You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

340 lines
12 KiB

//
// Copyright (c) 2017 The Khronos Group Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
#include "common.h"
#include "harness/mt19937.h"
#define GLOBAL_SIZE 65536
static const char *sources[] = {
"__kernel void migrate_kernel(__global uint * restrict a, __global uint * restrict b, __global uint * restrict c)\n"
"{\n"
" size_t i = get_global_id(0);\n"
" a[i] ^= 0x13579bdf;\n"
" b[i] ^= 0x2468ace0;\n"
" c[i] ^= 0x731fec8f;\n"
"}\n"
};
static void
fill_buffer(cl_uint* p, size_t n, MTdata seed)
{
for (size_t i=0; i<n; ++i)
p[i] = (cl_uint)genrand_int32(seed);
}
static bool
check(const char* s, cl_uint* a, cl_uint* e, size_t n)
{
bool ok = true;
for (size_t i=0; ok && i<n; ++i) {
if (a[i] != e[i]) {
log_error("ERROR: %s mismatch at word %u, *%08x vs %08x\n", s, (unsigned int)i, e[i], a[i]);
ok = false;
}
}
return ok;
}
static int
wait_and_release(const char* s, cl_event* evs, int n)
{
cl_int error = clWaitForEvents(n, evs);
if (error == CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST) {
for (int i=0; i<n; ++i) {
cl_int e;
error = clGetEventInfo(evs[i], CL_EVENT_COMMAND_EXECUTION_STATUS, sizeof(cl_int), &e, NULL);
test_error(error, "clGetEventInfo failed");
if (e != CL_COMPLETE) {
log_error("ERROR: %s event %d execution status was %s\n", s, i, IGetErrorString(e));
return e;
}
}
} else
test_error(error, "clWaitForEvents failed");
for (int i=0; i<n; ++i) {
error = clReleaseEvent(evs[i]);
test_error(error, "clReleaseEvent failed");
}
return 0;
}
int test_svm_migrate(cl_device_id deviceID, cl_context c, cl_command_queue queue, int num_elements)
{
cl_uint amem[GLOBAL_SIZE];
cl_uint bmem[GLOBAL_SIZE];
cl_uint cmem[GLOBAL_SIZE];
cl_uint ramem[GLOBAL_SIZE];
cl_uint rbmem[GLOBAL_SIZE];
cl_uint rcmem[GLOBAL_SIZE];
cl_event evs[20];
const size_t global_size = GLOBAL_SIZE;
RandomSeed seed(0);
clContextWrapper context = NULL;
clCommandQueueWrapper queues[MAXQ];
cl_uint num_devices = 0;
clProgramWrapper program;
cl_int error;
error = create_cl_objects(deviceID, &sources[0], &context, &program, &queues[0], &num_devices, CL_DEVICE_SVM_COARSE_GRAIN_BUFFER);
if (error)
return -1;
if (num_devices > 1) {
log_info(" Running on two devices.\n");
} else {
// Ensure we have two distinct queues
cl_device_id did;
error = clGetCommandQueueInfo(queues[0], CL_QUEUE_DEVICE, sizeof(did), (void *)&did, NULL);
test_error(error, "clGetCommandQueueInfo failed");
cl_command_queue_properties cqp;
error = clGetCommandQueueInfo(queues[0], CL_QUEUE_PROPERTIES, sizeof(cqp), &cqp, NULL);
test_error(error, "clGetCommandQueueInfo failed");
cl_queue_properties qp[3] = { CL_QUEUE_PROPERTIES, cqp, 0 };
queues[1] = clCreateCommandQueueWithProperties(context, did, qp, &error);
test_error(error, "clCteateCommandQueueWithProperties failed");
}
clKernelWrapper kernel = clCreateKernel(program, "migrate_kernel", &error);
test_error(error, "clCreateKernel failed");
char* asvm = (char*)clSVMAlloc(context, CL_MEM_READ_WRITE, global_size*sizeof(cl_uint), 16);
if (asvm == NULL) {
log_error("ERROR: clSVMAlloc returned NULL at %s:%d\n", __FILE__, __LINE__);
return -1;
}
char* bsvm = (char *)clSVMAlloc(context, CL_MEM_READ_WRITE, global_size*sizeof(cl_uint), 16);
if (bsvm == NULL) {
log_error("ERROR: clSVMAlloc returned NULL at %s:%d\n", __FILE__, __LINE__);
clSVMFree(context, asvm);
return -1;
}
char* csvm = (char *)clSVMAlloc(context, CL_MEM_READ_WRITE, global_size*sizeof(cl_uint), 16);
