|
15 | 15 |
|
16 | 16 | #include <umf/ipc.h> |
17 | 17 | #include <umf/memory_pool.h> |
| 18 | +#include <umf/memory_props.h> |
18 | 19 | #include <umf/pools/pool_disjoint.h> |
19 | 20 | #include <umf/pools/pool_proxy.h> |
20 | 21 | #include <umf/pools/pool_scalable.h> |
@@ -438,6 +439,47 @@ static void do_ipc_get_put_benchmark(alloc_t *allocs, size_t num_allocs, |
438 | 439 | } |
439 | 440 | } |
440 | 441 |
|
| 442 | +static void do_umf_mem_props_benchmark(ze_context_handle_t context, |
| 443 | + bool use_umf, alloc_t *allocs, |
| 444 | + size_t num_allocs, size_t repeats) { |
| 445 | + assert(context != NULL); |
| 446 | + |
| 447 | + for (size_t r = 0; r < repeats * 10; ++r) { |
| 448 | + for (size_t i = 0; i < num_allocs; ++i) { |
| 449 | + if (use_umf) { |
| 450 | + umf_memory_properties_handle_t props_handle = NULL; |
| 451 | + umf_result_t res = |
| 452 | + umfGetMemoryPropertiesHandle(allocs[i].ptr, &props_handle); |
| 453 | + (void)res; |
| 454 | + assert(res == UMF_RESULT_SUCCESS); |
| 455 | + |
| 456 | + umf_usm_memory_type_t type = UMF_MEMORY_TYPE_UNKNOWN; |
| 457 | + res = umfGetMemoryProperty( |
| 458 | + props_handle, UMF_MEMORY_PROPERTY_POINTER_TYPE, &type); |
| 459 | + assert(res == UMF_RESULT_SUCCESS); |
| 460 | + if (type != UMF_MEMORY_TYPE_DEVICE) { |
| 461 | + fprintf(stderr, |
| 462 | + "error: unexpected alloc_props.type value: %d\n", |
| 463 | + type); |
| 464 | + exit(-1); |
| 465 | + } |
| 466 | + } else { |
| 467 | + ze_memory_allocation_properties_t alloc_props = {0}; |
| 468 | + ze_device_handle_t device = 0; |
| 469 | + // calls zeMemGetAllocProperties() |
| 470 | + utils_ze_get_mem_props(context, allocs[i].ptr, &alloc_props, |
| 471 | + &device); |
| 472 | + if (alloc_props.type != ZE_MEMORY_TYPE_DEVICE) { |
| 473 | + fprintf(stderr, |
| 474 | + "error: unexpected alloc_props.type value: %d\n", |
| 475 | + alloc_props.type); |
| 476 | + exit(-1); |
| 477 | + } |
| 478 | + } |
| 479 | + } |
| 480 | + } |
| 481 | +} |
| 482 | + |
441 | 483 | static int create_level_zero_params(ze_context_handle_t *context, |
442 | 484 | ze_device_handle_t *device) { |
443 | 485 | uint32_t driver_idx = 0; |
@@ -623,6 +665,198 @@ UBENCH_EX(ipc, disjoint_pool_with_level_zero_provider) { |
623 | 665 | err_destroy_context: |
624 | 666 | utils_ze_destroy_context(context); |
625 | 667 | } |
| 668 | + |
| 669 | +UBENCH_EX(mem_props, level_zero) { |
| 670 | + const size_t BUFFER_SIZE = 100; |
| 671 | + const size_t N_BUFFERS = 1000; |
| 672 | + |
| 673 | + alloc_t *allocs = alloc_array(N_BUFFERS); |
| 674 | + if (allocs == NULL) { |
| 675 | + fprintf(stderr, "error: alloc_array() failed\n"); |
| 676 | + } |
| 677 | + |
| 678 | + ze_context_handle_t context = NULL; |
| 679 | + ze_device_handle_t device = NULL; |
| 680 | + int ret = create_level_zero_params(&context, &device); |
| 681 | + if (ret != 0) { |
| 682 | + fprintf(stderr, "error: create_level_zero_params() failed\n"); |
| 683 | + exit(-1); |
| 684 | + } |
| 685 | + |
| 686 | + ze_device_mem_alloc_desc_t dev_desc = { |
| 687 | + .stype = ZE_STRUCTURE_TYPE_DEVICE_MEM_ALLOC_DESC, |
| 688 | + .pNext = NULL, |
| 689 | + .flags = 0, |
| 690 | + .ordinal = 0}; |
| 691 | + |
| 692 | + for (size_t i = 0; i < N_BUFFERS; ++i) { |
| 693 | + ze_result_t ze_result = zeMemAllocDevice( |
| 694 | + context, &dev_desc, BUFFER_SIZE, 0, device, &allocs[i].ptr); |
| 695 | + if (ze_result != ZE_RESULT_SUCCESS) { |
| 696 | + fprintf(stderr, "error: zeMemAllocDevice() failed\n"); |
| 697 | + } |
| 698 | + allocs[i].size = BUFFER_SIZE; |
| 699 | + } |
| 700 | + |
| 701 | + do_umf_mem_props_benchmark(context, false, allocs, N_BUFFERS, |
| 702 | + 1); // WARMUP |
| 703 | + UBENCH_DO_BENCHMARK() { |
| 704 | + do_umf_mem_props_benchmark(context, false, allocs, N_BUFFERS, |
| 705 | + N_ITERATIONS); |
| 706 | + } |
| 707 | + |
| 708 | + for (size_t i = 0; i < N_BUFFERS; ++i) { |
| 709 | + zeMemFree(context, allocs[i].ptr); |
| 710 | + } |
| 711 | + |
| 712 | + free(allocs); |
