Skip to content

GitLab

Commit 6dfb4e78 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into cpu_avx2

parents 397a68f2 1ced00a5
Hide whitespace changes
Inline Side-by-side
Showing with 371 additions and 987 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.cpp deleted 100644 → 0
View file @ 397a68f2
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);
// Will be moved to use MultiBlockAtomicAdd
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i32_i8.cpp deleted 100644 → 0
View file @ 397a68f2
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.cpp deleted 100644 → 0
View file @ 397a68f2
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 6dfb4e78
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/include/ck/host_utility
${PROJECT_SOURCE_DIR}/include/ck/tensor_description
${PROJECT_SOURCE_DIR}/include/ck/tensor
${PROJECT_SOURCE_DIR}/include/ck/problem_transform
......
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ 6dfb4e78
......@@ -17,11 +17,21 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -119,19 +129,25 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, 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;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
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 d1_element_op = D1ElementOp{};
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<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
......@@ -155,15 +171,15 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
UnarySquareElementOp{}(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
......@@ -180,6 +196,9 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_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()));
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
......@@ -241,8 +260,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
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<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -252,7 +270,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/include/profile_conv_bwd_data_impl.hpp deleted 100644 → 0
View file @ 397a68f2
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
void profile_conv_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<InDataType> in_n_c_hi_wi_device_result(
f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvBwdDataInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvBwdDataInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdDataNoOpPtr =
ck::tensor_operation::device::DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamControl{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
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, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", out_n_k_ho_wo.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_host : ", in_n_c_hi_wi_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_device: ", in_n_c_hi_wi_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_impl.hpp
View file @ 6dfb4e78
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "check_err.hpp"
#include "config.hpp"
......@@ -42,14 +44,10 @@ void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<De
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_int8_int8_int8_mk_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_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>&);
......@@ -74,6 +72,21 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<Devic
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
......@@ -85,6 +98,7 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -125,7 +139,11 @@ void profile_gemm_impl(int do_verification,
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
// 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 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
......@@ -174,6 +192,9 @@ void profile_gemm_impl(int do_verification,
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);
}
......@@ -192,6 +213,9 @@ void profile_gemm_impl(int do_verification,
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);
}
......@@ -210,6 +234,9 @@ void profile_gemm_impl(int do_verification,
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);
}
......@@ -228,6 +255,9 @@ void profile_gemm_impl(int do_verification,
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);
}
......@@ -250,6 +280,9 @@ void profile_gemm_impl(int do_verification,
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);
}
......@@ -268,6 +301,9 @@ void profile_gemm_impl(int do_verification,
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);
......@@ -289,6 +325,9 @@ void profile_gemm_impl(int do_verification,
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);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
......@@ -307,6 +346,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
......@@ -353,28 +395,40 @@ void profile_gemm_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(gemm_ptrs);
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_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs);
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
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_int8_int8_int8_km_kn_mn_instances(gemm_ptrs);
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
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_int8_int8_int8_km_nk_mn_instances(gemm_ptrs);
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);
}
}
......@@ -458,8 +512,14 @@ void profile_gemm_impl(int do_verification,
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, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -491,6 +551,7 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
......@@ -523,12 +584,50 @@ void profile_gemm_impl(int do_verification,
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
std::cout << gemm_ptr->GetTypeString() << " does not support this GEMM 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_gemm_name << std::endl;
}
} // namespace profiler
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 6dfb4e78
......@@ -16,11 +16,22 @@ 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::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -112,24 +123,37 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, 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;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
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 d1_element_op = D1ElementOp{};
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<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
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{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -141,19 +165,23 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
d1_element_op(d1_val, d0_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);
}
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);
}
......@@ -165,6 +193,9 @@ bool profile_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_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()));
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......@@ -226,8 +257,7 @@ bool profile_gemm_reduce_impl(int do_verification,
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<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -237,7 +267,8 @@ bool profile_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op);
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ 6dfb4e78
......@@ -43,6 +43,7 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
......@@ -271,6 +272,7 @@ void profile_grouped_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
......
