Merge branch 'amd-develop' into amd-master
Change-Id: I32d41081ac065f2c50531dc2e420802d765665e2
This commit is contained in:
Maneesh Gupta
@@ -0,0 +1,5 @@
|
||||
vadd_amp_arrayview
|
||||
vadd_hc_am
|
||||
vadd_hc_array
|
||||
vadd_hc_arrayview
|
||||
vadd_hip
|
||||
@@ -5,8 +5,8 @@ OPT=-O2
|
||||
HCC_CFLAGS= `$(HCC_HOME)/bin/hcc-config --cxxflags` ${OPT}
|
||||
HCC_LDFLAGS= `$(HCC_HOME)/bin/hcc-config --ldflags` ${OPT}
|
||||
|
||||
CPPAMP_CFLAGS= -std=c++amp -stdlib=libc++ -I$(HCC_HOME)/include
|
||||
CPPAMP_LDFLAGS= -std=c++amp -L$(HCC_HOME)/lib -Wl,--rpath=$(HCC_HOME)/lib -lc++ -lc++abi -ldl -lpthread -Wl,--whole-archive -lmcwamp -Wl,--no-whole-archive
|
||||
CPPAMP_CFLAGS= `$(HCC_HOME)/bin/clamp-config --cxxflags`
|
||||
CPPAMP_LDFLAGS= `$(HCC_HOME)/bin/clamp-config --ldflags`
|
||||
|
||||
HIP_PATH?= $(wildcard /opt/rocm/hip)
|
||||
ifeq (,$(HIP_PATH))
|
||||
|
||||
@@ -0,0 +1,5 @@
|
||||
runKernel.hip.out
|
||||
vcpy_isa.code
|
||||
vcpy_isa.hsaco
|
||||
vcpy_kernel.co
|
||||
vcpy_kernel.code
|
||||
@@ -1,7 +1,4 @@
|
||||
HIP_PATH?= $(wildcard /opt/rocm/hip)
|
||||
ifeq (,$(HIP_PATH))
|
||||
HIP_PATH=../../..
|
||||
endif
|
||||
HIPCC=$(HIP_PATH)/bin/hipcc
|
||||
|
||||
all: square.hip.out
|
||||
@@ -11,9 +8,10 @@ square.cuda.out : square.cu
|
||||
|
||||
#hipify square.cu > square.cpp
|
||||
# Then review & finish port in square.cpp
|
||||
#
|
||||
|
||||
square.hip.out: square.hipref.cpp
|
||||
$(HIPCC) square.hipref.cpp -o $@
|
||||
$(HIPCC) $(CXXFLAGS) square.hipref.cpp -o $@
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -133,6 +133,15 @@ void printDeviceProp (int deviceId)
|
||||
}
|
||||
}
|
||||
cout << endl;
|
||||
cout << setw(w1) << "non-peers: ";
|
||||
for (int i=0; i<deviceCnt; i++) {
|
||||
int isPeer;
|
||||
hipDeviceCanAccessPeer(&isPeer, i, deviceId);
|
||||
if (!isPeer) {
|
||||
cout << "device#" << i << " ";
|
||||
}
|
||||
}
|
||||
cout << endl;
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -1,172 +0,0 @@
|
||||
/*
|
||||
Copyright (c) 2015-2016 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.
