Skip to content

GitLab

Commit 6ef4e211 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into contraction

parents b0a2afb9 9e4429f9
Hide whitespace changes
Inline Side-by-side
Showing with 1592 additions and 1683 deletions
profiler/include/profile_gemm_bias_relu_add_impl.hpp → profiler/include/profile_gemm_bilinear_impl.hpp
View file @ 6ef4e211
......@@ -3,63 +3,45 @@
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation_add.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/gemm_bilinear.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmBiasReluAddPtr = ck::tensor_operation::device::DeviceGemmBiasActivationAddPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddReluAdd>;
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename DDataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_add_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideC1,
int KBatch = 1)
typename DELayout> // assume Ds and E have same layout
bool profile_gemm_bilinear_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD,
int StrideE,
float alpha,
float beta)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
......@@ -77,21 +59,14 @@ void profile_gemm_bias_relu_add_impl(int do_verification,
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{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
// c1_m_n[m ,n]
Tensor<BDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
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;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
......@@ -99,161 +74,136 @@ void profile_gemm_bias_relu_add_impl(int do_verification,
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{});
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Bilinear;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{alpha, beta};
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddReluAdd;
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
// 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;
// run reference
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
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,
c0_n,
c1_m_n,
a_element_op,
b_element_op,
c_element_op);
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
d_m_n_device_buf.ToDevice(d_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasReluAddPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
c_element_op,
KBatch);
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, 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 * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N +
sizeof(CDataType) * M * N;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * 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_name << std::endl;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
......@@ -261,31 +211,22 @@ void profile_gemm_bias_relu_add_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a: ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0: ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c1: ", c1_m_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass &&
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
......
profiler/include/profile_gemm_impl.hpp
View file @ 6ef4e211
......@@ -12,112 +12,37 @@
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch)
int profile_gemm_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
......@@ -134,32 +59,25 @@ void profile_gemm_impl(int do_verification,
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_device_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
// case 0: break;
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{}, num_thread);
break;
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
......@@ -176,303 +94,72 @@ void profile_gemm_impl(int do_verification,
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmNoOpPtr> gemm_ptrs;
if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
is_same<CDataType, float>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
}
}
}
else if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
using DeviceOp = ck::tensor_operation::device::DeviceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
}
else if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
// Run reference GEMM
if(do_verification)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
is_same<CDataType, int8_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
auto ref_op = ReferenceGemmInstance{};
auto ref_invoker = ref_op.MakeInvoker();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
}
}
auto ref_argument = ref_op.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
ref_invoker.Run(ref_argument);
}
std::string best_gemm_name;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c_device_buf.SetZero();
std::string gemm_name = gemm_ptr->GetTypeString();
std::string op_name = op_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -487,11 +174,11 @@ void profile_gemm_impl(int do_verification,
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << gemm_name << std::endl;
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
......@@ -501,86 +188,15 @@ void profile_gemm_impl(int do_verification,
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_f32_m_k,
b_f32_k_n,
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
else
{
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
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);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
pass =
pass & ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
......@@ -588,8 +204,7 @@ void profile_gemm_impl(int do_verification,
}
else
{
std::cout << gemm_ptr->GetTypeString() << " does not support this GEMM problem"
<< std::endl;
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
......@@ -631,7 +246,9 @@ void profile_gemm_impl(int do_verification,
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_gemm_name << std::endl;
<< best_op_name << std::endl;
return pass ? 0 : 1;
}
} // namespace profiler
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 6ef4e211
......@@ -19,23 +19,19 @@
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
namespace instance {
using F32 = float;
using F16 = ck::half_t;
using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using ReduceInElementOps = ck::Tuple<Identity, Square>;
using ReduceOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr =
ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -49,7 +45,7 @@ void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -60,7 +56,7 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename ReduceDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -95,22 +91,22 @@ bool profile_gemm_reduce_impl(int do_verification,
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<DDataType> d0_m_host_result(
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
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;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
......@@ -130,34 +126,37 @@ bool profile_gemm_reduce_impl(int do_verification,
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto reduce0_op = ReduceOp0{};
const auto reduce1_op = ReduceOp1{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
ReduceDataType,
AElementOp,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
using ReduceAccDataType = ReduceDataType;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -169,44 +168,43 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val =
ReduceAccDataType d0_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
square(d1_val, d0_val);
reduce0_op(reduce0_acc, d0_val);
reduce1_op(reduce1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
gemm_ptrs;
std::vector<ck::tensor_operation::device::instance::DeviceGemmReduceNoOpPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
......@@ -215,7 +213,7 @@ bool profile_gemm_reduce_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
ck::tensor_operation::device::instance::
add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
gemm_ptrs);
}
......@@ -223,7 +221,7 @@ bool profile_gemm_reduce_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
ck::tensor_operation::device::instance::
add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
gemm_ptrs);
}
......@@ -231,7 +229,7 @@ bool profile_gemm_reduce_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
ck::tensor_operation::device::instance::
add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
gemm_ptrs);
}
......@@ -239,7 +237,7 @@ bool profile_gemm_reduce_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
ck::tensor_operation::device::instance::
add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
gemm_ptrs);
}
......@@ -258,30 +256,31 @@ bool profile_gemm_reduce_impl(int do_verification,
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op);
auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -311,12 +310,12 @@ bool profile_gemm_reduce_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
if(do_log)
{
......@@ -326,13 +325,17 @@ bool profile_gemm_reduce_impl(int do_verification,
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
LogRangeAsType<float>(
std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
<< std::endl;
}
}
......
