Skip to content

GitLab

"official/vision/dataloaders/utils_test.py" did not exist on "4242b54920a807db38aa2b0fb97d884433a7ee26"
Commit b134b7d6 authored by carlushuang's avatar carlushuang
Browse files

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

parents 090ba885 9f71ff48
Hide whitespace changes
Inline Side-by-side
Showing with 349 additions and 174 deletions
library/include/ck/library/host_tensor/host_reduction.hpp
View file @ b134b7d6
......@@ -211,7 +211,8 @@ struct ReductionHost
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
IndexDataType accuIndex = 0;
for(IndexDataType i = 0; i < reduce_dim_indexes.size(); i++)
for(IndexDataType i = 0; i < ck::type_convert<IndexDataType>(reduce_dim_indexes.size());
i++)
{
auto offset_reduce =
get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
......@@ -246,7 +247,9 @@ struct ReductionHost
auto offset_invariant =
get_offset_from_index<NumInvariantDim>(invariantStrides, invariant_index);
for(IndexDataType i = 0; i < reduce_dim_indexes.size(); i++)
for(IndexDataType i = 0;
i < ck::type_convert<IndexDataType>(reduce_dim_indexes.size());
i++)
{
auto offset_reduce =
get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
......
library/include/ck/library/host_tensor/host_tensor.hpp
View file @ b134b7d6
......@@ -154,7 +154,7 @@ struct ParallelTensorFunctor
{
std::array<std::size_t, NDIM> indices;
for(int idim = 0; idim < NDIM; ++idim)
for(std::size_t idim = 0; idim < NDIM; ++idim)
{
indices[idim] = i / mStrides[idim];
i -= indices[idim] * mStrides[idim];
......@@ -316,7 +316,7 @@ float check_error(const Tensor<T>& ref, const Tensor<T>& result)
constexpr float eps = 1e-10;
for(int i = 0; i < ref.mData.size(); ++i)
for(std::size_t i = 0; i < ref.mData.size(); ++i)
{
float ref_v = ck::type_convert<float>(ref.mData[i]);
float result_v = ck::type_convert<float>(result.mData[i]);
......
library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
View file @ b134b7d6
......@@ -84,7 +84,8 @@ struct ReferenceBatchedGemm : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp
View file @ b134b7d6
......@@ -70,18 +70,25 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
constexpr auto I1 = Number<1>{};
auto f_kcyx = [&](auto k, auto c, auto y, auto x) {
float v_acc = 0;
for(int n = 0; n < arg.out_n_k_ho_wo_.mDesc.GetLengths()[0]; ++n)
for(std::size_t n = 0; n < arg.out_n_k_ho_wo_.mDesc.GetLengths()[0]; ++n)
{
for(int ho = 0; ho < arg.out_n_k_ho_wo_.mDesc.GetLengths()[2]; ++ho)
for(std::size_t ho = 0; ho < arg.out_n_k_ho_wo_.mDesc.GetLengths()[2]; ++ho)
{
int hi = ho * arg.conv_strides_[I0] + y * arg.conv_dilations_[I0] -
arg.in_left_pads_[I0];
for(int wo = 0; wo < arg.out_n_k_ho_wo_.mDesc.GetLengths()[3]; ++wo)
auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[I0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[I0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[I0]);
for(std::size_t wo = 0; wo < arg.out_n_k_ho_wo_.mDesc.GetLengths()[3]; ++wo)
{
int wi = wo * arg.conv_strides_[I1] + x * arg.conv_dilations_[I1] -
arg.in_left_pads_[I1];
if(hi >= 0 && hi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] && wi >= 0 &&
wi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[I1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[I1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[I1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
wi >= 0 &&
ck::type_convert<std::size_t>(wi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
{
float v_out;
float v_in;