profiler/include/profile_reduce_impl.hpp
View file @ 6dfb4e78
......@@ -5,74 +5,77 @@
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_reduction.hpp"
#include "host_common_util.hpp"
#include "host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
template <int Rank, int NumReduceDim, int ReduceOpId, int NanOpt, int IndicesOpt>
template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription
{
static constexpr int Rank_ = Rank;
static constexpr int NumReduceDim_ = NumReduceDim;
static constexpr int ReduceOpId_ = ReduceOpId;
static constexpr int NanOpt_ = NanOpt;
static constexpr int IndicesOpt_ = IndicesOpt;
static constexpr int PropagateNan_ = PropagateNan;
static constexpr int UseIndex_ = UseIndex;
};
using reduce_description_instances = std::tuple<ReduceDescription<4, 3, 0, 0, 0>, // for ADD
ReduceDescription<4, 4, 0, 0, 0>,
ReduceDescription<4, 1, 0, 0, 0>,
ReduceDescription<2, 1, 0, 0, 0>,
ReduceDescription<4, 3, 5, 0, 0>, // for AVG
ReduceDescription<4, 4, 5, 0, 0>,
ReduceDescription<4, 1, 5, 0, 0>,
ReduceDescription<2, 1, 5, 0, 0>,
ReduceDescription<4, 3, 7, 0, 0>, // for NORM2
ReduceDescription<4, 4, 7, 0, 0>,
ReduceDescription<4, 1, 7, 0, 0>,
ReduceDescription<2, 1, 7, 0, 0>,
ReduceDescription<4, 3, 2, 0, 0>, // for MIN
ReduceDescription<4, 4, 2, 0, 0>,
ReduceDescription<4, 1, 2, 0, 0>,
ReduceDescription<2, 1, 2, 0, 0>,
ReduceDescription<4, 3, 3, 0, 0>, // for MAX
ReduceDescription<4, 4, 3, 0, 0>,
ReduceDescription<4, 1, 3, 0, 0>,
ReduceDescription<2, 1, 3, 0, 0>,
ReduceDescription<4, 3, 4, 0, 0>, // for AMAX
ReduceDescription<4, 4, 4, 0, 0>,
ReduceDescription<4, 1, 4, 0, 0>,
ReduceDescription<2, 1, 4, 0, 0>,
ReduceDescription<4, 3, 2, 0, 1>, // for MIN
ReduceDescription<4, 4, 2, 0, 1>,
ReduceDescription<4, 1, 2, 0, 1>,
ReduceDescription<2, 1, 2, 0, 1>,
ReduceDescription<4, 3, 3, 0, 1>, // for MAX
ReduceDescription<4, 4, 3, 0, 1>,
ReduceDescription<4, 1, 3, 0, 1>,
ReduceDescription<2, 1, 3, 0, 1>,
ReduceDescription<4, 3, 4, 0, 1>, // for AMAX
ReduceDescription<4, 4, 4, 0, 1>,
ReduceDescription<4, 1, 4, 0, 1>,
ReduceDescription<2, 1, 4, 0, 1>>;
using reduce_description_instances =
std::tuple<ReduceDescription<4, 3, 0, false, false>, // for ADD
ReduceDescription<4, 4, 0, false, false>,
ReduceDescription<4, 1, 0, false, false>,
ReduceDescription<2, 1, 0, false, false>,
ReduceDescription<4, 3, 5, false, false>, // for AVG
ReduceDescription<4, 4, 5, false, false>,
ReduceDescription<4, 1, 5, false, false>,
ReduceDescription<2, 1, 5, false, false>,
ReduceDescription<4, 3, 7, false, false>, // for NORM2
ReduceDescription<4, 4, 7, false, false>,
ReduceDescription<4, 1, 7, false, false>,
ReduceDescription<2, 1, 7, false, false>,
ReduceDescription<4, 3, 2, false, false>, // for MIN
ReduceDescription<4, 4, 2, false, false>,
ReduceDescription<4, 1, 2, false, false>,
ReduceDescription<2, 1, 2, false, false>,
ReduceDescription<4, 3, 3, false, false>, // for MAX
ReduceDescription<4, 4, 3, false, false>,
ReduceDescription<4, 1, 3, false, false>,
ReduceDescription<2, 1, 3, false, false>,
ReduceDescription<4, 3, 4, false, false>, // for AMAX
ReduceDescription<4, 4, 4, false, false>,
ReduceDescription<4, 1, 4, false, false>,
ReduceDescription<2, 1, 4, false, false>,
ReduceDescription<4, 3, 2, false, true>, // for MIN
ReduceDescription<4, 4, 2, false, true>,
ReduceDescription<4, 1, 2, false, true>,
ReduceDescription<2, 1, 2, false, true>,
ReduceDescription<4, 3, 3, false, true>, // for MAX
ReduceDescription<4, 4, 3, false, true>,
ReduceDescription<4, 1, 3, false, true>,
ReduceDescription<2, 1, 3, false, true>,
ReduceDescription<4, 3, 4, false, true>, // for AMAX
ReduceDescription<4, 4, 4, false, true>,
ReduceDescription<4, 1, 4, false, true>,
ReduceDescription<2, 1, 4, false, true>>;
template <typename DescriptionType>
bool description_match(const DescriptionType& description,
int Rank,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt)
bool PropagateNan,
bool UseIndex)
{
if(description.Rank_ != Rank || description.ReduceOpId_ != static_cast<int>(ReduceOpId) ||
description.NanOpt_ != static_cast<int>(NanOpt) ||
description.IndicesOpt_ != static_cast<int>(IndicesOpt))
description.PropagateNan_ != static_cast<int>(PropagateNan) ||
description.UseIndex_ != static_cast<int>(UseIndex))
return (false);
if(DescriptionType::NumReduceDim_ != reduceDims.size())
......@@ -116,46 +119,16 @@ static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduce
return invariantDims;
};
template <typename T>
static void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
{
std::ofstream outFile(fileName, std::ios::binary);
if(outFile)
{
outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
outFile.close();
std::cout << "Write output to file " << fileName << std::endl;
}
else
{
std::cout << "Could not open file " << fileName << " for writing" << std::endl;
}
};
// map the data type used by the GPU kernels to the corresponding type used by the host codes
template <typename InType>
struct type_mapping
{
using OutType = InType;
};
template <>
struct type_mapping<ck::half_t>
{
using OutType = half_float::half;
};
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
void profile_reduce_impl_impl(bool do_verification,
bool PropagateNan,
bool UseIndex>
bool profile_reduce_impl_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
bool time_kernel,