profiler/include/profile_gemm_bias_relu_impl.hpp → profiler/include/profile_gemm_splitk_impl.hpp
View file @ 6ef4e211
......@@ -3,63 +3,47 @@
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmBiasReluPtr = ck::tensor_operation::device::DeviceGemmBiasActivationPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch = 1)
bool profile_gemm_splitk_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
......@@ -79,160 +63,124 @@ void profile_gemm_bias_relu_impl(int do_verification,
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{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
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;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddRelu;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// 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;
// Run reference GEMM
if(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, c0_n, a_element_op, b_element_op, c_element_op);
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);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasReluPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
std::string op_name = op_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, 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 * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
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_name << std::endl;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
......@@ -242,13 +190,13 @@ void profile_gemm_bias_relu_impl(int do_verification,
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
pass =
pass & ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0 : ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
......@@ -258,12 +206,51 @@ void profile_gemm_bias_relu_impl(int do_verification,
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
if constexpr(is_same<CDataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<CDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<CDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<CDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
return pass;
}
} // namespace profiler
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ 6ef4e211
......@@ -20,7 +20,7 @@
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
namespace instance {
using DeviceGroupedGemmNoOpPtr = ck::tensor_operation::device::DeviceGroupedGemmPtr<
ck::tensor_operation::element_wise::PassThrough,
......@@ -36,7 +36,7 @@ void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
} // namespace device_grouped_gemm_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -171,9 +171,7 @@ void profile_grouped_gemm_impl(int do_verification,
}
// add device GEMM instances
std::vector<
ck::tensor_operation::device::device_grouped_gemm_instance::DeviceGroupedGemmNoOpPtr>
gemm_ptrs;
std::vector<ck::tensor_operation::device::instance::DeviceGroupedGemmNoOpPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
......@@ -182,28 +180,28 @@ void profile_grouped_gemm_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
......@@ -232,6 +230,10 @@ void profile_grouped_gemm_impl(int do_verification,
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
......
profiler/include/profile_normalization_impl.hpp 0 → 100644
View file @ 6ef4e211
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
enum struct NormType
{
LAYERNORM,
BATCHNORM,
SOFTMAX,
};
enum struct NormDataType
{
F32_F32, // in, out
F16_F16,
BF16_BF16,
INT8_INT8,
};
// clang-format off
template <typename NormDataType> std::string type_to_string();
template <> std::string type_to_string<float>() { return "f32"; }
template <> std::string type_to_string<half_t>() { return "f16"; }
template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
template <> std::string type_to_string<int8_t>() { return "int8"; }
template <> std::string type_to_string<int32_t>() { return "int32"; }
// clang-format on
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_normalization_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> in_length,
std::vector<index_t> in_strides,
std::vector<index_t> reduce_dims,
AccDataType alpha,
AccDataType beta,
NormType norm_type)
{
Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
: Tensor<InDataType>(in_length, in_strides);
Tensor<OutDataType> out(in.mDesc);
switch(init_method)
{
// case 0: break;
case 0:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{});
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
}
Tensor<OutDataType> out_ref(out);
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
out_dev.ToDevice(out.mData.data());
std::vector<index_t> i_in_lengths(in.mDesc.GetLengths().begin(), in.mDesc.GetLengths().end());
std::vector<index_t> i_in_strides(in.mDesc.GetStrides().begin(), in.mDesc.GetStrides().end());
// add device normalization instances
std::vector<tensor_operation::device::DeviceNormalizationPtr> instances;
if(norm_type == NormType::SOFTMAX)
{
if constexpr(is_same<InDataType, half_t>::value && is_same<OutDataType, half_t>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank3_instances(
instances);
if(in_length.size() == 4)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank4_instances(
instances);
}
else if constexpr(is_same<InDataType, float>::value && is_same<OutDataType, float>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank3_instances(
instances);
if(in_length.size() == 4)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank4_instances(
instances);
}
}
if(instances.size() <= 0)
{
throw std::runtime_error("wrong! no device normalization instance found");
}
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
for(auto& inst_ptr : instances)
{
// Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
// problem to rank 4 kernel) other than invoking IsSupportedArgument()?