......@@ -114,7 +121,8 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /*stream_config*/ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View file @ b134b7d6
......@@ -78,15 +78,18 @@ struct ReferenceConvBwdData : public device::BaseOperator
AccDataType v_acc = 0;
for(int x = 0; x < X; ++x)
for(std::size_t x = 0; x < X; ++x)
{
int w_tmp = wi + arg.in_left_pads_[0] - x * arg.conv_dilations_[0];
auto w_tmp = ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[0]);
if(w_tmp % arg.conv_strides_[0] == 0)
{
int wo = w_tmp / arg.conv_strides_[0];
if(wo >= 0 && wo < Wo)
auto wo = ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
{
for(int k = 0; k < K; ++k)
for(std::size_t k = 0; k < K; ++k)
{
AccDataType v_out = 0;
AccDataType v_wei = 0;
......@@ -128,24 +131,32 @@ struct ReferenceConvBwdData : public device::BaseOperator
AccDataType v_acc = 0;
for(int y = 0; y < Y; ++y)
for(std::size_t y = 0; y < Y; ++y)
{
int h_tmp = hi + arg.in_left_pads_[0] - y * arg.conv_dilations_[0];
auto h_tmp = ck::type_convert<ck::long_index_t>(hi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]);
if(h_tmp % arg.conv_strides_[0] == 0)
{
int ho = h_tmp / arg.conv_strides_[0];
if(ho >= 0 && ho < Ho)
auto ho = ck::type_convert<ck::long_index_t>(h_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
{
for(int x = 0; x < X; ++x)
for(std::size_t x = 0; x < X; ++x)
{
int w_tmp =
wi + arg.in_left_pads_[1] - x * arg.conv_dilations_[1];
auto w_tmp =
ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]) -
ck::type_convert<ck::long_index_t>(x *
arg.conv_dilations_[1]);
if(w_tmp % arg.conv_strides_[1] == 0)
{
int wo = w_tmp / arg.conv_strides_[1];
if(wo >= 0 && wo < Wo)
auto wo = ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[1]);
if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
{
for(int k = 0; k < K; ++k)
for(std::size_t k = 0; k < K; ++k)
{
AccDataType v_out = 0;
AccDataType v_wei = 0;
......@@ -194,33 +205,49 @@ struct ReferenceConvBwdData : public device::BaseOperator
AccDataType v_acc = 0;
for(int z = 0; z < Z; ++z)
for(std::size_t z = 0; z < Z; ++z)
{
int d_tmp = di + arg.in_left_pads_[0] - z * arg.conv_dilations_[0];
auto d_tmp = ck::type_convert<ck::long_index_t>(di) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]) -
ck::type_convert<ck::long_index_t>(z * arg.conv_dilations_[0]);
if(d_tmp % arg.conv_strides_[0] == 0)
{
int do_ = d_tmp / arg.conv_strides_[0];
if(do_ >= 0 && do_ < Do)
auto do_ = ck::type_convert<ck::long_index_t>(d_tmp) /
ck::type_convert<ck::long_index_t>(arg.conv_strides_[0]);
if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
{
for(int y = 0; y < Y; ++y)
for(std::size_t y = 0; y < Y; ++y)
{
int h_tmp =
hi + arg.in_left_pads_[1] - y * arg.conv_dilations_[1];
auto h_tmp =
ck::type_convert<ck::long_index_t>(hi) +
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]) -
ck::type_convert<ck::long_index_t>(y *
arg.conv_dilations_[1]);
if(h_tmp % arg.conv_strides_[1] == 0)
{
int ho = h_tmp / arg.conv_strides_[1];
if(ho >= 0 && ho < Ho)
auto ho = ck::type_convert<ck::long_index_t>(h_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[1]);
if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
{
for(int x = 0; x < X; ++x)
for(std::size_t x = 0; x < X; ++x)
{
int w_tmp = wi + arg.in_left_pads_[2] -
x * arg.conv_dilations_[2];
auto w_tmp =
ck::type_convert<ck::long_index_t>(wi) +
ck::type_convert<ck::long_index_t>(
arg.in_left_pads_[2]) -
ck::type_convert<ck::long_index_t>(
x * arg.conv_dilations_[2]);
if(w_tmp % arg.conv_strides_[2] == 0)
{
int wo = w_tmp / arg.conv_strides_[2];
if(wo >= 0 && wo < Wo)
auto wo =
