| // |
| // 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 <stdio.h> |
| #include <stdlib.h> |
| |
| #include "procs.h" |
| #include "harness/errorHelpers.h" |
| #include "harness/testHarness.h" |
| |
| #define MAX_SUB_DEVICES 16 // Limit the sub-devices to ensure no out of resource errors. |
| #define BUFFER_SIZE 1024 |
| |
| // Kernel source code |
| static const char *buffer_migrate_kernel_code = |
| "__kernel void test_buffer_migrate(__global uint *dst, __global uint *src1, __global uint *src2, uint x)\n" |
| "{\n" |
| " int tid = get_global_id(0);\n" |
| " dst[tid] = src1[tid] ^ src2[tid] ^ x;\n" |
| "}\n"; |
| |
| enum migrations { MIGRATE_PREFERRED, // migrate to the preferred sub-device |
| MIGRATE_NON_PREFERRED, // migrate to a randomly chosen non-preferred sub-device |
| MIGRATE_RANDOM, // migrate to a randomly chosen sub-device with randomly chosen flags |
| NUMBER_OF_MIGRATIONS }; |
| |
| static cl_mem init_buffer(cl_command_queue cmd_q, cl_mem buffer, cl_uint *data) |
| { |
| cl_int err; |
| |
| if (buffer) { |
| if ((err = clEnqueueWriteBuffer(cmd_q, buffer, CL_TRUE, 0, sizeof(cl_uint)*BUFFER_SIZE, data, 0, NULL, NULL)) != CL_SUCCESS) { |
| print_error(err, "Failed on enqueue write of buffer data."); |
| } |
| } |
| return buffer; |
| } |
| |
| static cl_int migrateMemObject(enum migrations migrate, cl_command_queue *queues, cl_mem *mem_objects, cl_uint num_devices, cl_mem_migration_flags *flags, MTdata d) |
| { |
| cl_uint i, j; |
| cl_int err = CL_SUCCESS; |
| |
| for (i=0; i<num_devices; i++) { |
| j = genrand_int32(d) % num_devices; |
| flags[i] = 0; |
| switch (migrate) { |
| case MIGRATE_PREFERRED: |
| // Force the device to be preferred |
| j = i; |
| break; |
| case MIGRATE_NON_PREFERRED: |
| // Coerce the device to be non-preferred |
| if ((j == i) && (num_devices > 1)) j = (j+1) % num_devices; |
| break; |
| case MIGRATE_RANDOM: |
| // Choose a random set of flags |
| flags[i] = (cl_mem_migration_flags)(genrand_int32(d) & (CL_MIGRATE_MEM_OBJECT_HOST | CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED));; |
| break; |
| default: log_error("Unhandled migration type: %d\n", migrate); return -1; |
| } |
| if ((err = clEnqueueMigrateMemObjects(queues[j], 1, (const cl_mem *)(&mem_objects[i]), flags[i], 0, NULL, NULL)) != CL_SUCCESS) { |
| print_error(err, "Failed migrating memory object."); |
| } |
| } |
| return err; |
| } |
| |
| static cl_int restoreBuffer(cl_command_queue *queues, cl_mem *buffers, cl_uint num_devices, cl_mem_migration_flags *flags, cl_uint *buffer) |
| { |
| cl_uint i, j; |
| cl_int err; |
| |
| // If the buffer was previously migrated with undefined content, reload the content. |
| |
| for (i=0; i<num_devices; i++) { |
| if (flags[i] & CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED) { |
| if ((err = clEnqueueWriteBuffer(queues[i], buffers[i], CL_TRUE, 0, sizeof(cl_uint)*BUFFER_SIZE, buffer, 0, NULL, NULL)) != CL_SUCCESS) { |
| print_error(err, "Failed on restoration enqueue write of buffer data."); |
| return err; |
| } |
| } |
| } |
| return CL_SUCCESS; |
| } |
| |
| int test_buffer_migrate(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements) |
| { |
| int failed = 0; |
| cl_uint i, j; |
| cl_int err; |
| cl_uint max_sub_devices = 0; |
| cl_uint num_devices, num_devices_limited; |
| cl_uint A[BUFFER_SIZE], B[BUFFER_SIZE], C[BUFFER_SIZE]; |
| cl_uint test_number = 1; |
| cl_device_affinity_domain domain, domains; |
| cl_device_id *devices; |
| cl_command_queue *queues; |
| cl_mem_migration_flags *flagsA, *flagsB, *flagsC; |