if(!(inst_ptr->GetRank() == static_cast<index_t>(i_in_lengths.size()) &&
inst_ptr->GetNumReduceDim() == static_cast<index_t>(reduce_dims.size())))
{
continue;
}
auto argument_ptr = inst_ptr->MakeArgumentPointer(i_in_lengths,
i_in_strides,
reduce_dims,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
return;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
in.mDesc.GetElementSize() * sizeof(InDataType) +
(beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
// TODO: factory method to dynamically switch between different reference normalizations
using ReferenceFactory =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
ReferenceFactory{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
out_dev.FromDevice(out.mData.data());
bool pass;
if(std::is_same<InDataType, int8_t>::value)
{
pass = ck::utils::check_err(
out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
if(do_log)
{
LogRangeAsType<int>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<int>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<int>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
else
{
pass = ck::utils::check_err(out.mData, out_ref.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
}
}
}
std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
<< type_to_string<OutDataType>() << ", ";
LogRange(std::cout << "length = ", i_in_lengths, ",") << ", ";
LogRange(std::cout << "stride = ", i_in_strides, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
std::cout << "alpha = " << alpha << ", "
<< "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
<< " GB/s, " << best_instance_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_reduce_impl.hpp
View file @ 6ef4e211
......@@ -16,7 +16,7 @@
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
namespace instance {
template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription
......@@ -91,7 +91,7 @@ bool description_match(const DescriptionType& description,
return (result);
};
} // namespace device_reduce_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -142,7 +142,7 @@ bool profile_reduce_impl_impl(bool do_verification,
float beta)
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::tensor_operation::device::instance;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
......@@ -464,7 +464,7 @@ bool profile_reduce_impl(bool do_verification,
bool pass = true;
using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances;
tensor_operation::device::instance::reduce_description_instances;
const auto tuple_object = tuple_of_description_instances{};
......
profiler/src/profile_batched_gemm.cpp
View file @ 6ef4e211
......@@ -15,10 +15,6 @@ enum struct GemmMatrixLayout
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
......@@ -31,8 +27,9 @@ enum struct GemmDataType
int profile_batched_gemm(int argc, char* argv[])
{
if(!(argc == 15))
if(argc != 18)
{
// clang-format off
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
......@@ -43,7 +40,8 @@ int profile_batched_gemm(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
// clang-format on
exit(1);
}
......@@ -62,332 +60,138 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
const int BatchCount = std::stoi(argv[17]);
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(BF16{}, BF16{}, BF16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(BF16{}, BF16{}, BF16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(BF16{}, BF16{}, BF16{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<int8_t,
int8_t,
int8_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<int8_t,
int8_t,
int8_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_impl<int8_t,
int8_t,
int8_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_impl<int8_t,
int8_t,
int8_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
return 1;
}
}
profiler/src/profile_convnd_fwd.cpp
View file @ 6ef4e211
......@@ -10,11 +10,10 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/utility/fill.hpp"
#include "profiler/include/profile_convnd_fwd.hpp"
namespace {
enum struct ConvDataType
......@@ -304,7 +303,7 @@ void profile_convnd_instances(ConvDataType data_type,
} // namespace
int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
int profile_convnd_fwd(int argc, char* argv[])
{
using namespace ck::utils::conv;
......
