mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
gemm/Conv xdlops + dlops quantization (#625)
* Add conv perlayer quantization * Add gemm_dlops quantization * Support int8 for innerproduct * Refine gemm dlops int8 kernel parameter * Support gfx908(MI100) and gfx90a(MI200) * clang-format * Rename example number * Support different layout for d tensor * Add conv dlops perchannel quantization example * Move to example 40 * Extract the common code for different platform (dlops and xdlops) * Move ot subfolder. Prepare to add other op of quantization * Refine the quantization instance library * Add conv dl instances and client example * Remove unnecessary type * Add gemm quantization instance * Add external api and client example * Refine num_bytes * Separete different layout to different cpp * Add more xdl instances * Revert "Remove unnecessary type" This reverts commit820869182f. * Remove CShuffleDataType in dlops Let acc and CShuffleDataType be the same in xdlops --------- Co-authored-by: zjing14 <zhangjing14@gmail.com> [ROCm/composable_kernel commit:16dc18e0f9]
This commit is contained in:
rocking5566
@@ -9,3 +9,6 @@ target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composab
|
||||
|
||||
add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
|
||||
target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_gemm_quantization gemm_quantization.cpp)
|
||||
target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_operations)
|
||||
|
||||
@@ -28,16 +28,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Cla
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4;
|
||||
static constexpr ck::index_t K = 64;
|
||||
static constexpr ck::index_t C = 32;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 71;
|
||||
static constexpr ck::index_t Wi = 71;
|
||||
static constexpr ck::index_t Ho = 36;
|
||||
static constexpr ck::index_t Wo = 36;
|
||||
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
@@ -64,8 +63,8 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
@@ -136,10 +135,11 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
|
||||
G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes =
|
||||
G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
|
||||
G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
@@ -162,11 +162,12 @@ int main(int argc, char* argv[])
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
// run the best intance
|
||||
if(best_op_id != -1)
|
||||
{
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
|
||||
@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4;
|
||||
static constexpr ck::index_t K = 64;
|
||||
static constexpr ck::index_t C = 32;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 71;
|
||||
static constexpr ck::index_t Wi = 71;
|
||||
static constexpr ck::index_t Ho = 36;
|
||||
static constexpr ck::index_t Wo = 36;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
|
||||
G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes =
|
||||
G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
// run the best intance
|
||||
if(best_op_id != -1)
|
||||
{
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
|
||||
@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_C
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4;
|
||||
static constexpr ck::index_t K = 64;
|
||||
static constexpr ck::index_t C = 32;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 71;
|
||||
static constexpr ck::index_t Wi = 71;
|
||||
static constexpr ck::index_t Ho = 36;
|
||||
static constexpr ck::index_t Wo = 36;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
|
||||
G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes =
|
||||
G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
|
||||
G * sizeof(RequantScaleDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
// run the best intance
|
||||
if(best_op_id != -1)
|
||||
{
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
|
||||
@@ -24,15 +24,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4;
|
||||
static constexpr ck::index_t K = 64;
|
||||
static constexpr ck::index_t C = 32;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 71;
|
||||
static constexpr ck::index_t Wi = 71;
|
||||
static constexpr ck::index_t Ho = 36;
|
||||
static constexpr ck::index_t Wo = 36;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -56,8 +56,8 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
@@ -150,11 +150,11 @@ int main(int argc, char* argv[])
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
// run the best intance
|
||||
if(best_op_id != -1)
|
||||
{
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
|
||||
193
client_example/09_quantization/gemm_quantization.cpp
Normal file
193
client_example/09_quantization/gemm_quantization.cpp
Normal file
@@ -0,0 +1,193 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp"
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using ActivationOp = PassThrough;
|
||||
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
using ADataType = int8_t;
|
||||
using BDataType = int8_t;
|
||||
using EDataType = int8_t;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using ELayout = Row;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
};
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
ck::index_t M = 1024;
|
||||
ck::index_t N = 1024;
|
||||
ck::index_t K = 1024;
|
||||
|
||||
ck::index_t StrideA = 1024;
|
||||
ck::index_t StrideB = 1024;
|
||||
ck::index_t StrideE = 1024;
|
||||
|
||||
float requant_scale = 0.03;
|
||||
|
||||
auto f_matrix_space_size =
|
||||
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
|
||||
using Layout = decltype(layout);
|
||||
|
||||
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return (nRow - 1) * stride + nCol;
|
||||
}
|
||||
else
|
||||
{
|
||||
return (nCol - 1) * stride + nRow;
|
||||
}
|
||||
};
|
||||
|
||||
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
|
||||
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
|
||||
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<>,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
|
||||
|
||||
std::string best_op_name;
|
||||
int best_op_id = -1;
|
||||
float best_avg_time = std::numeric_limits<float>::max();
|
||||
float best_gb_per_sec = 0;
|
||||
float best_tflops = 0;
|
||||
|
||||
// profile device operation instances
|
||||
std::cout << "Run all instances and do timing" << std::endl;
|
||||
|
||||
for(int i = 0; i < op_ptrs.size(); ++i)
|
||||
{
|
||||
auto& op_ptr = op_ptrs[i];
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
|
||||
b_device_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
{},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_bytes =
|
||||
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
|
||||
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
||||
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_avg_time = avg_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
best_tflops = tflops;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
if(best_op_id != -1)
|
||||
{
|
||||
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
|
||||
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
auto& op_ptr = op_ptrs[best_op_id];
|
||||
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
|
||||
<< std::endl;
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
|
||||
b_device_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
{},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
std::cout << "Done" << std::endl;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user