ck::type_convert<ck::long_index_t>(w_tmp) /
ck::type_convert<ck::long_index_t>(
arg.conv_strides_[2]);
if(wo >= 0 &&
ck::type_convert<std::size_t>(wo) < Wo)
{
for(int k = 0; k < K; ++k)
for(std::size_t k = 0; k < K; ++k)
{
AccDataType v_out = 0;
AccDataType v_wei = 0;
......@@ -264,7 +291,8 @@ struct ReferenceConvBwdData : public device::BaseOperator
}
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
View file @ b134b7d6
#ifndef REFERENCE_CONV_FWD_HPP
#define REFERENCE_CONV_FWD_HPP
#pragma once
#include <iostream>
#include <type_traits>
#include <sstream>
#include "stream_config.hpp"
#include "device_base.hpp"
#include "host_tensor.hpp"
......@@ -88,13 +89,16 @@ struct ReferenceConvFwd : public device::BaseOperator
auto f_ncw = [&](auto n, auto k, auto wo) {
float v_acc = 0;
for(int c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
for(std::size_t c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
{
for(int x = 0; x < arg.weight_.mDesc.GetLengths()[2]; ++x)
for(std::size_t x = 0; x < arg.weight_.mDesc.GetLengths()[2]; ++x)
{
int wi = wo * arg.conv_strides_[0] + x * arg.conv_dilations_[0] -
arg.in_left_pads_[0];
if(wi >= 0 && wi < arg.input_.mDesc.GetLengths()[2])
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
if(wi >= 0 &&
ck::type_convert<std::size_t>(wi) < arg.input_.mDesc.GetLengths()[2])
{
float v_in;
float v_wei;
......@@ -128,18 +132,26 @@ struct ReferenceConvFwd : public device::BaseOperator
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
float v_acc = 0;
for(int c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
for(std::size_t c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < arg.weight_.mDesc.GetLengths()[2]; ++y)
for(std::size_t y = 0; y < arg.weight_.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * arg.conv_strides_[0] + y * arg.conv_dilations_[0] -
arg.in_left_pads_[0];
for(int x = 0; x < arg.weight_.mDesc.GetLengths()[3]; ++x)
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t x = 0; x < arg.weight_.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * arg.conv_strides_[1] + x * arg.conv_dilations_[1] -
arg.in_left_pads_[1];
if(hi >= 0 && hi < arg.input_.mDesc.GetLengths()[2] && wi >= 0 &&
wi < arg.input_.mDesc.GetLengths()[3])
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) <
arg.input_.mDesc.GetLengths()[2] &&
wi >= 0 &&
ck::type_convert<std::size_t>(wi) <
arg.input_.mDesc.GetLengths()[3])
{
float v_in;
float v_wei;
......@@ -174,23 +186,37 @@ struct ReferenceConvFwd : public device::BaseOperator
auto f_nchw = [&](auto n, auto k, auto d_o, auto ho, auto wo) {
float v_acc = 0;
for(int c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
for(std::size_t c = 0; c < arg.weight_.mDesc.GetLengths()[1]; ++c)
{
for(int z = 0; z < arg.weight_.mDesc.GetLengths()[2]; ++z)
for(std::size_t z = 0; z < arg.weight_.mDesc.GetLengths()[2]; ++z)
{
int di = d_o * arg.conv_strides_[0] + z * arg.conv_dilations_[0] -
arg.in_left_pads_[0];
for(int y = 0; y < arg.weight_.mDesc.GetLengths()[3]; ++y)
auto di =
ck::type_convert<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(z * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t y = 0; y < arg.weight_.mDesc.GetLengths()[3]; ++y)
{
int hi = ho * arg.conv_strides_[1] + y * arg.conv_dilations_[1] -
arg.in_left_pads_[1];
for(int x = 0; x < arg.weight_.mDesc.GetLengths()[4]; ++x)
auto hi =
ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
for(std::size_t x = 0; x < arg.weight_.mDesc.GetLengths()[4]; ++x)
{
int wi = wo * arg.conv_strides_[2] +
x * arg.conv_dilations_[2] - arg.in_left_pads_[2];
