Files

忽略 .git-blame-ignore-revs 中的修訂。點擊 這裡 以繞過並查看正常的 Blame 視圖。

425 行
13 KiB
C++
原始文件 永久連結 標準檢視 歷史記錄

2016-01-08 16:48:40 -05:00
/*
* Copyright 2015 Advanced Micro Devices, Inc.
*
* 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 COPYRIGHT HOLDER(S) OR AUTHOR(S) 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 <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <errno.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include "hsakmt/hsakmt.h"
2016-01-08 16:48:40 -05:00
#include "amdp2ptest.h"
int rdma_fd = -1;
void rdma_open()
{
rdma_fd = open(AMDP2PTEST_DEVICE_PATH, O_RDWR);
if (-1 == rdma_fd ) {
int ret = errno;
fprintf(stderr, "error opening driver (errno=%d/%s)\n", ret, strerror(ret));
exit(EXIT_FAILURE);
}
}
void rdma_close()
{
int retcode = close(rdma_fd);
if (-1 == retcode) {
fprintf(stderr, "error closing driver (errno=%d/%s)\n", retcode, strerror(retcode));
exit(EXIT_FAILURE);
}
rdma_fd = -1;
}
int rdma_map(uint64_t gpu_ptr, size_t size, void **cpu_ptr)
{
int ret = 0;
*cpu_ptr = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_SHARED, rdma_fd, gpu_ptr);
if (*cpu_ptr == NULL) {
int __errno = errno;
*cpu_ptr = NULL;
fprintf(stderr, "Can't BAR, error=%s(%d) size=%zu offset=%llx\n",
strerror(__errno), __errno, size, (long long unsigned)gpu_ptr);
ret = __errno;
}
return ret;
}
int rdma_unmap(void *cpu_ptr, size_t size)
{
int ret = 0;
int retcode = munmap(cpu_ptr, size);
if (-1 == retcode) {
int __errno = errno;
fprintf(stderr, "can't unmap BAR, error=%s(%d) size=%zu\n",
strerror(__errno), __errno, size);
ret = __errno;
}
return ret;
}
void run_rdma_tests(HSAuint32 Node, HsaMemoryProperties *MemoryProperty)
{
printf("Size 0x%lx (%ld MB)\n", MemoryProperty->SizeInBytes,
MemoryProperty->SizeInBytes / (1024 * 1024));
printf("VirtualBaseAddress 0x%lx\n", MemoryProperty->VirtualBaseAddress);
void *cpu_ptr;
int ret = 0;
void *MemoryAddress = 0;
HSAuint64 SizeInBytes = 4096;
HsaMemFlags memFlags = {0};
memFlags.ui32.NonPaged = 1;
memFlags.ui32.CachePolicy = HSA_CACHING_WRITECOMBINED;
memFlags.ui32.NoSubstitute = 1;
memFlags.ui32.PageSize = HSA_PAGE_SIZE_4KB;
// memFlags.ui32.HostAccess = 1;
memFlags.ui32.CoarseGrain = 1;
HSAKMT_STATUS status = hsaKmtAllocMemory(Node,
SizeInBytes,
memFlags,
&MemoryAddress);
if (status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "Failure to allocate memory. Status %d\n", status);
exit(EXIT_FAILURE);
}
printf("Memory allocated. Address 0x%p\n", MemoryAddress);
struct AMDRDMA_IOCTL_GET_PAGE_SIZE_PARAM get_page_size = {0};
get_page_size.addr = (uint64_t) MemoryAddress;
get_page_size.length = SizeInBytes;
ret = ioctl(rdma_fd, AMD2P2PTEST_IOCTL_GET_PAGE_SIZE, &get_page_size);
2016-01-22 14:37:20 -05:00
if (ret != 0)
2016-01-08 16:48:40 -05:00
{
fprintf(stderr,
"AMD2P2PTEST_IOCTL_GET_PAGE_SIZE error (errno=%d/%s)\n",
ret, strerror(ret));
exit(EXIT_FAILURE);
}
printf("GPU Page size: 0x%ld\n", get_page_size.page_size);
struct AMDRDMA_IOCTL_GET_PAGES_PARAM get_cpu_ptr = {0};
get_cpu_ptr.addr = (uint64_t) MemoryAddress;
get_cpu_ptr.length = SizeInBytes;
ret = ioctl(rdma_fd, AMD2P2PTEST_IOCTL_GET_PAGES, &get_cpu_ptr);
2016-01-22 14:37:20 -05:00
if (ret != 0)
2016-01-08 16:48:40 -05:00
{
fprintf(stderr, "AMD2P2PTEST_IOCTL_GET_PAGES error (errno=%d/%s)\n",
ret, strerror(ret));
exit(EXIT_FAILURE);
}
ret = rdma_map((uint64_t)MemoryAddress, 4096, &cpu_ptr);
if (ret < 0)
{
exit(EXIT_FAILURE);
}