if (csvm == NULL) {
log_error("ERROR: clSVMAlloc returned NULL at %s:%d\n", __FILE__, __LINE__);
clSVMFree(context, bsvm);
clSVMFree(context, asvm);
return -1;
}
error = clSetKernelArgSVMPointer(kernel, 0, (void*)asvm);
test_error(error, "clSetKernelArgSVMPointer failed");
error = clSetKernelArgSVMPointer(kernel, 1, (void*)bsvm);
test_error(error, "clSetKernelArgSVMPointer failed");
error = clSetKernelArgSVMPointer(kernel, 2, (void*)csvm);
test_error(error, "clSetKernelArgSVMPointer failed");
// Initialize host copy of data (and result)
fill_buffer(amem, global_size, seed);
fill_buffer(bmem, global_size, seed);
fill_buffer(cmem, global_size, seed);
// Now we're ready to start
{
// First, fill in the data on device0
cl_uint patt[] = { 0, 0, 0, 0};
error = clEnqueueSVMMemFill(queues[0], (void *)asvm, patt, sizeof(patt), global_size*sizeof(cl_uint), 0, NULL, &evs[0]);
test_error(error, "clEnqueueSVMMemFill failed");
error = clEnqueueSVMMemFill(queues[0], (void *)bsvm, patt, sizeof(patt), global_size*sizeof(cl_uint), 0, NULL, &evs[1]);
test_error(error, "clEnqueueSVMMemFill failed");
error = clEnqueueSVMMemFill(queues[0], (void *)csvm, patt, sizeof(patt), global_size*sizeof(cl_uint), 0, NULL, &evs[2]);
test_error(error, "clEnqueueSVMMemFill failed");
}
{
// Now migrate fully to device 1 and discard the data
char* ptrs[] = { asvm, bsvm, csvm };
error = clEnqueueSVMMigrateMem(queues[1], 3, (const void**)ptrs, NULL, CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED, 1, &evs[2], &evs[3]);
test_error(error, "clEnqueueSVMMigrateMem failed");
}
{
// Test host flag
char *ptrs[] = { asvm+1, bsvm+3, csvm+5 };
const size_t szs[] = { 1, 1, 0 };
error = clEnqueueSVMMigrateMem(queues[0], 3, (const void**)ptrs, szs, CL_MIGRATE_MEM_OBJECT_HOST, 1, &evs[3], &evs[4]);
test_error(error, "clEnqueueSVMMigrateMem failed");
}
{
// Next fill with known data
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_WRITE, (void*)asvm, global_size*sizeof(cl_uint), 1, &evs[4], &evs[5]);
test_error(error, "clEnqueueSVMMap failed");
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_WRITE, (void*)bsvm, global_size*sizeof(cl_uint), 0, NULL, &evs[6]);
test_error(error, "clEnqueueSVMMap failed");
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_WRITE, (void*)csvm, global_size*sizeof(cl_uint), 0, NULL, &evs[7]);
test_error(error, "clEnqueueSVMMap failed");
}
error = clFlush(queues[0]);
test_error(error, "clFlush failed");
error = clFlush(queues[1]);
test_error(error, "clFlush failed");
// Check the event command type for clEnqueueSVMMigrateMem (OpenCL 3.0 and
// newer)
Version version = get_device_cl_version(deviceID);
if (version >= Version(3, 0))
{
cl_command_type commandType;
error = clGetEventInfo(evs[3], CL_EVENT_COMMAND_TYPE,
sizeof(commandType), &commandType, NULL);
test_error(error, "clGetEventInfo failed");
if (commandType != CL_COMMAND_SVM_MIGRATE_MEM)
{
log_error("Invalid command type returned for "
"clEnqueueSVMMigrateMem: %X\n",
commandType);
return TEST_FAIL;
}
}
error = wait_and_release("first batch", evs, 8);
if (error)
return -1;
memcpy((void *)asvm, (void *)amem, global_size*sizeof(cl_uint));
memcpy((void *)bsvm, (void *)bmem, global_size*sizeof(cl_uint));
memcpy((void *)csvm, (void *)cmem, global_size*sizeof(cl_uint));
{
error = clEnqueueSVMUnmap(queues[1], (void *)asvm, 0, NULL, &evs[0]);
test_error(error, "clEnqueueSVMUnmap failed");
error = clEnqueueSVMUnmap(queues[1], (void *)bsvm, 0, NULL, &evs[1]);