profiler/src/profile_gemm.cpp
View file @ 6ef4e211
......@@ -14,10 +14,6 @@ enum struct GemmMatrixLayout
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
......@@ -30,7 +26,7 @@ enum struct GemmDataType
int profile_gemm(int argc, char* argv[])
{
if(!(argc == 14 || argc == 15))
if(argc != 14)
{
printf("arg1: tensor operation (gemm: GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
......@@ -41,9 +37,8 @@ int profile_gemm(int argc, char* argv[])
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
}
......@@ -61,350 +56,125 @@ int profile_gemm(int argc, char* argv[])
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
int KBatch = 1;
if(argc == 15)
KBatch = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
using INT32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass =
ck::profiler::profile_gemm_impl<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(INT8{}, INT8{}, INT32{}, INT8{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(INT8{}, INT8{}, INT32{}, INT8{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(INT8{}, INT8{}, INT32{}, INT8{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
return profile(INT8{}, INT8{}, INT32{}, INT8{}, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
return 1;
}
}
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ 6ef4e211
......@@ -16,10 +16,6 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
MK_NK_MN_MN_MN, // 1
KM_KN_MN_MN_MN, // 2
KM_NK_MN_MN_MN, // 3
MK_KN_NM_MN_MN, // 4
MK_NK_NM_MN_MN, // 5
KM_KN_NM_MN_MN, // 6
KM_NK_NM_MN_MN, // 7
};
enum struct MatrixDataType
......@@ -33,7 +29,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
......@@ -43,7 +39,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
......@@ -79,9 +75,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto e_layout) {
auto de_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
......@@ -91,27 +85,23 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using ELayout = decltype(e_layout);
using DELayout = decltype(de_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>(
const int DefaultStrideD0 = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<DELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
DELayout>(
do_verification,
init_method,
do_log,
......@@ -124,31 +114,33 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::MK_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
return 1;
}
}
profiler/src/profile_gemm_bias_2d.cpp deleted 100644 → 0
View file @ b0a2afb9
// 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 "profiler/include/profile_gemm_bias_2d_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_2d(int argc, char* argv[])
{
if(!(argc == 16 || argc == 17))
{
printf("arg1: tensor operation (gemm: GEMM+Bias_2d)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: alpha\n");
printf("arg15: beta\n");
printf("arg16: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const float alpha = std::stof(argv[14]);
const float beta = std::stof(argv[15]);
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu_add.cpp deleted 100644 → 0
View file @ b0a2afb9
// 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 "profiler/include/profile_gemm_bias_relu_add_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_relu_add(int argc, char* argv[])
{
if(!(argc == 15 || argc == 16))
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU+Add)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int StrideC1 = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bilinear.cpp 0 → 100644
View file @ 6ef4e211
// 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 "profiler/include/profile_gemm_bilinear_impl.hpp"
int profile_gemm_bilinear(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F32_F32_F32_F32, // 0
F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8, // 3
};
if(argc != 17)
{
// clang-format off
printf("arg1: tensor operation (gemm_bilinear: GEMM+Bilinear)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
printf(" 1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
printf(" 2: E[m, n] = alpha * A[k, m] * B[k, n] + beta * D[m, n];\n");
printf(" 3: E[m, n] = alpha * A[k, m] * B[n, k] + beta * D[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
printf("arg15 to 16: alhpa, beta\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
const float alpha = std::stof(argv[15]);
const float beta = std::stof(argv[16]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d_type,
auto e_type,
auto a_layout,
auto b_layout,
auto de_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using DDataType = decltype(d_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using DELayout = decltype(de_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<DELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_bilinear_impl<ADataType,
BDataType,
AccDataType,
DDataType,
EDataType,
ALayout,
BLayout,
DELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD < 0) ? DefaultStrideD : StrideD,
(StrideE < 0) ? DefaultStrideE : StrideE,
alpha,
beta);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
profiler/src/profile_gemm_bias_relu.cpp → profiler/src/profile_gemm_splitk.cpp
View file @ 6ef4e211
......@@ -6,7 +6,7 @@
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_relu_impl.hpp"
#include "profiler/include/profile_gemm_splitk_impl.hpp"
enum struct GemmMatrixLayout
{
......@@ -14,24 +14,22 @@ enum struct GemmMatrixLayout
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
int profile_gemm_bias_relu(int argc, char* argv[])
int profile_gemm_splitk(int argc, char* argv[])
{
if(!(argc == 14 || argc == 15))
if(argc != 15)
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg1: tensor operation (gemm_splitk: Split-K GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
......@@ -39,7 +37,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
......@@ -59,15 +57,41 @@ int profile_gemm_bias_relu(int argc, char* argv[])
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int KBatch = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