|
||||
*/
|
||||
|
||||
#include<iostream>
|
||||
|
||||
// hip header file
|
||||
#include "hip/hip_runtime.h"
|
||||
|
||||
#define WIDTH 1024
|
||||
|
||||
#define NUM (WIDTH*WIDTH)
|
||||
|
||||
#define THREADS_PER_BLOCK_X 4
|
||||
#define THREADS_PER_BLOCK_Y 4
|
||||
#define THREADS_PER_BLOCK_Z 1
|
||||
|
||||
// Device (Kernel) function, it must be void
|
||||
// hipLaunchParm provides the execution configuration
|
||||
__global__ void matrixTranspose(hipLaunchParm lp,
|
||||
float *out,
|
||||
float *in,
|
||||
const int width)
|
||||
{
|
||||
int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
|
||||
int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
|
||||
|
||||
out[y * width + x] = in[x * width + y];
|
||||
}
|
||||
|
||||
// CPU implementation of matrix transpose
|
||||
void matrixTransposeCPUReference(
|
||||
float * output,
|
||||
float * input,
|
||||
const unsigned int width)
|
||||
{
|
||||
for(unsigned int j=0; j < width; j++)
|
||||
{
|
||||
for(unsigned int i=0; i < width; i++)
|
||||
{
|
||||
output[i*width + j] = input[j*width + i];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
int main() {
|
||||
|
||||
float* Matrix;
|
||||
float* TransposeMatrix;
|
||||
float* cpuTransposeMatrix;
|
||||
|
||||
float* gpuMatrix;
|
||||
float* gpuTransposeMatrix;
|
||||
|
||||
hipDeviceProp_t devProp;
|
||||
hipGetDeviceProperties(&devProp, 0);
|
||||
|
||||
std::cout << "Device name " << devProp.name << std::endl;
|
||||
|
||||
hipEvent_t start, stop;
|
||||
hipEventCreate(&start);
|
||||
hipEventCreate(&stop);
|
||||
float eventMs = 1.0f;
|
||||
|
||||
int i;
|
||||
int errors;
|
||||
|
||||
Matrix = (float*)malloc(NUM * sizeof(float));
|
||||
TransposeMatrix = (float*)malloc(NUM * sizeof(float));
|
||||
cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));
|
||||
|
||||
// initialize the input data
|
||||
for (i = 0; i < NUM; i++) {
|
||||
Matrix[i] = (float)i*10.0f;
|
||||
}
|
||||
|
||||
// allocate the memory on the device side
|
||||
hipMalloc((void**)&gpuMatrix, NUM * sizeof(float));
|
||||
hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float));
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Memory transfer from host to device
|
||||
hipMemcpy(gpuMatrix, Matrix, NUM*sizeof(float), hipMemcpyHostToDevice);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("hipMemcpyHostToDevice time taken = %6.3fms\n", eventMs);
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Lauching kernel from host
|
||||
hipLaunchKernel(matrixTranspose,
|
||||
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
|
||||
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
|
||||
0, 0,
|
||||
gpuTransposeMatrix , gpuMatrix, WIDTH);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("kernel Execution time = %6.3fms\n", eventMs);
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Memory transfer from device to host
|
||||
hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM*sizeof(float), hipMemcpyDeviceToHost);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("hipMemcpyDeviceToHost time taken = %6.3fms\n", eventMs);
|
||||
|
||||
// CPU MatrixTranspose computation
|
||||
matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);
|
||||
|
||||
// verify the results
|
||||
errors = 0;
|
||||
double eps = 1.0E-6;
|
||||
for (i = 0; i < NUM; i++) {
|
||||
if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps ) {
|
||||
errors++;
|
||||
}
|
||||
}
|
||||
if (errors!=0) {
|
||||
