AI/rocm-systems
2
0
Derivar 0
Código Questões Pedidos de integração Operações Pacotes Planeamentos Lançamentos Wiki Trabalho

SWDEV-543981 - Performance Test Improvement for Dispatch Speed and Kernel Latency (#527)

Ver a proveniência 
* SWDEV-543981 new kernel latency test with different timing modes and taking multiple iterations of same test

* SWDEV-543981 cleanup

* SWDEV-543981 removed outdated hit test

* SWDEV-543981 Updated timing kernel

[ROCm/hip-tests commit: d227a8110c]
Este cometimento está contido em:
Luo, Phoebe
2025-08-15 15:34:44 -04:00
cometido por GitHub
ascendente 5283a114b2
cometimento 12a1235939
2 ficheiros modificados com 157 adições e 330 eliminações
Descarregar ficheiro patch Descarregar ficheiro diff Expandir todos os ficheiros Colapsar todos os ficheiros
projects/hip-tests/catch/perftests/dispatch/hipPerfDispatchSpeed.cc
+157 -123
Ver ficheiro
@@ -18,10 +18,10 @@
*/
/**
* @addtogroup hipPerfDispatchSpeed hipPerfDispatchSpeed
* @{
* @ingroup perfDispatchTest
*/
* @addtogroup hipPerfDispatchSpeed hipPerfDispatchSpeed
* @{
* @ingroup perfDispatchTest
*/
// #define ENABLE_DEBUG 1
@@ -29,145 +29,179 @@
#include <string.h>
#include <complex>
typedef struct {
unsigned int iterations;
int flushEvery;
} testStruct;
testStruct testList[] = {
{ 1, -1},
{ 1, -1},
{ 10, 1},
{ 10, -1},
{ 100, 1},
{ 100, 10},
{ 100, -1},
{ 1000, 1},
{ 1000, 10},
{ 1000, 100},
{ 1000, -1},
{ 10000, 1},
{ 10000, 10},
{ 10000, 100},
{ 10000, 1000},
{ 10000, -1},
{ 100000, 1},
{ 100000, 10},
{ 100000, 100},
{ 100000, 1000},
{ 100000, 10000},
{ 100000, -1},
};
unsigned int mapTestList[] = {1, 1, 10, 100, 1000, 10000, 100000};
__global__ void _dispatchSpeed(float *outBuf) {
int i = (blockIdx.x * blockDim.x + threadIdx.x);
if (i < 0)
outBuf[i] = 0.0f;
};
/**
* Test Description
* ------------------------
* - Verify the hipPerf Dispatch speed.
* Test source
* ------------------------
* - perftests/compute/hipPerfMandelbrot.cc
* Test requirements
* ------------------------
* - HIP_VERSION >= 5.6
*/
* Test Description
* ------------------------
* - Verify the hipPerf Dispatch and Execution speed, AKA total kernel latency
* Test source
* ------------------------
* - perftests/dispatch/hipPerfDispatchSpeed.cc
* Test requirements
* ------------------------
* - HIP_VERSION >= 5.6
*/
TEST_CASE("Perf_hipPerfDispatchSpeed") {
int p_gpuDevice = 0;
int p_tests = -1;
unsigned int testList[] = {1, 10, 100, 1000, 10000};
// dummy kernel that just dispatches and does nothing
__global__ void _dispatchSpeed(float* outBuf) {
int i = (blockIdx.x * blockDim.x + threadIdx.x);
if (i < 0) outBuf[i] = 0.0f;
};
// kernel that has an execution of count, in GPU clock ticks
__global__ void _TimingKernel(uint64_t count) {
uint64_t begin_time = __builtin_amdgcn_s_memrealtime();
uint64_t curr_time = begin_time;
do {
curr_time = __builtin_amdgcn_s_memrealtime();
} while (begin_time + count > curr_time);
}
enum TimingMode { TimingMode_WallTime, TimingMode_HIPEvent, TimingMode_NumModes };
TEST_CASE("Perf_hipPerfDispatchAndExecutionSpeed") {