using F32 = float;
using F16 = ck::half_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_splitk_impl<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
do_verification,
init_method,
do_log,
......@@ -75,71 +99,50 @@ int profile_gemm_bias_relu(int argc, char* argv[])
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC);
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
KBatch);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC);
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC);
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC);
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
return 1;
}
}
profiler/src/profile_normalization.cpp 0 → 100644
View file @ 6ef4e211
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/include/profile_normalization_impl.hpp"
using ck::index_t;
using ck::profiler::NormDataType;
using ck::profiler::NormType;
struct ArgParser
{
std::unordered_map<std::string, NormType> norm_dict = {{"layernorm", NormType::LAYERNORM},
{"batchnorm", NormType::BATCHNORM},
{"softmax", NormType::SOFTMAX}};
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help()
{
std::cout << "arg1: tensor operation (layernorm/batchnorm/softmax)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg5: print tensor value (0: no; 1: yes)\n"
<< "arg6: time kernel (0=n0, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 8 4 256) \n"
<< "--stride: tensor strides (e.g, --stride 1024 256 1)\n"
<< "--reduce: to-reduce dimensions (e.g, --reduce 2)\n"
<< "--alpha: alpha scaling value\n"
<< "--beta: beta scaling value\n"
<< std::endl;
}
int profile_normalization(int argc, char* argv[])
{
if(argc <= 2)
{
print_help();
return 0;
}
ArgParser arg_parser;
// short unnamed options
const NormType norm_type = arg_parser.norm_dict[argv[1]];
const NormDataType data_type = static_cast<NormDataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
const std::vector<index_t> stride = arg_parser.long_opts["stride"];
const std::vector<index_t> reduce = arg_parser.long_opts["reduce"];
const index_t alpha =
arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
// hijack main() for quick debugging
// int main(int argc, char* argv[])
// {
// profile_normalization(argc, argv);
// return 0;
// }
profiler/src/profiler.cpp
View file @ 6ef4e211
// 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 <cstring>
#include "profiler/include/profile_convnd_fwd.hpp"
int profile_gemm(int, char*[]);
int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_splitk(int, char*[]);
int profile_gemm_bilinear(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_convnd_fwd(int argc, char* argv[]);
int profile_convnd_bwd_data(int, char*[], int);
int profile_reduce(int, char*[]);
int profile_conv_bwd_weight(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_normalization(int, char*[]);
int profile_reduce(int, char*[]);
static void print_helper_message()
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_reduce: GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n");
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_splitk: Split-K GEMM\n"
" gemm_bilinear: GEMM+Bilinear\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" gemm_reduce: GEMM+Reduce\n"
" gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
" batched_gemm: Batched GEMM\n"
" batched_gemm_reduce: Batched GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
" reduce: Reduce\n");
// clang-format on
}
......@@ -60,17 +57,17 @@ int main(int argc, char* argv[])
{
return profile_gemm(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_2d") == 0)
else if(strcmp(argv[1], "gemm_splitk") == 0)
{
return profile_gemm_bias_2d(argc, argv);
return profile_gemm_splitk(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu") == 0)
else if(strcmp(argv[1], "gemm_bilinear") == 0)
{
return profile_gemm_bias_relu(argc, argv);
return profile_gemm_bilinear(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu_add") == 0)
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
return profile_gemm_bias_relu_add(argc, argv);
return profile_gemm_add_add_fastgelu(argc, argv);
}
else if(strcmp(argv[1], "gemm_reduce") == 0)
{
......@@ -94,7 +91,7 @@ int main(int argc, char* argv[])
}
else if(strcmp(argv[1], "conv_fwd") == 0)
{
return ck::profiler::profile_convnd_fwd(argc, argv);
return profile_convnd_fwd(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0)
{
......@@ -116,17 +113,18 @@ int main(int argc, char* argv[])
{
return profile_convnd_bwd_data(argc, argv, 3);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "conv2d_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_gemm_add_add_fastgelu(argc, argv);
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 ||
strcmp(argv[1], "softmax") == 0)
{
return profile_normalization(argc, argv);
}
else
{
......
script/docker-rocm4.1.sh deleted 100755 → 0
View file @ b0a2afb9
WORKSPACE=$1
echo "workspace: " $WORKSPACE
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v $WORKSPACE:/root/workspace \
rocm/tensorflow:rocm4.1-tf1.15-dev \
/bin/bash
#--network host \
script/docker-rocm4.3.1.sh deleted 100755 → 0
View file @ b0a2afb9
WORKSPACE=$1
echo "workspace: " $WORKSPACE
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v $WORKSPACE:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
#--network host \
test/CMakeLists.txt
View file @ 6ef4e211
......@@ -13,6 +13,7 @@ function(add_test_executable TEST_NAME)
add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> )
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
endfunction(add_test_executable TEST_NAME)
include(GoogleTest)
......@@ -26,6 +27,7 @@ function(add_gtest_executable TEST_NAME)
target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
gtest_discover_tests(${TEST_NAME})
rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
endfunction(add_gtest_executable TEST_NAME)
......
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