test_error(error, "clEnqueueSVMUnmap failed");
error = clEnqueueSVMUnmap(queues[1], (void *)csvm, 0, NULL, &evs[2]);
test_error(error, "clEnqueueSVMUnmap failed");
}
{
// Now try some overlapping regions, and operate on the result
char *ptrs[] = { asvm+100, bsvm+17, csvm+1000, asvm+101, bsvm+19, csvm+1017 };
const size_t szs[] = { 13, 23, 43, 3, 7, 11 };
error = clEnqueueSVMMigrateMem(queues[0], 3, (const void**)ptrs, szs, 0, 1, &evs[2], &evs[3]);
test_error(error, "clEnqueueSVMMigrateMem failed");
error = clEnqueueNDRangeKernel(queues[0], kernel, 1, NULL, &global_size, NULL, 0, NULL, &evs[4]);
test_error(error, "clEnqueueNDRangeKernel failed");
}
{
// Now another pair
char *ptrs[] = { asvm+8, bsvm+17, csvm+31, csvm+83 };
const size_t szs[] = { 0, 1, 3, 7 };
error = clEnqueueSVMMigrateMem(queues[1], 4, (const void**)ptrs, szs, 0, 1, &evs[4], &evs[5]);
test_error(error, "clEnqueueSVMMigrateMem failed");
error = clEnqueueNDRangeKernel(queues[1], kernel, 1, NULL, &global_size, NULL, 0, NULL, &evs[6]);
test_error(error, "clEnqueueNDRangeKernel failed");
}
{
// Another pair
char *ptrs[] = { asvm+64, asvm+128, bsvm+64, bsvm+128, csvm, csvm+64 };
const size_t szs[] = { 64, 64, 64, 64, 64, 64 };
error = clEnqueueSVMMigrateMem(queues[0], 6, (const void**)ptrs, szs, 0, 1, &evs[6], &evs[7]);
test_error(error, "clEnqueueSVMMigrateMem failed");
error = clEnqueueNDRangeKernel(queues[0], kernel, 1, NULL, &global_size, NULL, 0, NULL, &evs[8]);
test_error(error, "clEnqueueNDRangeKernel failed");
}
{
// Final pair
char *ptrs[] = { asvm, asvm, bsvm, csvm, csvm };
const size_t szs[] = { 0, 1, 0, 1, 0 };
error = clEnqueueSVMMigrateMem(queues[1], 5, (const void**)ptrs, szs, 0, 1, &evs[8], &evs[9]);
test_error(error, "clEnqueueSVMMigrateMem failed");
error = clEnqueueNDRangeKernel(queues[1], kernel, 1, NULL, &global_size, NULL, 0, NULL, &evs[10]);
test_error(error, "clEnqueueNDRangeKernel failed");
}
{
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_READ, (void*)asvm, global_size*sizeof(cl_uint), 0, NULL, &evs[11]);
test_error(error, "clEnqueueSVMMap failed");
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_READ, (void*)bsvm, global_size*sizeof(cl_uint), 0, NULL, &evs[12]);
test_error(error, "clEnqueueSVMMap failed");
error = clEnqueueSVMMap(queues[1], CL_FALSE, CL_MAP_READ, (void*)csvm, global_size*sizeof(cl_uint), 0, NULL, &evs[13]);
test_error(error, "clEnqueueSVMMap failed");
}
error = clFlush(queues[0]);
test_error(error, "clFlush failed");
error = clFlush(queues[1]);
test_error(error, "clFlush failed");
error = wait_and_release("batch 2", evs, 14);
if (error)
return -1;
// Check kernel results
bool ok = check("memory a", (cl_uint *)asvm, amem, global_size);
ok &= check("memory b", (cl_uint *)bsvm, bmem, global_size);
ok &= check("memory c", (cl_uint *)csvm, cmem, global_size);
{
void *ptrs[] = { asvm, bsvm, csvm };
error = clEnqueueSVMUnmap(queues[1], (void *)asvm, 0, NULL, &evs[0]);
test_error(error, "clEnqueueSVMUnmap failed");
error = clEnqueueSVMUnmap(queues[1], (void *)bsvm, 0, NULL, &evs[1]);
test_error(error, "clEnqueueSVMUnmap failed");
error = clEnqueueSVMUnmap(queues[1], (void *)csvm, 0, NULL, &evs[2]);
test_error(error, "clEnqueueSVMUnmap failed");
error = clEnqueueSVMFree(queues[1], 3, ptrs, NULL, NULL, 0, NULL, &evs[3]);
}
error = clFlush(queues[1]);
test_error(error, "clFlush failed");
error = wait_and_release("batch 3", evs, 4);
if (error)
return -1;
// The wrappers will clean up the rest
return ok ? 0 : -1;
}