const std::vector<size_t>& inLengths,
......@@ -165,16 +138,13 @@ void profile_reduce_impl_impl(bool do_verification,
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
constexpr bool PropagateNan = (NanOpt == NanPropagation::PROPAGATE_NAN);
constexpr bool OutputIndex = (op_support_indices && UseIndex);
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add =
......@@ -195,8 +165,7 @@ void profile_reduce_impl_impl(bool do_verification,
(op_support_indices && !std::is_same<AccDataType, float>::value);
// 1) The indices can only be used when the reduction operation is indexable
constexpr bool invalid_reduce_3 =
(!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
// 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
// 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
......@@ -219,6 +188,8 @@ void profile_reduce_impl_impl(bool do_verification,
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
bool pass = true;
if constexpr(!invalid_reduce)
{
Tensor<InDataType> in(inLengths);
......@@ -282,42 +253,26 @@ void profile_reduce_impl_impl(bool do_verification,
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0;
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
float best_avg_time = 0;
float best_gb_per_sec = 0;
using InElementwiseOperation_0 =
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
InElementwiseOperation;
using AccElementwiseOperation_0 =
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using InElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
InElementwiseOperation;
using AccElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
AccElementwiseOperation;
using InElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType,
AccDataType,
......@@ -325,8 +280,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
......@@ -334,8 +289,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
if constexpr(use_atomic_add)
{
......@@ -345,35 +300,11 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
}
else
{
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
};
// used for secondary reduction
if constexpr(!use_atomic_add)
{
add_device_reduce_instance_blockwise_second_call<AccDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
};
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
if(reduce0_ptrs.empty())
{
throw std::runtime_error("Wrong! No device REDUCE instance found");
};
......@@ -383,31 +314,34 @@ void profile_reduce_impl_impl(bool do_verification,
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
};
const auto i_inLengths = to_int_vector(inLengths);
const auto i_inStrides = to_int_vector(inStrides);
const auto i_outLengths = to_int_vector(outLengths);
const auto i_outStrides = to_int_vector(outStrides);
std::vector<ck::index_t> i_inLengths;
std::vector<ck::index_t> i_inStrides;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
for(auto& reduce_ptr : reduce0_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(
static_cast<int32_t>(reduce_total_length));
InElementwiseOperation in_elementwise_op(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation acc_elementwise_op(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
......@@ -417,11 +351,11 @@ void profile_reduce_impl_impl(bool do_verification,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
in_elementwise_op,
acc_elementwise_op);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
......@@ -439,8 +373,9 @@ void profile_reduce_impl_impl(bool do_verification,
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
if(time_kernel)
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< reduce_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
......@@ -450,22 +385,24 @@ void profile_reduce_impl_impl(bool do_verification,
if(do_verification)
{
bool single_pass;
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
single_pass = ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
if(OutputIndex)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
single_pass = single_pass &&
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
if(do_log)
if(!single_pass)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",") << std::endl;
};
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
}
pass = pass && single_pass;
};
if(do_dumpout)
......@@ -474,7 +411,7 @@ void profile_reduce_impl_impl(bool do_verification,
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
if(OutputIndex)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
......@@ -486,158 +423,34 @@ void profile_reduce_impl_impl(bool do_verification,
};
};
for(auto& reduce_ptr : reduce1_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_1 acc_elementwise_op_1(
static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_1,
acc_elementwise_op_1);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
std::string reduce_name = reduce_ptr->GetTypeString();
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
std::vector<int> inStrides2{inLengths2[1], 1};
for(auto& reduce2_ptr : reduce2_ptrs)
{
InElementwiseOperation_2 in_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_2 acc_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
auto argument2_ptr =
reduce2_ptr->MakeArgumentPointer(inLengths2,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_2,
acc_elementwise_op_2);
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
continue;
