Fix HIP Streams Duplication Error (#313)
* Fix stream duplication and fixed tests * Added comments to explain stream.cpp code, change stream nullptr check to occur in update table to prevent readding null stream, simplified hip-streams bin file code, add destroyStreams to hip-streams bin file code * Removed roctx from CMakeLists.txt * Updated documentation * Fix documentation * Removed update_table for HIP compiler table and updated stream.cpp to remove support for HIP compiler table * Added runtime initialization check for HIP * Changed tool name, working on fixing memory management * Added context for counter collection kernel rename combination * Changed name from map to set and changed description * Fix documentation description for group-by-queue * Merged memory copy and kernel operations onto a single track when on the same stream * Updated perfetto output to remove hardware information from track name to merge all memory copy and kernel operations on the same stream to the same track: * Most pr comments addressed * Added filter for counter collection and removed kernel buffer tracing hack * Added PR comment fixes --------- Co-authored-by: Madsen, Jonathan <Jonathan.Madsen@amd.com>
This commit is contained in:
Trowbridge, Ian
GitHub
@@ -0,0 +1,41 @@
|
||||
#
|
||||
#
|
||||
#
|
||||
cmake_minimum_required(VERSION 3.21.0 FATAL_ERROR)
|
||||
|
||||
if(NOT CMAKE_HIP_COMPILER)
|
||||
find_program(
|
||||
amdclangpp_EXECUTABLE
|
||||
NAMES amdclang++
|
||||
HINTS ${ROCM_PATH} ENV ROCM_PATH /opt/rocm
|
||||
PATHS ${ROCM_PATH} ENV ROCM_PATH /opt/rocm
|
||||
PATH_SUFFIXES bin llvm/bin NO_CACHE)
|
||||
mark_as_advanced(amdclangpp_EXECUTABLE)
|
||||
|
||||
if(amdclangpp_EXECUTABLE)
|
||||
set(CMAKE_HIP_COMPILER "${amdclangpp_EXECUTABLE}")
|
||||
endif()
|
||||
endif()
|
||||
|
||||
project(rocprofiler-sdk-tests-bin-hip-streams LANGUAGES CXX HIP)
|
||||
|
||||
foreach(_TYPE DEBUG MINSIZEREL RELEASE RELWITHDEBINFO)
|
||||
if("${CMAKE_HIP_FLAGS_${_TYPE}}" STREQUAL "")
|
||||
set(CMAKE_HIP_FLAGS_${_TYPE} "${CMAKE_CXX_FLAGS_${_TYPE}}")
|
||||
endif()
|
||||
endforeach()
|
||||
|
||||
set(CMAKE_CXX_STANDARD 17)
|
||||
set(CMAKE_CXX_EXTENSIONS OFF)
|
||||
set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
||||
set(CMAKE_HIP_STANDARD 17)
|
||||
set(CMAKE_HIP_EXTENSIONS OFF)
|
||||
set(CMAKE_HIP_STANDARD_REQUIRED ON)
|
||||
|
||||
set_source_files_properties(compute_comm_overlap.cpp PROPERTIES LANGUAGE HIP)
|
||||
add_executable(hip-streams)
|
||||
target_sources(hip-streams PRIVATE compute_comm_overlap.cpp)
|
||||
target_link_libraries(hip-streams PRIVATE rocprofiler-sdk::tests-build-flags)
|
||||
|
||||
find_package(Threads REQUIRED)
|
||||
target_link_libraries(hip-streams PRIVATE Threads::Threads)
|
||||
@@ -0,0 +1,138 @@
|
||||
#include <math.h>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include "hip/hip_runtime.h"
|
||||
|
||||
#define BLOCKDIM 64
|
||||
|
||||
/* Macro for checking GPU API return values */
|
||||
#define HIP_ASSERT(call) \
|
||||
do \
|
||||
{ \
|
||||
hipError_t gpuErr = call; \
|
||||
if(hipSuccess != gpuErr) \
|
||||
{ \
|
||||
printf( \
|
||||
"GPU API Error - %s:%d: '%s'\n", __FILE__, __LINE__, hipGetErrorString(gpuErr)); \
|
||||
exit(1); \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
// HIP kernel. Each thread takes care of one element of input
|
||||
__global__ void
|
||||
cube(double* input, double* output, size_t offset, size_t elements_per_stream)
|
||||
{
|
||||