printf("CPU Virtual address 0x%p\n", cpu_ptr);
hsaKmtFreeMemory(MemoryAddress, SizeInBytes);
}
int getSysMemorySize(unsigned long *memSize)
{
FILE *meminfo = fopen("/proc/meminfo", "r");
if(meminfo == NULL)
return -1;
char buff[256];
while (fgets(buff, sizeof(buff), meminfo))
{
long ramKB;
if (sscanf(buff, "MemTotal: %ld kB", &ramKB) == 1)
{
*memSize = ramKB * 1024;
break;
}
}
fclose(meminfo);
printf("Total system memory size 0x%lx\n", *memSize);
return 0;
}
/*
* RDMA contiguous memory allocation test
*
* Test steps:
* 1. fragment the entire VRAM, alloc all VRAM using multiple buffers, then free 1 buffer from every
other buffers
* 2. alloc memFlags.ui32.Contiguous=1 buffer for contiguous VRAM allocation
* 3. Call AMD2P2PTEST_IOCTL_GET_PAGES to get contiguous VRAM buffer pages
* 4. Test fails if any above step failed
*/
void run_rdma_contiguous_mem_tests(HSAuint32 Node, HsaMemoryProperties *MemoryProperty)
{
unsigned int *nullPtr = NULL;
unsigned long bufSize = 512ULL << 20;
unsigned long nBuf;
HSAuint64 vramSize;
unsigned long sysMemSize;
HsaMemFlags memFlags = {0};
HSAKMT_STATUS status;
if (getSysMemorySize(&sysMemSize) < 0) {
fprintf(stderr, "Failed to get system memory size\n");
exit(EXIT_FAILURE);
}
status = hsaKmtAvailableMemory(Node, &vramSize);
if (status != HSAKMT_STATUS_SUCCESS) {
fprintf(stderr, "Failed %d to get VRAM size\n", status);
exit(EXIT_FAILURE);
}
if (sysMemSize < (16UL << 30) || vramSize < (4UL << 30)) {
fprintf(stderr, "No enough system memory or VRAM\n");
exit(0);
}
nBuf = vramSize / bufSize;
void **pBuf = (void **)malloc(sizeof(*pBuf) * nBuf);
memFlags.ui32.NonPaged = 1;
for (int i = 0; i < nBuf; i++) {
status = hsaKmtAllocMemory(Node, bufSize, memFlags, &pBuf[i]);
if (status != HSAKMT_STATUS_SUCCESS) {
fprintf(stderr, "Failed %d to alloc buf %d\n", status, i);
exit(EXIT_FAILURE);
}
status = hsaKmtMapMemoryToGPU(pBuf[i], bufSize, NULL);
if (status != HSAKMT_STATUS_SUCCESS) {
fprintf(stderr, "Failed %d to map buf %d\n", status, i);
exit(EXIT_FAILURE);
}
}
//printf("Freeing every other BO to fragment VRAM\n");
for (int i = 0; i < nBuf; i+=2) {
status = hsaKmtUnmapMemoryToGPU(pBuf[i]);
if (status != HSAKMT_STATUS_SUCCESS) {
fprintf(stderr, "Failed %d to unmap buf %d from GPU\n", status, i);
exit(EXIT_FAILURE);
}
status = hsaKmtFreeMemory(pBuf[i], bufSize);
if (status != HSAKMT_STATUS_SUCCESS) {
fprintf(stderr, "Failed %d to free buf %d\n", status, i);
exit(EXIT_FAILURE);
}
}
printf("Node %d Size 0x%lx (%ld MB)\n", Node, MemoryProperty->SizeInBytes,
MemoryProperty->SizeInBytes / (1024 * 1024));
void *cpu_ptr;
int ret = 0;
void *MemoryAddress = 0;
HSAuint64 SizeInBytes = 1UL << 30;
memFlags.ui32.Contiguous = 1;
status = hsaKmtAllocMemory(Node, SizeInBytes, memFlags, &MemoryAddress);
if (status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "Failure to allocate memory 0x%lx. Status %d\n", SizeInBytes, status);
exit(EXIT_FAILURE);
}
status = hsaKmtMapMemoryToGPU(MemoryAddress, SizeInBytes, NULL);
if (status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "Failure to map memory. Status %d\n", status);
exit(EXIT_FAILURE);
}
printf("VRAM allocated. Address %p size 0x%lx bytes\n", MemoryAddress, SizeInBytes);
//printf("Press Enter key to continue\n");
//getchar();
struct AMDRDMA_IOCTL_GET_PAGE_SIZE_PARAM get_page_size = {0};
get_page_size.addr = (uint64_t) MemoryAddress;
get_page_size.length = SizeInBytes;