printf("FAILED: %d errors\n",errors);
|
||||
} else {
|
||||
printf ("PASSED!\n");
|
||||
}
|
||||
|
||||
//free the resources on device side
|
||||
hipFree(gpuMatrix);
|
||||
hipFree(gpuTransposeMatrix);
|
||||
|
||||
//free the resources on host side
|
||||
free(Matrix);
|
||||
free(TransposeMatrix);
|
||||
free(cpuTransposeMatrix);
|
||||
|
||||
return errors;
|
||||
}
|
||||
@@ -0,0 +1,53 @@
|
||||
HIP_PATH?= $(wildcard /opt/rocm/hip)
|
||||
|
||||
HIPCC=$(HIP_PATH)/bin/hipcc
|
||||
|
||||
|
||||
HIPPROFILER=/opt/rocm/bin/rocm-profiler
|
||||
PROFILER_OPT=-A -o MT.atp -e HIP_PROFILE_API=1
|
||||
HIPPROFILER_POST_CMD=$(HIP_PATH)/bin/hipdemangleatp MT.atp
|
||||
|
||||
TARGET=hcc
|
||||
|
||||
SOURCES = MatrixTranspose.cpp
|
||||
OBJECTS = $(SOURCES:.cpp=.o)
|
||||
|
||||
EXECUTABLE=./MatrixTranspose
|
||||
|
||||
.PHONY: test
|
||||
|
||||
|
||||
all: $(EXECUTABLE) profile
|
||||
|
||||
|
||||
|
||||
OPT =-g
|
||||
CXXFLAGS =$(OPT)
|
||||
CXX=$(HIPCC)
|
||||
|
||||
|
||||
$(EXECUTABLE): $(OBJECTS)
|
||||
$(HIPCC) $(OBJECTS) -o $@
|
||||
|
||||
|
||||
profile: $(EXECUTABLE)
|
||||
$(HIPPROFILER) $(PROFILER_OPT) $(EXECUTABLE)
|
||||
$(HIPPROFILER_POST_CMD)
|
||||
|
||||
|
||||
# Pass option to control start and stop iterations for profiling - see MatrixTranspose.cpp for implementation:
|
||||
# Note we start profiler in --startdisabled mode - no timing collected until app enabled it via hipProfilerStart()
|
||||
profile_trigger: $(EXECUTABLE)
|
||||
$(HIPPROFILER) $(PROFILER_OPT) --startdisabled $(EXECUTABLE) 3 6
|
||||
$(HIPPROFILER_POST_CMD)
|
||||
|
||||
|
||||
run: $(EXECUTABLE)
|
||||
$(EXECUTABLE)
|
||||
|
||||
|
||||
clean:
|
||||
rm -f $(EXECUTABLE)
|
||||
rm -f $(OBJECTS)
|
||||
rm -f $(HIP_PATH)/src/*.o
|
||||
|
||||
@@ -0,0 +1,233 @@
|
||||
/*
|
||||
Copyright (c) 2015-2016 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.
|
||||
*/
|
||||
|
||||
#include<iostream>
|
||||
|
||||
// hip header file
|
||||
#include "hip/hip_runtime.h"
|
||||
#include "hip/hip_profile.h"
|
||||
|
||||
#define WIDTH 1024
|
||||
|
||||
#define NUM (WIDTH*WIDTH)
|
||||
|
||||
#define THREADS_PER_BLOCK_X 4
|
||||
#define THREADS_PER_BLOCK_Y 4
|
||||
#define THREADS_PER_BLOCK_Z 1
|
||||
|
||||
#define ITERATIONS 10
|
||||
|
||||
// Cmdline parms to control start and stop triggers
|
||||
int startTriggerIteration=-1;
|
||||
int stopTriggerIteration=-1;
|
||||
|
||||
// Device (Kernel) function, it must be void
|
||||
// hipLaunchParm provides the execution configuration
|
||||
__global__ void matrixTranspose(hipLaunchParm lp,
|
||||
float *out,
|
||||
float *in,
|
||||
const int width)
|
||||
{
|
||||
int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
|
||||
int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
|
||||
|
||||
out[y * width + x] = in[x * width + y];
|
||||
}
|
||||
|
||||
// CPU implementation of matrix transpose
|
||||
void matrixTransposeCPUReference(
|
||||
float * output,
|
||||
float * input,
|
||||
const unsigned int width)
|
||||
{
|
||||
for(unsigned int j=0; j < width; j++)
|
||||
{
|
||||
for(unsigned int i=0; i < width; i++)
|
||||
{
|
||||
output[i*width + j] = input[j*width + i];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
// Use a separate function to demonstrate how to use function name as part of scoped marker:
|
||||
void runGPU(float *Matrix, float *TransposeMatrix,
|
||||
float* gpuMatrix, float* gpuTransposeMatrix) {
|
||||
|
||||