| cl_device_partition_property property[] = {CL_DEVICE_PARTITION_BY_AFFINITY_DOMAIN, 0, 0}; |
| cl_mem *bufferA, *bufferB, *bufferC; |
| cl_program program = NULL; |
| cl_kernel kernel = NULL; |
| cl_context ctx = NULL; // context for all sub-devices |
| enum migrations migrateA, migrateB, migrateC; |
| MTdata d = init_genrand(gRandomSeed); |
| const size_t wgs[1] = {BUFFER_SIZE}; |
| |
| /* Allocate arrays whose size varies according to the maximum number of sub-devices */ |
| if ((err = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(max_sub_devices), &max_sub_devices, NULL)) != CL_SUCCESS) { |
| print_error(err, "clGetDeviceInfo(CL_DEVICE_MAX_COMPUTE_UNITS) failed"); |
| return -1; |
| } |
| if (max_sub_devices < 1) { |
| log_error("ERROR: Invalid number of compute units returned.\n"); |
| return -1; |
| } |
| devices = (cl_device_id *)malloc(max_sub_devices * sizeof(cl_device_id)); |
| queues = (cl_command_queue *)malloc(max_sub_devices * sizeof(cl_command_queue)); |
| flagsA = (cl_mem_migration_flags *)malloc(max_sub_devices * sizeof(cl_mem_migration_flags)); |
| flagsB = (cl_mem_migration_flags *)malloc(max_sub_devices * sizeof(cl_mem_migration_flags)); |
| flagsC = (cl_mem_migration_flags *)malloc(max_sub_devices * sizeof(cl_mem_migration_flags)); |
| bufferA = (cl_mem *)malloc(max_sub_devices * sizeof(cl_mem)); |
| bufferB = (cl_mem *)malloc(max_sub_devices * sizeof(cl_mem)); |
| bufferC = (cl_mem *)malloc(max_sub_devices * sizeof(cl_mem)); |
| |
| if ((devices == NULL) || (queues == NULL) || |
| (flagsA == NULL) || (flagsB == NULL) || (flagsC == NULL) || |
| (bufferA == NULL) || (bufferB == NULL) || (bufferC == NULL)) { |
| log_error("ERROR: Failed to successfully allocate required local buffers.\n"); |
| failed = -1; |
| goto cleanup_allocations; |
| } |
| |
| for (i=0; i<max_sub_devices; i++) { |
| devices[i] = NULL; |
| queues [i] = NULL; |
| bufferA[i] = bufferB[i] = bufferC[i] = NULL; |
| } |
| |
| for (i=0; i<BUFFER_SIZE; i++) { |
| A[i] = genrand_int32(d); |
| B[i] = genrand_int32(d); |
| } |
| |
| // Attempt to partition the device along each of the allowed affinity domain. |
| if ((err = clGetDeviceInfo(deviceID, CL_DEVICE_PARTITION_AFFINITY_DOMAIN, sizeof(domains), &domains, NULL)) != CL_SUCCESS) { |
| print_error(err, "clGetDeviceInfo(CL_PARTITION_AFFINITY_DOMAIN) failed"); |
| return -1; |
| } |
| |
| domains &= (CL_DEVICE_AFFINITY_DOMAIN_L4_CACHE | CL_DEVICE_AFFINITY_DOMAIN_L3_CACHE | |
| CL_DEVICE_AFFINITY_DOMAIN_L2_CACHE | CL_DEVICE_AFFINITY_DOMAIN_L1_CACHE | CL_DEVICE_AFFINITY_DOMAIN_NUMA); |
| |
| do { |
| if (domains) { |
| for (domain = 1; (domain & domains) == 0; domain <<= 1) {}; |
| domains &= ~domain; |
| } else { |
| domain = 0; |
| } |
| |
| // Determine the number of partitions for the device given the specific domain. |
| if (domain) { |
| property[1] = domain; |
| err = clCreateSubDevices(deviceID, (const cl_device_partition_property *)property, -1, NULL, &num_devices); |
| if ((err != CL_SUCCESS) || (num_devices == 0)) { |
| print_error(err, "Obtaining the number of partions by affinity failed."); |
| failed = 1; |
| goto cleanup; |
| } |
| } else { |
| num_devices = 1; |
| } |
| |
| if (num_devices > 1) { |
| // Create each of the sub-devices and a corresponding context. |
| if ((err = clCreateSubDevices(deviceID, (const cl_device_partition_property *)property, num_devices, devices, &num_devices)) != CL_SUCCESS) { |
| print_error(err, "Failed creating sub devices."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| // Create a context containing all the sub-devices |
| ctx = clCreateContext(NULL, num_devices, devices, notify_callback, NULL, &err); |