size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
size_t gstride = blockDim.x * gridDim.x;
|
||||
|
||||
// Span all elements assigned to this stream
|
||||
for(size_t id = tid + offset; id < offset + elements_per_stream; id += gstride)
|
||||
for(size_t i = 0; i < 1000; ++i)
|
||||
output[id] = input[id] * input[id] * input[id];
|
||||
}
|
||||
|
||||
int
|
||||
main()
|
||||
{
|
||||
// number of streams
|
||||
const int num_streams = 8;
|
||||
|
||||
// Number of threads in each thread block
|
||||
const int blockSize = 512;
|
||||
|
||||
// Size of vectors
|
||||
int n = 100000000;
|
||||
int elements_per_stream = n / num_streams;
|
||||
int bytes_per_stream = elements_per_stream * sizeof(double);
|
||||
|
||||
// Host input vectors
|
||||
double* h_input1{nullptr};
|
||||
// Host output vector
|
||||
double* h_output1{nullptr};
|
||||
|
||||
// Device input vectors
|
||||
double* d_input1{nullptr};
|
||||
// Device output vector
|
||||
double* d_output1{nullptr};
|
||||
|
||||
// Creating events for timers
|
||||
hipEvent_t start{}, stop{};
|
||||
HIP_ASSERT(hipEventCreate(&start));
|
||||
HIP_ASSERT(hipEventCreate(&stop));
|
||||
|
||||
// Creating streams
|
||||
hipStream_t streams[num_streams];
|
||||
for(int i = 0; i < num_streams; ++i)
|
||||
{
|
||||
HIP_ASSERT(hipStreamCreate(&streams[i]));
|
||||
}
|
||||
|
||||
// Size, in bytes, of each vector
|
||||
size_t bytes = n * sizeof(double);
|
||||
|
||||
// Allocate page locked memory for these vectors on host
|
||||
HIP_ASSERT(hipHostMalloc(&h_input1, bytes));
|
||||
HIP_ASSERT(hipHostMalloc(&h_output1, bytes));
|
||||
|
||||
// Allocate memory for each vector on GPU
|
||||
HIP_ASSERT(hipMalloc(&d_input1, bytes));
|
||||
HIP_ASSERT(hipMalloc(&d_output1, bytes));
|
||||
|
||||
// Initialize vectors on host
|
||||
for(int i = 0; i < n; i++)
|
||||
{
|
||||
h_input1[i] = sin(i);
|
||||
}
|
||||
|
||||
// Number of thread blocks in grid
|
||||
const int gridSizePerStream = 104; //(int)ceil((float)elements_per_stream/blockSize);
|
||||
|
||||
HIP_ASSERT(hipEventRecord(start));
|
||||
// split H2D copies and kernel calls into separate loops
|
||||
for(int i = 0; i < num_streams; i++)
|
||||
{
|
||||
int offset = i * elements_per_stream;
|
||||
HIP_ASSERT(hipMemcpyAsync(&d_input1[offset],
|
||||
&h_input1[offset],
|
||||
bytes_per_stream,
|
||||
hipMemcpyHostToDevice,
|
||||
streams[i]));
|
||||
}
|
||||
for(int i = 0; i < num_streams; i++)
|
||||
{
|
||||
int offset = i * elements_per_stream;
|
||||
cube<<<gridSizePerStream, blockSize, 0, streams[i]>>>(
|
||||
d_input1, d_output1, offset, elements_per_stream);
|
||||
}
|
||||
for(int i = 0; i < num_streams; i++)
|
||||
{
|
||||
int offset = i * elements_per_stream;
|
||||
HIP_ASSERT(hipMemcpyAsync(&h_output1[offset],
|
||||
&d_output1[offset],
|
||||
bytes_per_stream,
|
||||
hipMemcpyDeviceToHost,
|
||||
streams[i]));
|
||||
}
|
||||
|
||||
HIP_ASSERT(hipEventRecord(stop));
|
||||
HIP_ASSERT(hipEventSynchronize(stop));
|
||||
|
||||
float milliseconds = 0;
|
||||
HIP_ASSERT(hipEventElapsedTime(&milliseconds, start, stop));
|
||||
|
||||
// Release device memory
|
||||
HIP_ASSERT(hipFree(d_input1));
|
||||
HIP_ASSERT(hipFree(d_output1));
|
||||
|
||||
// Release host memory
|
||||
HIP_ASSERT(hipHostFree(h_input1));
|
||||
HIP_ASSERT(hipHostFree(h_output1));
|
||||
|
||||
// Destroy streams
|
||||
for(int i = 0; i < num_streams; ++i)
|
||||
{
|
||||
HIP_ASSERT(hipStreamDestroy(streams[i]));
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user