::index_t k_batch = 4;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using namespace ck::literals;
auto& [M, N, K, StrideA, StrideB, StrideC, KBatch] = problem_size;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
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}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
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, config.time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
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_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(config.do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
if(std::is_same<CDataType, ck::half_t>::value)
{
return ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"fp16 incorrect result",
3e-3,
1e-3);
}
else
{
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
}
}
return true;
}
bool run_splitK_gemm_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
if(argc == 1)
{
// use default case
}
else if(argc == 5)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.k_batch = std::stoi(argv[4]);
}
else if(argc == 11)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.k_batch = std::stoi(argv[4]);
problem_size.M = std::stoi(argv[5]);
problem_size.N = std::stoi(argv[6]);
problem_size.K = std::stoi(argv[7]);
problem_size.stride_A = std::stoi(argv[8]);
problem_size.stride_B = std::stoi(argv[9]);
problem_size.stride_C = std::stoi(argv[10]);
}
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: KBatch\n");
printf("arg5 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
return run_splitK_gemm(problem_size, config);
}
example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp 0 → 100644
View file @ 0d2aafb2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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/device_gemm_xdl_splitk_c_shuffle.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_gemm.hpp"
#include "ck/library/utility/literals.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 BDataType = BF16;
using AccDataType = F32;
using CDataType = F32;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp 0 → 100644
View file @ 0d2aafb2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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/device_gemm_xdl_splitk_c_shuffle.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_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
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;
using BDataType = F16;
using AccDataType = F32;
using CDataType = F16;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp 0 → 100644
View file @ 0d2aafb2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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/device_gemm_xdl_splitk_c_shuffle.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_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
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 = F32;
using BDataType = F32;
using AccDataType = F32;
using CDataType = F32;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp 0 → 100644
View file @ 0d2aafb2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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/device_gemm_xdl_splitk_c_shuffle.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_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CDataType = int32_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 16, 16, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 16, 16, true, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
example/CMakeLists.txt
View file @ 0d2aafb2
......@@ -38,7 +38,7 @@ add_subdirectory(20_convnd_bwd_weight)
add_subdirectory(21_gemm_layernorm)
add_subdirectory(22_cgemm)
add_subdirectory(23_softmax)
add_subdirectory(24_batched_gemm_e_permute)
add_subdirectory(24_batched_gemm)
add_subdirectory(25_gemm_bias_e_permute)
add_subdirectory(26_contraction)
add_subdirectory(27_layernorm)
......@@ -49,4 +49,4 @@ add_subdirectory(31_batched_gemm_gemm)
add_subdirectory(32_batched_gemm_scale_softmax_gemm)
add_subdirectory(33_multiple_reduce)
add_subdirectory(34_batchnorm)
add_subdirectory(35_splitK_gemm)
include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 0d2aafb2
......@@ -129,6 +129,25 @@ namespace device {
// B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
// D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// FIXME: TensorSpecialization::Packed specialization does not cover all packed tensor cases, it
// merely degenerates into TensorSpecialization::Default with NumDimG/M/N/K = 1
//
// Detail- Packed tensor satisfies
// stride_0 = 1
// stride_i = stride_{i - 1} * extent_{i - 1}
// So tensor
// [G0, G1, G2, M, N]
// transposed into tensor
// [G0, G2, G1, M, N]
// with strides
// [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
//
// Might need to expose dimension order to the interface to fully support
// TensorSpecialization::Packed.