hipError_t err = hipSuccess;
hipDeviceProp_t props;
HIP_CHECK(hipGetDeviceProperties(&props, p_gpuDevice));
unsigned int testListSize = sizeof(testList) / sizeof(testStruct);
int numTests = (p_tests == -1) ? (2*2*testListSize - 1) : p_tests;
int test = (p_tests == -1) ? 0 : p_tests;
unsigned int testListSize = sizeof(testList) / sizeof(testList[0]);
int numTests = testListSize;
int warmup = 10; // number of warmup iterations
DEBUG_PRINT("numTests %d", numTests);
// set up timing kernel
uint64_t timer_freq_in_hz;
int clock_rate = 0; // in kHz
HIP_CHECK(hipDeviceGetAttribute(&clock_rate, hipDeviceAttributeWallClockRate, 0));
timer_freq_in_hz = clock_rate * 1000;
uint64_t timing_in_us = 4; // CHANGE THIS TO CHANGE EXECUTION TIME
const uint64_t timing_count = timer_freq_in_hz * timing_in_us / 1000000;
int iterations = 100; // number of times to run the test to get an average time
float* srcBuffer = NULL;
unsigned int bufSize_ = 64*sizeof(float);
unsigned int bufSize_ = 64 * sizeof(float);
err = hipMalloc(&srcBuffer, bufSize_);
REQUIRE(err == hipSuccess);
for (; test <= numTests; test++) {
int openTest = test % testListSize;
bool sleep = false;
hipEvent_t startEvent, stopEvent;
if (test >= (testListSize * 2)) {
sleep = true;
HIP_CHECK(hipEventCreate(&startEvent));
HIP_CHECK(hipEventCreate(&stopEvent));
// run twice for both dispatch speed and full kernel latency
for (int j = 0; j < 2; j++) {
bool useTimingKernel = (j == 1);
if (useTimingKernel) {
CONSOLE_PRINT("\nTIMING KERNEL TEST ()");
CONSOLE_PRINT("--------------------------------------------------------------");
} else {
CONSOLE_PRINT("EMPTY KERNEL TEST");
CONSOLE_PRINT("--------------------------------------------------------------");
}
int threads = (bufSize_ / sizeof(float));
int threads_per_block = 64;
int blocks = (threads/threads_per_block) + (threads % threads_per_block);
// warmup
hipLaunchKernelGGL(_dispatchSpeed, dim3(blocks), dim3(threads_per_block),
0, hipStream_t(0), srcBuffer);
err = hipDeviceSynchronize();
REQUIRE(err == hipSuccess);
auto start = std::chrono::high_resolution_clock::now();
for (unsigned int i = 0; i < testList[openTest].iterations; i++) {
hipLaunchKernelGGL(_dispatchSpeed, dim3(blocks),
dim3(threads_per_block), 0, hipStream_t(0), srcBuffer);
if ((testList[openTest].flushEvery > 0) &&
(((i + 1) % testList[openTest].flushEvery) == 0)) {
if (sleep) {
err = hipDeviceSynchronize();
REQUIRE(err == hipSuccess);
} else {
do {
err = hipStreamQuery(NULL);
} while (err == hipErrorNotReady);
// loop through all possible timing methods
for (unsigned int i = 0; i < TimingMode_NumModes; i++) {
TimingMode mode = static_cast<TimingMode>(i);
CONSOLE_PRINT("\nTIMING METHOD:");
switch (mode) {
case TimingMode_WallTime:
CONSOLE_PRINT("Wall Time");
break;
case TimingMode_HIPEvent:
CONSOLE_PRINT("HIP Events");
break;
default:
CONSOLE_PRINT("Unknown Mode");
}
// go through test iterations
for (int test = 0; test < numTests; test++) {
int openTest = test % testListSize;
int threads = (bufSize_ / sizeof(float));