| if (ctx == NULL) { |
| print_error(err, "Failed creating context containing the sub-devices."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| // Create a command queue for each sub-device |
| for (i=0; i<num_devices; i++) { |
| if (devices[i]) { |
| if ((queues[i] = clCreateCommandQueue(ctx, devices[i], 0, &err)) == NULL) { |
| print_error(err, "Failed creating command queues."); |
| failed = 1; |
| goto cleanup; |
| } |
| } |
| } |
| } else { |
| // No partitioning available. Just exercise the APIs on a single device. |
| devices[0] = deviceID; |
| queues[0] = queue; |
| ctx = context; |
| } |
| |
| // Build the kernel program. |
| if ((err = create_single_kernel_helper(ctx, &program, &kernel, 1, |
| &buffer_migrate_kernel_code, |
| "test_buffer_migrate"))) |
| { |
| print_error(err, "Failed creating kernel."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| num_devices_limited = num_devices; |
| |
| // Allocate memory buffers. 3 buffers (2 input, 1 output) for each sub-device. |
| // If we run out of memory, then restrict the number of sub-devices to be tested. |
| for (i=0; i<num_devices; i++) { |
| bufferA[i] = init_buffer(queues[i], clCreateBuffer(ctx, (CL_MEM_READ_ONLY | CL_MEM_ALLOC_HOST_PTR), sizeof(cl_uint) * BUFFER_SIZE, NULL, &err), A); |
| bufferB[i] = init_buffer(queues[i], clCreateBuffer(ctx, (CL_MEM_READ_ONLY | CL_MEM_ALLOC_HOST_PTR), sizeof(cl_uint) * BUFFER_SIZE, NULL, &err), B); |
| bufferC[i] = clCreateBuffer(ctx, (CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR), sizeof(cl_uint) * BUFFER_SIZE, NULL, &err); |
| |
| if ((bufferA[i] == NULL) || (bufferB[i] == NULL) || (bufferC[i] == NULL)) { |
| if (i == 0) { |
| log_error("Failed to allocate even 1 set of buffers.\n"); |
| failed = 1; |
| goto cleanup; |
| } |
| num_devices_limited = i; |
| break; |
| } |
| } |
| |
| // For each partition, we will execute the test kernel with each of the 3 buffers migrated to one of the migrate options |
| for (migrateA=(enum migrations)(0); migrateA<NUMBER_OF_MIGRATIONS; migrateA = (enum migrations)((int)migrateA + 1)) { |
| if (migrateMemObject(migrateA, queues, bufferA, num_devices_limited, flagsA, d) != CL_SUCCESS) { |
| failed = 1; |
| goto cleanup; |
| } |
| for (migrateC=(enum migrations)(0); migrateC<NUMBER_OF_MIGRATIONS; migrateC = (enum migrations)((int)migrateC + 1)) { |
| if (migrateMemObject(migrateC, queues, bufferC, num_devices_limited, flagsC, d) != CL_SUCCESS) { |
| failed = 1; |
| goto cleanup; |
| } |
| for (migrateB=(enum migrations)(0); migrateB<NUMBER_OF_MIGRATIONS; migrateB = (enum migrations)((int)migrateB + 1)) { |
| if (migrateMemObject(migrateB, queues, bufferB, num_devices_limited, flagsB, d) != CL_SUCCESS) { |
| failed = 1; |
| goto cleanup; |
| } |
| // Run the test on each of the partitions. |
| for (i=0; i<num_devices_limited; i++) { |
| cl_uint x; |
| |
| x = i + test_number; |
| |
| if ((err = clSetKernelArg(kernel, 0, sizeof(cl_mem), (const void *)&bufferC[i])) != CL_SUCCESS) { |
| print_error(err, "Failed set kernel argument 0."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| if ((err = clSetKernelArg(kernel, 1, sizeof(cl_mem), (const void *)&bufferA[i])) != CL_SUCCESS) { |
| print_error(err, "Failed set kernel argument 1."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| if ((err = clSetKernelArg(kernel, 2, sizeof(cl_mem), (const void *)&bufferB[i])) != CL_SUCCESS) { |
| print_error(err, "Failed set kernel argument 2."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| if ((err = clSetKernelArg(kernel, 3, sizeof(cl_uint), (const void *)&x)) != CL_SUCCESS) { |