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
View file @ 0d2aafb2
......@@ -54,33 +54,6 @@ struct DeviceBatchedGemmGemm : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CLayout,
typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
using DeviceBatchedGemmGemmPtr = std::unique_ptr<DeviceBatchedGemmGemm<ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AElementwiseOperation,
B0ElementwiseOperation,
Acc0ElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm_xdl_cshuffle.hpp
View file @ 0d2aafb2
......@@ -12,6 +12,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -188,6 +189,10 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto matrix_padder =
GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
......@@ -203,92 +208,18 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
const auto MPad = M - MRaw;
const auto KPad = K - KRaw;
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
assert(KRaw % AK1 == 0);
const auto AK0 = K / AK1;
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
return transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
......@@ -306,84 +237,18 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
}
}();
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
const auto BK0 = KRaw / BK1;
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
const auto BK0 = K / BK1;
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
return transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
// Args: Gemm1KRaw, Gemm1NRaw, StrideB1
......@@ -402,47 +267,19 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
}
}();
const auto N = math::integer_divide_ceil(NRaw, Gemm1NPerBlock) * Gemm1NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, Gemm1KPerBlock) * Gemm1KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
// TODO: implement finer-grained padding
if constexpr(GemmSpec == GemmSpecialization::Default)
{
const auto B1K0 = KRaw / B1K1;
const auto N = b1_grid_desc_n_k.GetLength(I0);
const auto K = b1_grid_desc_n_k.GetLength(I1);
const auto b1_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b1_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
const auto B1K0 = K / B1K1;
return b1_grid_desc_bk0_n_bk1;
}
else
{
// pad both B1N and B1K
const auto B1K0 = K / B1K1;
const auto b1_grid_desc_n_k =
transform_tensor_descriptor(b1_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b1_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b1_grid_desc_bk0_n_bk1;
}
return transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
......@@ -460,47 +297,7 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, Gemm1NPerBlock) * Gemm1NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
}
struct ComputeBasePtrOfStridedBatch
......@@ -651,13 +448,15 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
b1_element_op_{b1_element_op},
c_element_op_{c_element_op},
batch_count_(Batch),
compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC}
compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC},
raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw}
{
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
b1_grid_desc_bk0_n_bk1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
block_2_ctile_map_,
raw_lengths_m_n_k_o_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
......@@ -684,6 +483,9 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
CElementwiseOperation c_element_op_;
index_t batch_count_;
ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
// For robust IsSupportedArgument() check
std::vector<index_t> raw_lengths_m_n_k_o_;
};
// Invoker
......@@ -697,7 +499,8 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
arg.block_2_ctile_map_,
arg.raw_lengths_m_n_k_o_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
......@@ -787,11 +590,37 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
return false;
}
// Note: we need raw lengths since threadwise copy can not handle vector load when part of
// vector is out of bounds
const auto MRaw = arg.raw_lengths_m_n_k_o_[0];
const auto NRaw = arg.raw_lengths_m_n_k_o_[1];
const auto KRaw = arg.raw_lengths_m_n_k_o_[2];
const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
// Check scalar per vector requirement
const auto a_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
const auto b_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
const auto b1_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
const auto c_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
arg.block_2_ctile_map_,
arg.raw_lengths_m_n_k_o_);
}
// polymorphic
......@@ -903,7 +732,8 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
<< MPerBlock << ", "
<< Gemm1NPerBlock << ", "
<< Gemm1KPerBlock << ", "
<< B1K1 << ">";
<< B1K1 << ", "
<< getGemmSpecializationString(GemmSpec) << ">";
// clang-format on
return str.str();
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
View file @ 0d2aafb2
......@@ -54,34 +54,6 @@ struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CLayout,
typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
using DeviceBatchedGemmSoftmaxGemmPtr =
std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AElementwiseOperation,
B0ElementwiseOperation,
Acc0ElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp 0 → 100644
View file @ 0d2aafb2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b0,
const void* p_b1,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t O,
ck::index_t Batch,
std::vector<index_t> c_gs_ms_os_lengths,
std::vector<index_t> c_gs_ms_os_strides,
ck::index_t StrideA,
ck::index_t StrideB0,
ck::index_t StrideB1,
ck::index_t BatchStrideA,
ck::index_t BatchStrideB0,
ck::index_t BatchStrideB1,
AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op,
Acc0ElementwiseOperation acc0_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ 0d2aafb2
......@@ -95,7 +95,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(M)),
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
......
include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
View file @ 0d2aafb2
......@@ -9,6 +9,7 @@ namespace device {
enum struct GemmSpecialization
{
// Gemm
Default,
MPadding,
NPadding,
......@@ -17,6 +18,15 @@ enum struct GemmSpecialization
MKPadding,
NKPadding,
MNKPadding,
// Gemm + Gemm
OPadding,
MOPadding,
NOPadding,
KOPadding,
MNOPadding,
MKOPadding,
NKOPadding,
MNKOPadding,
};
inline std::string getGemmSpecializationString(const GemmSpecialization& s)
......@@ -31,6 +41,14 @@ inline std::string getGemmSpecializationString(const GemmSpecialization& s)
case GemmSpecialization::MKPadding: return "MKPadding";
case GemmSpecialization::NKPadding: return "NKPadding";
case GemmSpecialization::MNKPadding: return "MNKPadding";
case GemmSpecialization::OPadding: return "OPadding";
case GemmSpecialization::MOPadding: return "MOPadding";
case GemmSpecialization::NOPadding: return "NOPadding";
case GemmSpecialization::KOPadding: return "KOPadding";
case GemmSpecialization::MNOPadding: return "MNOPadding";
case GemmSpecialization::MKOPadding: return "MKOPadding";
case GemmSpecialization::NKOPadding: return "NKOPadding";
case GemmSpecialization::MNKOPadding: return "MNKOPadding";
default: return "Unrecognized specialization!";
}
}
......
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