int threads_per_block = 64;
int blocks = (threads / threads_per_block) + (threads % threads_per_block);
double finalPerf = 0.0;
double wallMicroSec = 0.0;
std::chrono::high_resolution_clock::time_point startWall, stopWall;
// warmup
for (int i = 0; i < warmup; i++) {
hipLaunchKernelGGL(_TimingKernel, dim3(blocks), dim3(threads_per_block), 0,
hipStream_t(0), timing_count);
}
HIP_CHECK(hipStreamSynchronize(0));
for (int it = 0; it < iterations; it++) {
switch (mode) {
case TimingMode_WallTime:
startWall = std::chrono::high_resolution_clock::now();
break;
case TimingMode_HIPEvent:
HIP_CHECK(hipEventRecord(startEvent, 0));
break;
default:
CONSOLE_PRINT("Unknown Mode");
}
for (unsigned int i = 0; i < testList[openTest]; i++) {
if (useTimingKernel) {
// use the timing kernel to measure dispatch and execution speed
hipLaunchKernelGGL(_TimingKernel, dim3(blocks), dim3(threads_per_block), 0,
hipStream_t(0), timing_count);
} else {
// use the dispatch speed kernel
hipLaunchKernelGGL(_dispatchSpeed, dim3(blocks), dim3(threads_per_block), 0,
hipStream_t(0), srcBuffer);
}
}
switch (mode) {
case TimingMode_WallTime: {
err = hipStreamSynchronize(0);
REQUIRE(err == hipSuccess);
stopWall = std::chrono::high_resolution_clock::now();
wallMicroSec =
std::chrono::duration<double, std::micro>(stopWall - startWall).count();
finalPerf += wallMicroSec / testList[openTest];
break;
}
case TimingMode_HIPEvent: {
HIP_CHECK(hipEventRecord(stopEvent, 0));
HIP_CHECK(hipEventSynchronize(stopEvent));
float elapsed;
HIP_CHECK(hipEventElapsedTime(&elapsed, startEvent, stopEvent)); // in milliseconds
finalPerf += (elapsed * 1000.0f) / testList[openTest]; // convert ms to µs
break;
}
default:
CONSOLE_PRINT("Unknown Mode");
}
}
finalPerf /= iterations; // average the performance over all iterations
CONSOLE_PRINT("HIPPerfDispatchSpeed[%3d] %7d dispatches (us/disp) %3f", test,
testList[openTest], (float)finalPerf);
}
}
if (sleep) {
err = hipDeviceSynchronize();
REQUIRE(err == hipSuccess);
} else {
do {
err = hipStreamQuery(NULL);
} while (err == hipErrorNotReady);
}
auto stop = std::chrono::high_resolution_clock::now();
double microSec = std::chrono::duration<double, std::micro>(stop - start).count();
// microseconds per launch
double perf = (microSec/testList[openTest].iterations);
const char *waitType;
const char *extraChar;
const char *n;
if (sleep) {
waitType = "sleep";
extraChar = "";
n = "";
} else {
waitType = "spin";
n = "n";
extraChar = " ";
}
if (testList[openTest].flushEvery > 0) {
CONSOLE_PRINT("HIPPerfDispatchSpeed[%3d] %7d dispatches %s%sing every %5d (us/disp) %3f",
test, testList[openTest].iterations, waitType, n, testList[openTest].flushEvery,
(float)perf);
} else {
CONSOLE_PRINT("HIPPerfDispatchSpeed[%3d] %7d dispatches (%s%s) (us/disp) %3f",
test, testList[openTest].iterations, waitType, extraChar, (float)perf);
}
}
HIP_CHECK(hipEventDestroy(startEvent));
HIP_CHECK(hipEventDestroy(stopEvent));
HIP_CHECK(hipFree(srcBuffer));
}
/**
* End doxygen group perfDispatchTest.
* @}
*/
* End doxygen group perfDispatchTest.