ret = ioctl(rdma_fd, AMD2P2PTEST_IOCTL_GET_PAGE_SIZE, &get_page_size);
if (ret != 0)
{
fprintf(stderr, "AMD2P2PTEST_IOCTL_GET_PAGE_SIZE error (errno=%d/%s)\n",
ret, strerror(ret));
exit(EXIT_FAILURE);
}
printf("GPU Page size: 0x%ld\n", get_page_size.page_size);
struct AMDRDMA_IOCTL_GET_PAGES_PARAM get_cpu_ptr = {0};
get_cpu_ptr.addr = (uint64_t) MemoryAddress;
get_cpu_ptr.length = SizeInBytes;
ret = ioctl(rdma_fd, AMD2P2PTEST_IOCTL_GET_PAGES, &get_cpu_ptr);
if (ret != 0)
{
fprintf(stderr, "AMD2P2PTEST_IOCTL_GET_PAGES error (errno=%d/%s)\n",
ret, strerror(ret));
//printf("IOCTL_GET_PAGES failed, Press Enter key to continue\n");
//getchar();
exit(EXIT_FAILURE);
}
printf("IOCTL_GET_PAGES return contiguous VRAM address %p size 0x%lx bytes\n", MemoryAddress, SizeInBytes);
printf("Pause to dump page table to check if allocation is contiguous\n");
printf("Press Enter key to continue\n");
getchar();
ret = rdma_map((uint64_t)MemoryAddress, 4096, &cpu_ptr);
if (ret < 0)
{
exit(EXIT_FAILURE);
}
hsaKmtFreeMemory(MemoryAddress, SizeInBytes);
}
2016-01-08 16:48:40 -05:00
int main(void)
{
HsaVersionInfo VersionInfo;
HSAKMT_STATUS status = hsaKmtOpenKFD();
if( status == HSAKMT_STATUS_SUCCESS)
{
status = hsaKmtGetVersion(&VersionInfo);
if(status == HSAKMT_STATUS_SUCCESS)
{
printf("Kernel Interface Major Version: %d\n", VersionInfo.KernelInterfaceMajorVersion);
printf("Kernel Interface Minor Version: %d\n", VersionInfo.KernelInterfaceMinorVersion);
}
}
rdma_open();
HsaSystemProperties SystemProperties = {0};
status = hsaKmtAcquireSystemProperties(&SystemProperties);
if(status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "hsaKmtAcquireSystemProperties call failed. Error: %d\n", status);
exit(EXIT_FAILURE);
}
printf("System properties: Number of nodes: %d\n", SystemProperties.NumNodes);
for (HSAuint32 iNode = 0; iNode < SystemProperties.NumNodes; iNode++)
{
HsaNodeProperties NodeProperties = {0};
status = hsaKmtGetNodeProperties(iNode, &NodeProperties);
if(status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "hsaKmtGetNodeProperties (Node = %d) call failed. Error: %d\n",
iNode, status);
exit(EXIT_FAILURE);
}
printf("Node %d -> Number of Memory Banks = %d\n", iNode,
NodeProperties.NumMemoryBanks);
HsaMemoryProperties* MemoryProperties =
new HsaMemoryProperties[NodeProperties.NumMemoryBanks];
status = hsaKmtGetNodeMemoryProperties(iNode,
NodeProperties.NumMemoryBanks,
MemoryProperties);
if(status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "hsaKmtGetNodeMemoryProperties (Node = %d) call failed. Error: %d\n",
iNode, status);
exit(EXIT_FAILURE);
}
for (HSAuint32 iMemBank = 0; iMemBank < NodeProperties.NumMemoryBanks; iMemBank++)
{
printf("Heap type: %d\n", MemoryProperties[iMemBank].HeapType);
if (MemoryProperties[iMemBank].HeapType == HSA_HEAPTYPE_FRAME_BUFFER_PUBLIC)
{
// We found local memory available for RDMA operation.
// Run some tests on it.
run_rdma_tests(iNode, &MemoryProperties[iMemBank]);
run_rdma_contiguous_mem_tests(iNode, &MemoryProperties[iMemBank]);
2016-01-08 16:48:40 -05:00
}
}
}
status = hsaKmtReleaseSystemProperties();
if(status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "hsaKmtReleaseSystemProperties call failed. Error: %d\n",
status);
exit(EXIT_FAILURE);
}
rdma_close();
status = hsaKmtCloseKFD();
if(status != HSAKMT_STATUS_SUCCESS)
{
fprintf(stderr, "hsaKmtCloseKFD call failed. Error: %d\n", status);
exit(EXIT_FAILURE);
}
return EXIT_SUCCESS;
}