// __func__ is a standard C++ macro which expands to the name of the function, in this case "runGPU"
|
||||
HIP_SCOPED_MARKER(__func__, "MyGroup");
|
||||
|
||||
for (int i=0; i<ITERATIONS; i++) {
|
||||
|
||||
if (i==startTriggerIteration) {
|
||||
hipProfilerStart();
|
||||
}
|
||||
if (i==stopTriggerIteration) {
|
||||
hipProfilerStop();
|
||||
}
|
||||
|
||||
float eventMs = 0.0f;
|
||||
|
||||
hipEvent_t start, stop;
|
||||
hipEventCreate(&start);
|
||||
hipEventCreate(&stop);
|
||||
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Memory transfer from host to device
|
||||
hipMemcpy(gpuMatrix, Matrix, NUM*sizeof(float), hipMemcpyHostToDevice);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("hipMemcpyHostToDevice time taken = %6.3fms\n", eventMs);
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Lauching kernel from host
|
||||
hipLaunchKernel(matrixTranspose,
|
||||
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
|
||||
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
|
||||
0, 0,
|
||||
gpuTransposeMatrix , gpuMatrix, WIDTH);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("kernel Execution time = %6.3fms\n", eventMs);
|
||||
|
||||
// Record the start event
|
||||
hipEventRecord(start, NULL);
|
||||
|
||||
// Memory transfer from device to host
|
||||
hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM*sizeof(float), hipMemcpyDeviceToHost);
|
||||
|
||||
// Record the stop event
|
||||
hipEventRecord(stop, NULL);
|
||||
hipEventSynchronize(stop);
|
||||
|
||||
hipEventElapsedTime(&eventMs, start, stop);
|
||||
|
||||
printf ("hipMemcpyDeviceToHost time taken = %6.3fms\n", eventMs);
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
|
||||
if (argc >= 2) {
|
||||
startTriggerIteration = atoi(argv[1]);
|
||||
printf ("info : will start tracing at iteration:%d\n", startTriggerIteration);
|
||||
}
|
||||
if (argc >= 3) {
|
||||
stopTriggerIteration = atoi(argv[2]);
|
||||
printf ("info : will stop tracing at iteration:%d\n", stopTriggerIteration);
|
||||
}
|
||||
|
||||
float* Matrix;
|
||||
float* TransposeMatrix;
|
||||
float* cpuTransposeMatrix;
|
||||
|
||||
float* gpuMatrix;
|
||||
float* gpuTransposeMatrix;
|
||||
|
||||
hipDeviceProp_t devProp;
|
||||
hipGetDeviceProperties(&devProp, 0);
|
||||
|
||||
std::cout << "Device name " << devProp.name << std::endl;
|
||||
|
||||
{
|
||||
// Show example of how to create a "scoped marker".
|
||||
// The scoped marker records the time spent inside the { scope } of the marker - the begin timestamp is at the
|
||||
// beginning of the code scope, and the end is recorded when the SCOPE exits. This can be viewed in CodeXL
|
||||
// timeline relative to other GPU and CPU events.
|
||||
// This marker captures the time spent in setup including host allocation, initialization, and device memory allocation.
|
||||
HIP_SCOPED_MARKER("Setup", "MyGroup");
|
||||
|
||||
|
||||
|
||||
Matrix = (float*)malloc(NUM * sizeof(float));
|
||||
TransposeMatrix = (float*)malloc(NUM * sizeof(float));
|
||||
cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));
|
||||
|
||||
// initialize the input data
|
||||
for (int i = 0; i < NUM; i++) {
|
||||
Matrix[i] = (float)i*10.0f;
|
||||
}
|
||||
|
||||
|
||||
// allocate the memory on the device side
|
||||
hipMalloc((void**)&gpuMatrix, NUM * sizeof(float));
|
||||
hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float));
|
||||
|
||||
// FYI, the scoped-marker will be destroyed here when the scope exits, and will record its "end" timestamp.