* @}
*/
projects/hip-tests/perftests/dispatch/hipPerfDispatchSpeed.cpp
-207
Ver ficheiro
@@ -1,207 +0,0 @@
/*
Copyright (c) 2015 - 2021 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
/* HIT_START
* BUILD: %t %s ../../src/test_common.cpp ../../src/timer.cpp
* TEST: %t
* HIT_END
*/
#include <stdio.h>
#include <assert.h>
#include <string.h>
#include <complex>
#include "timer.h"
#include "test_common.h"
// Quiet pesky warnings
#ifdef WIN_OS
#define SNPRINTF sprintf_s
#else
#define SNPRINTF snprintf
#endif
#define CHAR_BUF_SIZE 512
#define CHECK_RESULT(test, msg) \
if ((test)) \
{ \
printf("\n%s\n", msg); \
abort(); \
}
typedef struct {
unsigned int iterations;
int flushEvery;
} testStruct;
testStruct testList[] =
{
{ 1, -1},
{ 1, -1},
{ 10, 1},
{ 10, -1},
{ 100, 1},
{ 100, 10},
{ 100, -1},
{ 1000, 1},
{ 1000, 10},
{ 1000, 100},
{ 1000, -1},
{ 10000, 1},
{ 10000, 10},
{ 10000, 100},
{ 10000, 1000},
{ 10000, -1},
{ 100000, 1},
{ 100000, 10},
{ 100000, 100},
{ 100000, 1000},
{ 100000, 10000},
{ 100000, -1},
};
unsigned int mapTestList[] = {1, 1, 10, 100, 1000, 10000, 100000};
__global__ void _dispatchSpeed(float *outBuf)
{
int i = (blockIdx.x * blockDim.x + threadIdx.x);
if (i < 0)
outBuf[i] = 0.0f;
};
int main(int argc, char* argv[]) {
HipTest::parseStandardArguments(argc, argv, true);
hipError_t err = hipSuccess;
hipDeviceProp_t props = {0};
hipGetDeviceProperties(&props, p_gpuDevice);
CHECK_RESULT(err != hipSuccess, "hipGetDeviceProperties failed" );
printf("Set device to %d : %s\n", p_gpuDevice, props.name);
unsigned int testListSize = sizeof(testList) / sizeof(testStruct);
int numTests = (p_tests == -1) ? (2*2*testListSize - 1) : p_tests;
int test = (p_tests == -1) ? 0 : p_tests;
float* srcBuffer = NULL;
unsigned int bufSize_ = 64*sizeof(float);
err = hipMalloc(&srcBuffer, bufSize_);
CHECK_RESULT(err != hipSuccess, "hipMalloc failed");
for(;test <= numTests; test++)
{
int openTest = test % testListSize;
bool sleep = false;
if (test >= (testListSize * 2))
{
sleep = true;
}
int threads = (bufSize_ / sizeof(float));
int threads_per_block = 64;
int blocks = (threads/threads_per_block) + (threads % threads_per_block);
// warmup
hipLaunchKernelGGL(_dispatchSpeed, dim3(blocks), dim3(threads_per_block),
0, hipStream_t(0), srcBuffer);
err = hipDeviceSynchronize();
CHECK_RESULT(err != hipSuccess, "hipDeviceSynchronize failed");
CPerfCounter timer;
timer.Reset();
timer.Start();
for (unsigned int i = 0; i < testList[openTest].iterations; i++)
{
hipLaunchKernelGGL(_dispatchSpeed, dim3(blocks), dim3(threads_per_block),
0, hipStream_t(0), srcBuffer);
if ((testList[openTest].flushEvery > 0) &&
(((i + 1) % testList[openTest].flushEvery) == 0))
{
if (sleep)
{
err = hipDeviceSynchronize();
CHECK_RESULT(err != hipSuccess, "hipDeviceSynchronize failed");
}
else
{
do {
err = hipStreamQuery(NULL);
} while (err == hipErrorNotReady);
}
}
}
if (sleep)
{
err = hipDeviceSynchronize();
CHECK_RESULT(err != hipSuccess, "hipDeviceSynchronize failed");
}
else
{
do {
err = hipStreamQuery(NULL);
} while (err == hipErrorNotReady);
}
timer.Stop();
double sec = timer.GetElapsedTime();
// microseconds per launch
double perf = (1000000.f*sec/testList[openTest].iterations);
const char *waitType;
const char *extraChar;
const char *n;
if (sleep)
{
waitType = "sleep";
extraChar = "";
n = "";
}
else
{
waitType = "spin";
n = "n";
extraChar = " ";
}
char buf[256];
if (testList[openTest].flushEvery > 0)
{
SNPRINTF(buf, sizeof(buf),
"HIPPerfDispatchSpeed[%3d] %7d dispatches %s%sing every %5d (us/disp) %3f",
test, testList[openTest].iterations,
waitType, n, testList[openTest].flushEvery, (float)perf);
}
else
{
SNPRINTF(buf, sizeof(buf),
"HIPPerfDispatchSpeed[%3d] %7d dispatches (%s%s) (us/disp) %3f",
test, testList[openTest].iterations, waitType, extraChar, (float)perf);
}
printf("%s\n", buf);
}
hipFree(srcBuffer);
passed();
}