if(di >= 0 && di < arg.input_.mDesc.GetLengths()[2] &&
hi >= 0 && hi < arg.input_.mDesc.GetLengths()[3] &&
wi >= 0 && wi < arg.input_.mDesc.GetLengths()[4])
auto wi =
ck::type_convert<ck::long_index_t>(wo *
arg.conv_strides_[2]) +
ck::type_convert<ck::long_index_t>(x *
arg.conv_dilations_[2]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[2]);
if(di >= 0 &&
ck::type_convert<std::size_t>(di) <
arg.input_.mDesc.GetLengths()[2] &&
hi >= 0 &&
ck::type_convert<std::size_t>(hi) <
arg.input_.mDesc.GetLengths()[3] &&
wi >= 0 &&
ck::type_convert<std::size_t>(wi) <
arg.input_.mDesc.GetLengths()[4])
{
float v_in;
float v_wei;
......@@ -226,7 +252,8 @@ struct ReferenceConvFwd : public device::BaseOperator
}
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /*stream_config*/ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......@@ -286,4 +313,3 @@ struct ReferenceConvFwd : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
View file @ b134b7d6
......@@ -73,18 +73,25 @@ struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
float v_acc = 0;
for(int c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * arg.conv_strides_[0] + y * arg.conv_dilations_[0] -
arg.in_left_pads_[0];
for(int x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * arg.conv_strides_[1] + x * arg.conv_dilations_[1] -
arg.in_left_pads_[1];
if(hi >= 0 && hi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] && wi >= 0 &&
wi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
wi >= 0 &&
ck::type_convert<std::size_t>(wi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
{
float v_in;
float v_wei;
......@@ -117,7 +124,8 @@ struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
View file @ b134b7d6
......@@ -76,18 +76,25 @@ struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
float v_acc = 0;
for(int c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * arg.conv_strides_[0] + y * arg.conv_dilations_[0] -
arg.in_left_pads_[0];
for(int x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * arg.conv_strides_[1] + x * arg.conv_dilations_[1] -
arg.in_left_pads_[1];
if(hi >= 0 && hi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] && wi >= 0 &&
wi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
auto wi =
ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
if(hi >= 0 &&
ck::type_convert<std::size_t>(hi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
wi >= 0 &&
ck::type_convert<std::size_t>(wi) <
arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
{
float v_in;
float v_wei;
......@@ -123,7 +130,8 @@ struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /*stream_config*/ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ b134b7d6
#ifndef REFERENCE_GEMM_HPP
#define REFERENCE_GEMM_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device_base.hpp"
......@@ -82,7 +80,8 @@ struct ReferenceGemm : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......@@ -129,4 +128,3 @@ struct ReferenceGemm : public device::BaseOperator
} // namespace host
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
View file @ b134b7d6
......@@ -82,7 +82,8 @@ struct ReferenceGemmBias2D : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
View file @ b134b7d6
......@@ -85,7 +85,8 @@ struct ReferenceGemmBiasActivation : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
View file @ b134b7d6
......@@ -91,7 +91,8 @@ struct ReferenceGemmBiasActivationAdd : public device::BaseOperator
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
float Run(const device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
......
library/include/ck/library/utility/conv_fwd_util.hpp → library/include/ck/library/utility/conv_util.hpp
View file @ b134b7d6
......@@ -146,19 +146,19 @@ struct ConvParams
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial;
ck::index_t N;
ck::index_t K;
ck::index_t C;
ck::index_t num_dim_spatial_;
ck::index_t N_;
ck::index_t K_;
ck::index_t C_;
std::vector<ck::index_t> filter_spatial_lengths;
std::vector<ck::index_t> input_spatial_lengths;
std::vector<ck::index_t> filter_spatial_lengths_;
std::vector<ck::index_t> input_spatial_lengths_;
std::vector<ck::index_t> conv_filter_strides;
std::vector<ck::index_t> conv_filter_dilations;
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;
std::vector<ck::index_t> input_left_pads_;
std::vector<ck::index_t> input_right_pads_;
std::vector<ck::index_t> GetOutputSpatialLengths() const;
};
......@@ -268,10 +268,10 @@ void run_reference_convolution_forward(const ConvParams& params,
auto ref_argument = ref_conv.MakeArgument(input,
weights,
output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
......@@ -437,17 +437,17 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual InTensorsTuple GetInputTensors() const override
{
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params_.N),
static_cast<std::size_t>(params_.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params_.input_spatial_lengths),
std::end(params_.input_spatial_lengths));
std::begin(params_.input_spatial_lengths_),
std::end(params_.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params_.K),
static_cast<std::size_t>(params_.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params_.K_),
static_cast<std::size_t>(params_.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params_.filter_spatial_lengths),
std::end(params_.filter_spatial_lengths));
std::begin(params_.filter_spatial_lengths_),
std::end(params_.filter_spatial_lengths_));
auto input = std::make_unique<Tensor<InDataType>>(
get_host_tensor_descriptor(input_dims, InLayout{}));
......@@ -465,8 +465,8 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual TensorPtr<OutDataType> GetOutputTensor() const override
{
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params_.N),
static_cast<std::size_t>(params_.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths_),
std::end(output_spatial_lengths_));
......@@ -522,16 +522,16 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
static_cast<InDataType*>(in_device_buffers[0]->GetDeviceBuffer()),
static_cast<WeiDataType*>(in_device_buffers[1]->GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buffer->GetDeviceBuffer()),
params_.N,
params_.K,
params_.C,
params_.input_spatial_lengths,
params_.filter_spatial_lengths,
params_.N_,
params_.K_,
params_.C_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_,
params_.conv_filter_strides,
params_.conv_filter_dilations,
params_.input_left_pads,
params_.input_right_pads,
params_.conv_filter_strides_,
params_.conv_filter_dilations_,
params_.input_left_pads_,
params_.input_right_pads_,
InElementwiseOp{},
WeiElementwiseOp{},
OutElementwiseOp{});
......@@ -539,20 +539,20 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual std::size_t GetFlops() const override
{
return get_flops(params_.N,
params_.C,
params_.K,
params_.filter_spatial_lengths,
return get_flops(params_.N_,
params_.C_,
params_.K_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
virtual std::size_t GetBtype() const override
{
return get_btype<InDataType, WeiDataType, OutDataType>(params_.N,
params_.C,
params_.K,
params_.input_spatial_lengths,
params_.filter_spatial_lengths,
return get_btype<InDataType, WeiDataType, OutDataType>(params_.N_,
params_.C_,
params_.K_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
......
library/include/ck/library/utility/op_instance_engine.hpp
View file @ b134b7d6
......@@ -128,7 +128,7 @@ class OpInstanceRunEngine
template <typename OpInstancePtr>
ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
int nrepeat = 100,
bool time_kernel = false,
bool do_verification = false,
bool do_log = false)
{
......@@ -143,7 +143,7 @@ class OpInstanceRunEngine
if(op_ptr->IsSupportedArgument(argument.get()))
{
std::string op_name = op_ptr->GetTypeString();
float avg_time = invoker->Run(argument.get(), nrepeat);
float avg_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flops = op_instance_.GetFlops();
std::size_t num_btype = op_instance_.GetBtype();
......
library/src/host_tensor/CMakeLists.txt
View file @ b134b7d6
......@@ -10,10 +10,31 @@ set(HOST_TENSOR_SOURCE
host_tensor.cpp
)
add_library(host_tensor SHARED ${HOST_TENSOR_SOURCE})
add_library(host_tensor STATIC ${HOST_TENSOR_SOURCE})
add_library(composable_kernel::host_tensor ALIAS host_tensor)
target_compile_features(host_tensor PUBLIC)
set_target_properties(host_tensor PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(host_tensor SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
install(TARGETS host_tensor LIBRARY DESTINATION lib)
target_include_directories(host_tensor PUBLIC
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/host_tensor>"
)
install(TARGETS host_tensor
EXPORT host_tensorTargets
LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
)
install(EXPORT host_tensorTargets
FILE composable_kernelhost_tensorTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
clang_tidy_check(host_tensor)
library/src/host_tensor/device.cpp
View file @ b134b7d6
......@@ -3,7 +3,7 @@
DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
{
hipGetErrorString(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void* DeviceMem::GetDeviceBuffer() { return mpDeviceBuf; }
......@@ -12,18 +12,17 @@ std::size_t DeviceMem::GetBufferSize() { return mMemSize; }
void DeviceMem::ToDevice(const void* p)
{
hipGetErrorString(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
}
void DeviceMem::FromDevice(void* p)
{
hipGetErrorString(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
}
void DeviceMem::SetZero() { hipGetErrorString(hipMemset(mpDeviceBuf, 0, mMemSize)); }
void DeviceMem::SetZero() { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
DeviceMem::~DeviceMem() { hipGetErrorString(hipFree(mpDeviceBuf)); }
DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
DeviceAlignedMemCPU::DeviceAlignedMemCPU(std::size_t mem_size, std::size_t alignment)
: mMemSize(mem_size), mAlignment(alignment)
......@@ -68,32 +67,32 @@ struct KernelTimerImpl
{
KernelTimerImpl()
{
hipGetErrorString(hipEventCreate(&mStart));
hipGetErrorString(hipEventCreate(&mEnd));
hip_check_error(hipEventCreate(&mStart));
hip_check_error(hipEventCreate(&mEnd));
}
~KernelTimerImpl()
{
hipGetErrorString(hipEventDestroy(mStart));
hipGetErrorString(hipEventDestroy(mEnd));
hip_check_error(hipEventDestroy(mStart));
hip_check_error(hipEventDestroy(mEnd));
}
void Start()
{
hipGetErrorString(hipDeviceSynchronize());
hipGetErrorString(hipEventRecord(mStart, nullptr));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(mStart, nullptr));
}
void End()
{
hipGetErrorString(hipEventRecord(mEnd, nullptr));
hipGetErrorString(hipEventSynchronize(mEnd));
hip_check_error(hipEventRecord(mEnd, nullptr));
hip_check_error(hipEventSynchronize(mEnd));
}
float GetElapsedTime() const
{
float time;
hipGetErrorString(hipEventElapsedTime(&time, mStart, mEnd));
hip_check_error(hipEventElapsedTime(&time, mStart, mEnd));
return time;
}
......
library/src/host_tensor/host_tensor.cpp
View file @ b134b7d6
......@@ -25,7 +25,7 @@ std::size_t HostTensorDescriptor::GetElementSize() const
std::size_t HostTensorDescriptor::GetElementSpace() const
{
std::size_t space = 1;
for(int i = 0; i < mLens.size(); ++i)
for(std::size_t i = 0; i < mLens.size(); ++i)
{
space += (mLens[i] - 1) * mStrides[i];
}
......@@ -68,7 +68,7 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream
// FIXME: remove
void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst)
{
for(int i = 0; i < src.mData.size(); ++i)
for(std::size_t i = 0; i < src.mData.size(); ++i)
dst.mData[i] = ck::type_convert<float>(src.mData[i]);
}
#endif
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ b134b7d6
......@@ -11,6 +11,7 @@ include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/thread
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/element
${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
${PROJECT_SOURCE_DIR}/library/include/ck/library/host
${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance
${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance/gpu/reduce
${PROJECT_SOURCE_DIR}/external/include/half
......@@ -18,7 +19,7 @@ include_directories(BEFORE
function(add_instance_library INSTANCE_NAME)
message("adding instance ${INSTANCE_NAME}")
add_library(${INSTANCE_NAME} SHARED ${ARGN})
add_library(${INSTANCE_NAME} OBJECT ${ARGN})
target_compile_features(${INSTANCE_NAME} PUBLIC)
set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
endfunction(add_instance_library INSTANCE_NAME)
......@@ -41,3 +42,73 @@ add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_gemm)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(batched_gemm_reduce)
add_library(device_operations STATIC
$<TARGET_OBJECTS:device_conv1d_fwd_instance>
$<TARGET_OBJECTS:device_batched_gemm_instance>
$<TARGET_OBJECTS:device_conv2d_bwd_data_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_bias_relu_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_bias_relu_add_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_bias_relu_atomic_add_instance>
$<TARGET_OBJECTS:device_gemm_instance>
$<TARGET_OBJECTS:device_gemm_bias_relu_instance>
$<TARGET_OBJECTS:device_gemm_bias_relu_add_instance>
$<TARGET_OBJECTS:device_gemm_bias2d_instance>
$<TARGET_OBJECTS:device_reduce_instance>
$<TARGET_OBJECTS:device_convnd_bwd_data_instance>
$<TARGET_OBJECTS:device_grouped_gemm_instance>
$<TARGET_OBJECTS:device_conv2d_bwd_weight_instance>
$<TARGET_OBJECTS:device_batched_gemm_reduce_instance>
$<TARGET_OBJECTS:device_conv3d_fwd_instance>
device_conv2d.cpp
)
add_library(composablekernels::device_operations ALIAS device_operations)
set(DEV_OPS_INC_DIRS
${PROJECT_SOURCE_DIR}/include/ck/
${PROJECT_SOURCE_DIR}/library/include/ck/
${PROJECT_SOURCE_DIR}/external/include/
)
target_compile_features(device_operations PUBLIC)
set_target_properties(device_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(device_operations PUBLIC
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_description>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/problem_transform>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/grid>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/block>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/warp>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/thread>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/element>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/host_tensor>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/host>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/reduce>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/half>
)
#once new arches are enabled make this an option on the main cmake file
# and pass down here to be exported
target_compile_options(device_operations
PRIVATE --offload-arch=gfx908
)
# install(TARGETS device_operations LIBRARY DESTINATION lib)
install(TARGETS device_operations
EXPORT device_operationsTargets
LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
)
install(DIRECTORY ${DEV_OPS_INC_DIRS} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck)
install(EXPORT device_operationsTargets
FILE composable_kerneldevice_operationsTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
View file @ b134b7d6
......@@ -18,9 +18,9 @@ set(DEVICE_BATCHED_GEMM_INSTANCE_SOURCE
device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp;
)
add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
target_compile_features(device_batched_gemm_instance PUBLIC)
add_library(device_batched_gemm_instance OBJECT ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
# target_compile_features(device_batched_gemm_instance PUBLIC)
set_target_properties(device_batched_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_batched_gemm_instance LIBRARY DESTINATION lib)
# install(TARGETS device_batched_gemm_instance LIBRARY DESTINATION lib)
clang_tidy_check(device_batched_gemm_instance)
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
View file @ b134b7d6
......@@ -5,7 +5,8 @@ set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
)
add_instance_library(device_batched_gemm_reduce_instance ${DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE})
install(TARGETS device_batched_gemm_reduce_instance LIBRARY DESTINATION lib)
add_instance_library(device_batched_gemm_reduce_instance OBJECT ${DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE})
target_compile_features(device_batched_gemm_reduce_instance PUBLIC)
set_target_properties(device_batched_gemm_reduce_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_batched_gemm_reduce_instance)
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