|
||||
}
|
||||
|
||||
runGPU(Matrix, TransposeMatrix, gpuMatrix, gpuTransposeMatrix);
|
||||
|
||||
|
||||
// show how to use explicit begin/end markers:
|
||||
// We begin the timed region with HIP_BEGIN_MARKER, passing in the markerName and group:
|
||||
// The region will stop when HIP_END_MARKER is called
|
||||
// This is another way to mark begin/end - as an alternative to scoped markers.
|
||||
HIP_BEGIN_MARKER("Check&TearDown", "MyGroup");
|
||||
|
||||
int errors = 0;
|
||||
|
||||
// CPU MatrixTranspose computation
|
||||
matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);
|
||||
|
||||
// verify the results
|
||||
double eps = 1.0E-6;
|
||||
for (int i = 0; i < NUM; i++) {
|
||||
if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps ) {
|
||||
errors++;
|
||||
}
|
||||
}
|
||||
if (errors!=0) {
|
||||
printf("FAILED: %d errors\n",errors);
|
||||
} else {
|
||||
printf ("PASSED!\n");
|
||||
}
|
||||
|
||||
//free the resources on device side
|
||||
hipFree(gpuMatrix);
|
||||
hipFree(gpuTransposeMatrix);
|
||||
|
||||
//free the resources on host side
|
||||
free(Matrix);
|
||||
free(TransposeMatrix);
|
||||
free(cpuTransposeMatrix);
|
||||
|
||||
// This ends the last marker started in this thread, in this case "Check&TearDown"
|
||||
HIP_END_MARKER();
|
||||
|
||||
return errors;
|
||||
}
|
||||
+5
-9
@@ -1,6 +1,6 @@
|
||||
## Using hipEvents to measure performance ###
|
||||
|
||||
This tutorial is follow-up of the previous two tutorial where we learn how to write our first hip program, in which we compute Matrix Transpose and in second one, we added feature to measure time taken for memory transfer and kernel execution. In this tutorial, we won't make amy changes to the source code. We'll explain how to use the codexl/rocm-profiler for hip timeline tracing.
|
||||
This tutorial is follow-up of the previous two tutorial where we learn how to write our first hip program, in which we compute Matrix Transpose and in second one, we added feature to measure time taken for memory transfer and kernel execution. In this tutorial, we'll explain how to use the codexl/rocm-profiler for hip timeline tracing. Also, we will augment the source code with additional markers so we can see the high-level application flow alongside the information that CodeXL automatically collects.
|
||||
|
||||
|
||||
## Introduction:
|
||||
@@ -24,15 +24,11 @@ HIP can generate markers at function being/end which are displayed on the CodeXL
|
||||
|
||||
1. Install ROCm-Profiler Installing HIP from the rocm pre-built packages, installs the ROCm-Profiler as well. Alternatively, you can build ROCm-Profiler using the instructions given below.
|
||||
|
||||
2. Build HIP with ATP markers enabled HIP pre-built packages are enabled with ATP marker support by default. To enable ATP marker support when building HIP from source, use the option -DCOMPILE_HIP_ATP_MARKER=1 during the cmake configure step.
|
||||
|
||||
3. Set HIP_ATP_MARKER
|
||||
`export HIP_ATP_MARKER=1`
|
||||
|
||||
4. Recompile the target application
|
||||
|
||||
5. Run with profiler enabled to generate ATP file.
|
||||
`/opt/rocm/bin/rocm-profiler -o <outputATPFileName> -A <applicationName> <applicationArguments>`
|
||||
2. Run with profiler enabled to generate ATP file.
|
||||
(These steps are also captured in the Makefile)
|
||||
The HIP_PROFILE_API enables display of the HIP APIs on the CodeXL trimeline view.
|
||||
`/opt/rocm/bin/rocm-profiler -o <outputATPFileName> -A <applicationName> -e HIP_PROFILE_API=1 <applicationArguments>`
|
||||
|
||||
##Using HIP_TRACE_API
|
||||
|
||||
+36
-36
@@ -1,36 +1,36 @@
|
||||
HIP_PATH?= $(wildcard /opt/rocm/hip)
|
||||
ifeq (,$(HIP_PATH))
|
||||
HIP_PATH=../../..
|
||||
endif
|
||||
|
||||
HIPCC=$(HIP_PATH)/bin/hipcc
|
||||
|
||||
TARGET=hcc
|
||||
|
||||
SOURCES = MatrixTranspose.cpp
|
||||
OBJECTS = $(SOURCES:.cpp=.o)
|
||||
|
||||
EXECUTABLE=./MatrixTranspose
|
||||
|
||||
.PHONY: test
|
||||
|
||||
|
||||
all: $(EXECUTABLE) test
|
||||
|
||||
CXXFLAGS =-g
|
||||
CXX=$(HIPCC)
|
||||
|
||||
|
||||
$(EXECUTABLE): $(OBJECTS)
|
||||
$(HIPCC) $(OBJECTS) -o $@
|
||||
|
||||
|
||||
test: $(EXECUTABLE)
|
||||
$(EXECUTABLE)
|
||||
|
||||
|
||||
clean:
|
||||
rm -f $(EXECUTABLE)
|
||||
rm -f $(OBJECTS)
|
||||
rm -f $(HIP_PATH)/src/*.o
|
||||
|
||||
HIP_PATH?= $(wildcard /opt/rocm/hip)
|
||||
ifeq (,$(HIP_PATH))
|
||||
HIP_PATH=../../..
|
||||
endif
|
||||
|
||||
HIPCC=$(HIP_PATH)/bin/hipcc
|
||||
|
||||
TARGET=hcc
|
||||
|
||||
SOURCES = peer2peer.cpp
|
||||
OBJECTS = $(SOURCES:.cpp=.o)
|
||||
|
||||
EXECUTABLE=./peer2peer
|
||||
|
||||
.PHONY: test
|
||||
|
||||
|
||||
all: $(EXECUTABLE) test
|
||||
|
||||
CXXFLAGS =-g
|
||||
CXX=$(HIPCC)
|
||||
|
||||
|
||||
$(EXECUTABLE): $(OBJECTS)
|
||||
$(HIPCC) $(OBJECTS) -o $@
|
||||
|
||||
|
||||
test: $(EXECUTABLE)
|
||||
$(EXECUTABLE)
|
||||
|
||||
|
||||
clean:
|
||||
rm -f $(EXECUTABLE)
|
||||
rm -f $(OBJECTS)
|
||||
rm -f $(HIP_PATH)/src/*.o
|
||||
|
||||
@@ -0,0 +1,241 @@
|
||||
/*
|
||||
Copyright (c) 2015-2016 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 WARRANUMTY OF ANY KIND, EXPRESS OR
|
||||
IMPLIED, INUMCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
FITNUMESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANUMY CLAIM, DAMAGES OR OTHER
|
||||
LIABILITY, WHETHER INUM AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR INUM CONUMECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
||||
THE SOFTWARE.
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
#include <hip/hip_runtime.h>
|
||||
#include <assert.h>
|
||||
#define WIDTH 32
|
||||
|
||||
#define NUM (WIDTH*WIDTH)
|
||||
|
||||
#define THREADS_PER_BLOCK_X 4
|
||||
#define THREADS_PER_BLOCK_Y 4
|
||||
#define THREADS_PER_BLOCK_Z 1
|
||||
|
||||
using namespace std;
|
||||
|
||||
#define KNRM "\x1B[0m"
|
||||
#define KRED "\x1B[31m"
|
||||
|
||||
#define failed(...) \
|
||||
printf ("%serror: ", KRED);\
|
||||
printf (__VA_ARGS__);\
|
||||
printf ("\n");\
|
||||
printf ("error: TEST FAILED\n%s", KNRM );\
|
||||
abort();
|
||||
|
||||
#define HIPCHECK(error) \
|
||||
{\
|
||||
hipError_t localError = error; \
|
||||
if (localError != hipSuccess) { \
|
||||
printf("%serror: '%s'(%d) from %s at %s:%d%s\n", \
|
||||
KRED, hipGetErrorString(localError), localError,\
|
||||
#error,__FILE__, __LINE__, KNRM); \
|
||||
failed("API returned error code.");\
|
||||
}\
|
||||
}
|
||||
|
||||
void checkPeer2PeerSupport()
|
||||
{
|
||||
int gpuCount;
|
||||
int canAccessPeer;
|
||||
int p2pCapableDeviceCount=0;
|
||||
|
||||
HIPCHECK(hipGetDeviceCount(&gpuCount));
|
||||
|
||||
if (gpuCount < 2)
|
||||
printf("Peer2Peer application requires atleast 2 gpu devices");
|
||||
|
||||
for (int currentGpu=0; currentGpu<gpuCount; currentGpu++)
|
||||
{
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
|
||||
for (int peerGpu=0; peerGpu<currentGpu; peerGpu++)
|
||||
{
|
||||
if (currentGpu!=peerGpu)
|
||||
{
|
||||
HIPCHECK(hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu));
|
||||
printf ("currentGpu#%d canAccessPeer: peerGpu#%d=%d\n", currentGpu, peerGpu, canAccessPeer);
|
||||
}
|
||||
|
||||
HIPCHECK(hipSetDevice(peerGpu));
|
||||
HIPCHECK(hipDeviceReset());
|
||||
}
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
HIPCHECK(hipDeviceReset());
|
||||
}
|
||||
}
|
||||
|
||||
void enablePeer2Peer(int currentGpu, int peerGpu)
|
||||
{
|
||||
int canAccessPeer;
|
||||
|
||||
// Must be on a multi-gpu system:
|
||||
assert (currentGpu != peerGpu);
|
||||
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu);
|
||||
|
||||
if(canAccessPeer==1){
|
||||
HIPCHECK(hipDeviceEnablePeerAccess(peerGpu, 0));
|
||||
}
|
||||
else
|
||||
printf("peer2peer transfer not possible between the selected gpu devices");
|
||||
}
|
||||
|
||||
void disablePeer2Peer(int currentGpu, int peerGpu)
|
||||
{
|
||||
int canAccessPeer;
|
||||
|
||||
// Must be on a multi-gpu system:
|
||||
assert (currentGpu != peerGpu);
|
||||
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
hipDeviceCanAccessPeer(&canAccessPeer, currentGpu, peerGpu);
|
||||
|
||||
if(canAccessPeer==1){
|
||||
HIPCHECK(hipDeviceDisablePeerAccess(peerGpu));
|
||||
}
|
||||
else
|
||||
printf("peer2peer disable not required");
|
||||
}
|
||||
|
||||
|
||||
__global__ void matrixTranspose_static_shared(hipLaunchParm lp,
|
||||
float *out,
|
||||
float *in,
|
||||
const int width)
|
||||
{
|
||||
__shared__ float sharedMem[WIDTH*WIDTH];
|
||||
|
||||
int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
|
||||
int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
|
||||
|
||||
sharedMem[y * width + x] = in[x * width + y];
|
||||
|
||||
__syncthreads();
|
||||
|
||||
out[y * width + x] = sharedMem[y * width + x];
|
||||
}
|
||||
|
||||
__global__ void matrixTranspose_dynamic_shared(hipLaunchParm lp,
|
||||
float *out,
|
||||
float *in,
|
||||
const int width)
|
||||
{
|
||||
// declare dynamic shared memory
|
||||
HIP_DYNAMIC_SHARED(float, sharedMem)
|
||||
|
||||
int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
|
||||
int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
|
||||
|
||||
sharedMem[y * width + x] = in[x * width + y];
|
||||
|
||||
__syncthreads();
|
||||
|
||||
out[y * width + x] = sharedMem[y * width + x];
|
||||
}
|
||||
|
||||
int main(){
|
||||
|
||||
checkPeer2PeerSupport();
|
||||
|
||||
int gpuCount;
|
||||
int currentGpu, peerGpu;
|
||||
|
||||
HIPCHECK(hipGetDeviceCount(&gpuCount));
|
||||
|
||||
currentGpu = 0;
|
||||
peerGpu = (currentGpu + 1);
|
||||
|
||||
printf ("currentGpu=%d peerGpu=%d (Total no. of gpu = %d)\n", currentGpu, peerGpu, gpuCount);
|
||||
|
||||
float *data[2], *TransposeMatrix[2], *gpuTransposeMatrix[2], *randArray;
|
||||
|
||||
int width = WIDTH;
|
||||
|
||||
randArray = (float*)malloc(NUM * sizeof(float));
|
||||
|
||||
for(int i = 0; i < NUM; i++)
|
||||
{
|
||||
randArray[i] = (float)i*1.0f;
|
||||
}
|
||||
|
||||
enablePeer2Peer(currentGpu, peerGpu);
|
||||
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
TransposeMatrix[0] = (float*)malloc(NUM * sizeof(float));
|
||||
hipMalloc((void**)&gpuTransposeMatrix[0], NUM * sizeof(float));
|
||||
hipMalloc((void**)&data[0], NUM * sizeof(float));
|
||||
hipMemcpy(data[0], randArray, NUM * sizeof(float), hipMemcpyHostToDevice);
|
||||
|
||||
hipLaunchKernel(matrixTranspose_static_shared,
|
||||
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
|
||||
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
|
||||
0, 0,
|
||||
gpuTransposeMatrix[0], data[0], width);
|
||||
|
||||
HIPCHECK(hipSetDevice(peerGpu));
|
||||
TransposeMatrix[1] = (float*)malloc(NUM * sizeof(float));
|
||||
hipMalloc((void**)&gpuTransposeMatrix[1], NUM * sizeof(float));
|
||||
hipMalloc((void**)&data[1], NUM * sizeof(float));
|
||||
hipMemcpy(data[1], gpuTransposeMatrix[0], NUM * sizeof(float), hipMemcpyDeviceToDevice);
|
||||
|
||||
hipLaunchKernel(matrixTranspose_dynamic_shared,
|
||||
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
|
||||
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
|
||||
sizeof(float)*WIDTH*WIDTH, 0,
|
||||
gpuTransposeMatrix[1], data[1], width);
|
||||
|
||||
hipMemcpy(TransposeMatrix[1], gpuTransposeMatrix[1], NUM*sizeof(float), hipMemcpyDeviceToHost);
|
||||
|
||||
hipDeviceSynchronize();
|
||||
|
||||
disablePeer2Peer(currentGpu, peerGpu);
|
||||
|
||||
// verify the results
|
||||
int errors = 0;
|
||||
double eps = 1.0E-6;
|
||||
for (int i = 0; i < NUM; i++) {
|
||||
if (std::abs(randArray[i] - TransposeMatrix[1][i]) > eps ) {
|
||||
printf("%d cpu: %f gpu peered data %f\n",i,randArray[i],TransposeMatrix[1][i]);
|
||||
errors++;
|
||||
}
|
||||
}
|
||||
if (errors!=0) {
|
||||
printf("FAILED: %d errors\n",errors);
|
||||
} else {
|
||||
printf ("Peer2Peer PASSED!\n");
|
||||
}
|
||||
|
||||
free(randArray);
|
||||
for(int i=0;i<2;i++){
|
||||
hipFree(data[i]);
|
||||
hipFree(gpuTransposeMatrix[i]);
|
||||
free(TransposeMatrix[i]);
|
||||
}
|
||||
|
||||
HIPCHECK(hipSetDevice(peerGpu));
|
||||
HIPCHECK(hipDeviceReset());
|
||||
|
||||
HIPCHECK(hipSetDevice(currentGpu));
|
||||
HIPCHECK(hipDeviceReset());
|
||||
|
||||
return 0;
|
||||
}
|
||||
Fai riferimento in un nuovo problema
Block a user