std::string reduce2_name = reduce2_ptr->GetTypeString();
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
float avg_time_2 =
invoker2_ptr->Run(argument2_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes_2 =
static_cast<size_t>(inLengths2[0]) * inLengths2[1] * sizeof(AccDataType);
float gb_per_sec = (num_bytes + num_bytes_2) / 1.E6 / (avg_time + avg_time_2);
std::cout << "Perf: " << (avg_time + avg_time_2) << " ms, " << gb_per_sec
<< " GB/s, " << reduce_name << " => " << reduce2_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
best_avg_time = avg_time + avg_time_2;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
};
if(do_log)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",")
<< std::endl;
}
}
if(do_dumpout)
{
dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
out_indices.mDesc.GetElementSize());
dumpBufferToFile("dump_indices_host.bin",
out_indices_ref.mData.data(),
out_indices_ref.mDesc.GetElementSize());
};
};
};
};
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
if(time_kernel)
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
}
else
{
std::cout << "The requested reduction operation is not supported, please check !!!"
<< std::endl;
};
return pass;
};
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_reduce_impl(bool do_verification,
bool profile_reduce_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt,
bool PropagateNan,
bool UseIndex,
float alpha,
float beta)
{
bool matched = false;
bool pass = true;
using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances;
......@@ -651,29 +464,30 @@ void profile_reduce_impl(bool do_verification,
using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
if(!description_match(
descType{}, inLengths.size(), reduceDims, ReduceOpId, NanOpt, IndicesOpt))
descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
return;
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<NanPropagation>(descType::NanOpt_),
static_cast<ReduceTensorIndices>(descType::IndicesOpt_)>(
do_verification,
init_method,
do_log,
do_dumpout,
time_kernel,
inLengths,
reduceDims,
alpha,
beta);
pass = pass &&
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<bool>(descType::PropagateNan_),
static_cast<bool>(descType::UseIndex_)>(do_verification,
init_method,
do_dumpout,
time_kernel,
inLengths,
reduceDims,
alpha,
beta);
matched = true;
});
return pass;
};
} // namespace profiler
......
profiler/src/profile_batched_gemm.cpp
View file @ 6dfb4e78
......@@ -396,5 +396,5 @@ int profile_batched_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_batched_gemm_reduce.cpp
View file @ 6dfb4e78
......@@ -149,5 +149,5 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_bwd_data.cpp deleted 100644 → 0
View file @ 397a68f2
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv_bwd_data_impl.hpp"
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int profile_conv_bwd_data(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_bwd: BackwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
float,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::BF16_BF16_BF16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
uint16_t,
uint16_t,
uint16_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::INT8_INT8_INT8 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
}
profiler/src/profile_conv_bwd_weight.cpp
View file @ 6dfb4e78
......@@ -142,5 +142,5 @@ int profile_conv_bwd_weight(int argc, char* argv[])
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu.cpp
View file @ 6dfb4e78
......@@ -110,5 +110,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View file @ 6dfb4e78
......@@ -111,5 +111,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu_atomic_add.cpp
View file @ 6dfb4e78
......@@ -112,5 +112,5 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_convnd_fwd.cpp
View file @ 6dfb4e78
......@@ -347,5 +347,5 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
std::to_string(num_dim_spatial));
}
return 1;
return 0;
}
profiler/src/profile_gemm.cpp
View file @ 6dfb4e78
......@@ -68,6 +68,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -88,6 +89,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -108,6 +110,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -128,6 +131,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -146,6 +150,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
......@@ -166,6 +171,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
......@@ -186,6 +192,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -206,6 +213,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -228,6 +236,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -248,6 +257,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -268,6 +278,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -288,6 +299,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -308,6 +320,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -328,6 +341,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -348,6 +362,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -368,6 +383,7 @@ int profile_gemm(int argc, char* argv[])
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>(
......@@ -388,5 +404,5 @@ int profile_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_gemm_bias_2d.cpp
View file @ 6dfb4e78
......@@ -252,5 +252,5 @@ int profile_gemm_bias_2d(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment