ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix bug of layernorm ckProfiler and refine code (#448)

Browse Source 
* Fix bug of profiler for layernorm

* 1. Rename layernorm into normalization
2. Decouple softmax from normalization

* clang-format

[ROCm/composable_kernel commit: 1b62bfaa2a]
This commit is contained in:
rocking5566
2022-10-13 10:06:39 +08:00
committed by GitHub
parent 096571bea9
commit 1dcaa3991f
29 changed files with 423 additions and 461 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
profiler/src/profile_layernorm.cpp
View File

@@ -12,8 +12,7 @@ using ck::index_t;
struct LayernormArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"strideXY", {}}, {"strideGamma", {}}, {"strideBeta", {}}};
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
@@ -52,9 +51,6 @@ void print_help_layernorm()
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1024 1024) \n"
<< "--strideXY: tensor strides (e.g, --strideXY 1024 1)\n"
<< "--strideGamma: tensor strides (e.g, --strideGamma 1)\n"
<< "--strideBeta: tensor strides (e.g, --strideBeta 1)\n"
<< std::endl;
}
@@ -77,10 +73,7 @@ int profile_layernorm(int argc, char* argv[])
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
const std::vector<index_t> strideXY = arg_parser.long_opts["strideXY"];
const std::vector<index_t> strideGamma = arg_parser.long_opts["strideGamma"];
const std::vector<index_t> strideBeta = arg_parser.long_opts["strideBeta"];
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F16 = ck::half_t;
using F32 = float;
@@ -88,25 +81,13 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(do_verification,
init_method,
do_log,
time_kernel,
length,
strideXY,
strideGamma,
strideBeta);
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(do_verification,
init_method,
do_log,
time_kernel,
length,
strideXY,
strideGamma,
strideBeta);
ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{

84
profiler/src/profile_normalization.cpp → profiler/src/profile_softmax.cpp
View File

@@ -5,7 +5,7 @@
#include <vector>
#include <unordered_map>
#include "profiler/include/profile_normalization_impl.hpp"
#include "profiler/include/profile_softmax_impl.hpp"
using ck::index_t;
using ck::profiler::NormDataType;
@@ -95,30 +95,29 @@ int profile_normalization(int argc, char* argv[])
{
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t, 3>(
do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float, 3>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
ck::profiler::profile_softmax_impl<float, float, float, 3>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
@@ -129,30 +128,29 @@ int profile_normalization(int argc, char* argv[])
{
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t, 4>(
do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float, 4>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
ck::profiler::profile_softmax_impl<float, float, float, 4>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{