| print_error(err, "Failed set kernel argument 3."); |
| failed = 1; |
| goto cleanup; |
| } |
| |
| if ((err = clEnqueueNDRangeKernel(queues[i], kernel, 1, NULL, wgs, NULL, 0, NULL, NULL)) != CL_SUCCESS) { |
| print_error(err, "Failed enqueueing the NDRange kernel."); |
| failed = 1; |
| goto cleanup; |
| } |
| } |
| // Verify the results as long as neither input is an undefined migration |
| for (i=0; i<num_devices_limited; i++, test_number++) { |
| if (((flagsA[i] | flagsB[i]) & CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED) == 0) { |
| if ((err = clEnqueueReadBuffer(queues[i], bufferC[i], CL_TRUE, 0, sizeof(cl_uint)*BUFFER_SIZE, C, 0, NULL, NULL)) != CL_SUCCESS) { |
| print_error(err, "Failed reading output buffer."); |
| failed = 1; |
| goto cleanup; |
| } |
| for (j=0; j<BUFFER_SIZE; j++) { |
| cl_uint expected; |
| |
| expected = A[j] ^ B[j] ^ test_number; |
| if (C[j] != expected) { |
| log_error("Failed on device %d, work item %4d, expected 0x%08x got 0x%08x (0x%08x ^ 0x%08x ^ 0x%08x)\n", i, j, expected, C[j], A[j], B[j], test_number); |
| failed = 1; |
| } |
| } |
| if (failed) goto cleanup; |
| } |
| } |
| |
| if (restoreBuffer(queues, bufferB, num_devices_limited, flagsB, B) != CL_SUCCESS) { |
| failed = 1; |
| goto cleanup; |
| } |
| } |
| } |
| if (restoreBuffer(queues, bufferA, num_devices_limited, flagsA, A) != CL_SUCCESS) { |
| failed = 1; |
| goto cleanup; |
| } |
| } |
| |
| cleanup: |
| // Clean up all the allocted resources create by the test. This includes sub-devices, |
| // command queues, and memory buffers. |
| |
| for (i=0; i<max_sub_devices; i++) { |
| // Memory buffer cleanup |
| if (bufferA[i]) { |
| if ((err = clReleaseMemObject(bufferA[i])) != CL_SUCCESS) { |
| print_error(err, "Failed releasing memory object."); |
| failed = 1; |
| } |
| } |
| if (bufferB[i]) { |
| if ((err = clReleaseMemObject(bufferB[i])) != CL_SUCCESS) { |
| print_error(err, "Failed releasing memory object."); |
| failed = 1; |
| } |
| } |
| if (bufferC[i]) { |
| if ((err = clReleaseMemObject(bufferC[i])) != CL_SUCCESS) { |
| print_error(err, "Failed releasing memory object."); |
| failed = 1; |
| } |
| } |
| |
| |
| if (num_devices > 1) { |
| // Command queue cleanup |
| if (queues[i]) { |
| if ((err = clReleaseCommandQueue(queues[i])) != CL_SUCCESS) { |
| print_error(err, "Failed releasing command queue."); |
| failed = 1; |
| } |
| } |
| |
| // Sub-device cleanup |
| if (devices[i]) { |
| if ((err = clReleaseDevice(devices[i])) != CL_SUCCESS) { |
| print_error(err, "Failed releasing sub device."); |
| failed = 1; |
| } |
| } |
| devices[i] = 0; |
| } |
| } |
| |
| // Context, program, and kernel cleanup |
| if (program) { |
| if ((err = clReleaseProgram(program)) != CL_SUCCESS) { |
| print_error(err, "Failed releasing program."); |
| failed = 1; |
| } |
| program = NULL; |
| } |
| |
| if (kernel) { |
| if ((err = clReleaseKernel(kernel)) != CL_SUCCESS) { |
| print_error(err, "Failed releasing kernel."); |
| failed = 1; |
| } |
| kernel = NULL; |
| } |
| |
| if (ctx && (ctx != context)) { |
| if ((err = clReleaseContext(ctx)) != CL_SUCCESS) { |
| print_error(err, "Failed releasing context."); |
| failed = 1; |
| } |
| } |
| ctx = NULL; |
| |
| if (failed) goto cleanup_allocations; |
| } while (domains); |
| |
| cleanup_allocations: |
| if (devices) free(devices); |
| if (queues) free(queues); |
| if (flagsA) free(flagsA); |
| if (flagsB) free(flagsB); |
| if (flagsC) free(flagsC); |
| if (bufferA) free(bufferA); |
| if (bufferB) free(bufferB); |
| if (bufferC) free(bufferC); |
| |
| return ((failed) ? -1 : 0); |
| } |
