NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-05-25 15:24:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUTLASS 2.1 (#83)

Browse Source 
CUTLASS 2.1 contributes:
- BLAS-style host-side API added to CUTLASS Library
- Planar Complex GEMM kernels targeting Volta and Turing Tensor Cores
- Minor enhancements and bug fixes
This commit is contained in:
Andrew Kerr
2020-04-07 13:51:25 -07:00
committed by GitHub
parent 7c0cd26d13
commit 96dab34ad9
196 changed files with 20653 additions and 1995 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
tools/profiler/src/performance_report.cpp
View File

@@ -29,9 +29,15 @@
#include <iostream>
#include <stdexcept>
#include <iomanip>
#include <algorithm>
#include <cstring>
#include "cutlass/library/util.h"
#include "cutlass/library/util.h"
#include "performance_report.h"
#include "debug.h"
namespace cutlass {
namespace profiler {
@@ -57,12 +63,17 @@ namespace profiler {
PerformanceReport::PerformanceReport(
Options const &options,
std::vector<std::string> const &argument_names
std::vector<std::string> const &argument_names,
library::OperationKind const &op_kind
):
options_(options), argument_names_(argument_names), problem_index_(0), good_(true) {
options_(options), argument_names_(argument_names), problem_index_(0), good_(true), op_kind_(op_kind) {
std::string file_name = options_.report.output_path.substr(0, options_.report.output_path.rfind("."));
std::string file_extension = options_.report.output_path.substr(options_.report.output_path.rfind(".") + 1);
op_file_name_ = file_name + "." + to_string(op_kind_) + "." + file_extension;
//
// Open output file
// Open output file for operation of PerformanceReport::op_kind
//
if (!options_.report.output_path.empty()) {
@@ -70,17 +81,17 @@ PerformanceReport::PerformanceReport(
if (options_.report.append) {
std::ifstream test_output_file(options_.report.output_path.c_str());
std::ifstream test_output_file(op_file_name_);
if (test_output_file.is_open()) {
print_header = false;
test_output_file.close();
}
output_file_.open(options_.report.output_path.c_str(), std::ios::app);
output_file_.open(op_file_name_, std::ios::app);
}
else {
output_file_.open(options_.report.output_path.c_str());
output_file_.open(op_file_name_);
}
if (!output_file_.good()) {
@@ -148,7 +159,7 @@ void PerformanceReport::close() {
}
}
else if (output_file_.is_open() && options_.report.verbose) {
std::cout << "\n\nWrote results to '" << options_.report.output_path << "'" << std::endl;
std::cout << "\n\nWrote results to '" << op_file_name_ << "'" << std::endl;
}
}
@@ -184,19 +195,30 @@ std::ostream & PerformanceReport::print_result_pretty_(
out
<< "\n"
<< " Provider: " << SHELL_COLOR_BRIGHT() << to_string(result.provider, true) << SHELL_COLOR_END() << "\n"
<< " Operation: " << result.operation_name << "\n\n"
<< " Disposition: " << disposition_status_color(result.disposition) << to_string(result.disposition, true) << SHELL_COLOR_END() << "\n"
<< " Status: " << SHELL_COLOR_BRIGHT() << library::to_string(result.status, true) << SHELL_COLOR_END() << "\n";
<< " Provider: " << SHELL_COLOR_BRIGHT() << library::to_string(result.provider, true) << SHELL_COLOR_END() << "\n"
<< " Operation: " << result.operation_name << "\n\n"
<< " Status: " << SHELL_COLOR_BRIGHT() << library::to_string(result.status, true) << SHELL_COLOR_END() << "\n"
<< " Verification: " << SHELL_COLOR_BRIGHT() << (options_.verification.enabled ? "ON":"OFF") << SHELL_COLOR_END() << "\n"
<< " Disposition: " << disposition_status_color(result.disposition) << to_string(result.disposition, true) << SHELL_COLOR_END() << "\n\n";
// Display individual verification results for each verification-provider
if (options_.verification.enabled) {
static int const indent_spaces = 22;
for(auto & m : result.verification_map) {
out << std::right << std::setw(indent_spaces) << library::to_string(m.first, true) << ": " << to_string(m.second, true) << "\n";
}
}
out
<< "\n Arguments: ";
<< "\n Arguments: ";
int column_idx = 0;
for (auto const &arg : result.arguments) {
if (!arg.second.empty()) {
out << " --" << arg.first << "=" << arg.second;
column_idx += 4 + arg.first.size() + arg.second.size();
column_idx += int(4 + arg.first.size() + arg.second.size());
if (column_idx > 90) {
out << " \\\n ";
column_idx = 0;
@@ -206,15 +228,15 @@ std::ostream & PerformanceReport::print_result_pretty_(
out << "\n\n";
out
<< " Bytes: " << result.bytes << " bytes\n"
<< " FLOPs: " << result.flops << " flops\n\n";
<< " Bytes: " << result.bytes << " bytes\n"
<< " FLOPs: " << result.flops << " flops\n\n";
if (result.good()) {
out
<< " Runtime: " << result.runtime << " ms\n"
<< " Memory: " << result.gbytes_per_sec() << " GiB/s\n"
<< "\n Math: " << result.gflops_per_sec() << " GFLOP/s\n";
<< " Runtime: " << result.runtime << " ms\n"
<< " Memory: " << result.gbytes_per_sec() << " GiB/s\n"
<< "\n Math: " << result.gflops_per_sec() << " GFLOP/s\n";
}