| 713 | + utils_ze_destroy_context(context); |
| 714 | +} |
| 715 | + |
| 716 | +UBENCH_EX(mem_props, disjoint_pool_with_level_zero_provider_use_umf) { |
| 717 | + const size_t BUFFER_SIZE = 4 * 1024; |
| 718 | + const size_t N_BUFFERS = 1000; |
| 719 | + umf_result_t umf_result; |
| 720 | + ze_context_handle_t context = NULL; |
| 721 | + ze_device_handle_t device = NULL; |
| 722 | + umf_level_zero_memory_provider_params_handle_t level_zero_params = NULL; |
| 723 | + |
| 724 | + int ret = create_level_zero_params(&context, &device); |
| 725 | + if (ret != 0) { |
| 726 | + fprintf(stderr, "error: create_level_zero_params() failed\n"); |
| 727 | + exit(-1); |
| 728 | + } |
| 729 | + |
| 730 | + umf_result = umfLevelZeroMemoryProviderParamsCreate(&level_zero_params); |
| 731 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 732 | + fprintf(stderr, |
| 733 | + "error: umfLevelZeroMemoryProviderParamsCreate() failed\n"); |
| 734 | + goto err_destroy_context; |
| 735 | + } |
| 736 | + |
| 737 | + umf_result = |
| 738 | + umfLevelZeroMemoryProviderParamsSetContext(level_zero_params, context); |
| 739 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 740 | + fprintf(stderr, |
| 741 | + "error: umfLevelZeroMemoryProviderParamsSetContext() failed\n"); |
| 742 | + goto err_destroy_params; |
| 743 | + } |
| 744 | + |
| 745 | + umf_result = |
| 746 | + umfLevelZeroMemoryProviderParamsSetDevice(level_zero_params, device); |
| 747 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 748 | + fprintf(stderr, |
| 749 | + "error: umfLevelZeroMemoryProviderParamsSetDevice() failed\n"); |
| 750 | + goto err_destroy_params; |
| 751 | + } |
| 752 | + |
| 753 | + umf_result = umfLevelZeroMemoryProviderParamsSetMemoryType( |
| 754 | + level_zero_params, UMF_MEMORY_TYPE_DEVICE); |
| 755 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 756 | + fprintf( |
| 757 | + stderr, |
| 758 | + "error: umfLevelZeroMemoryProviderParamsSetMemoryType() failed\n"); |
| 759 | + goto err_destroy_params; |
| 760 | + } |
| 761 | + |
| 762 | + alloc_t *allocs = alloc_array(N_BUFFERS); |
| 763 | + if (allocs == NULL) { |
| 764 | + fprintf(stderr, "error: alloc_array() failed\n"); |
| 765 | + goto err_destroy_context; |
| 766 | + } |
| 767 | + |
| 768 | + umf_memory_provider_handle_t provider = NULL; |
| 769 | + umf_result = umfMemoryProviderCreate(umfLevelZeroMemoryProviderOps(), |
| 770 | + level_zero_params, &provider); |
| 771 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 772 | + fprintf(stderr, "error: umfMemoryProviderCreate() failed\n"); |
| 773 | + goto err_free_allocs; |
| 774 | + } |
| 775 | + |
| 776 | + umf_disjoint_pool_params_handle_t disjoint_params = NULL; |
| 777 | + umf_result = umfDisjointPoolParamsCreate(&disjoint_params); |
| 778 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 779 | + fprintf(stderr, "ERROR: umfDisjointPoolParamsCreate failed\n"); |
| 780 | + goto err_provider_destroy; |
| 781 | + } |
| 782 | + |
| 783 | + umf_result = |
| 784 | + umfDisjointPoolParamsSetSlabMinSize(disjoint_params, BUFFER_SIZE * 10); |
| 785 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 786 | + fprintf(stderr, |
| 787 | + "error: umfDisjointPoolParamsSetSlabMinSize() failed\n"); |
| 788 | + goto err_params_destroy; |
| 789 | + } |
| 790 | + |
| 791 | + umf_result = umfDisjointPoolParamsSetMaxPoolableSize( |
| 792 | + disjoint_params, 4ull * 1024ull * 1024ull); |
| 793 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 794 | + fprintf(stderr, |
| 795 | + "error: umfDisjointPoolParamsSetMaxPoolableSize() failed\n"); |
| 796 | + goto err_params_destroy; |
| 797 | + } |
| 798 | + |
| 799 | + umf_result = |
| 800 | + umfDisjointPoolParamsSetCapacity(disjoint_params, 64ull * 1024ull); |
| 801 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 802 | + fprintf(stderr, "error: umfDisjointPoolParamsSetCapacity() failed\n"); |
| 803 | + goto err_params_destroy; |
| 804 | + } |
| 805 | + |
| 806 | + umf_result = umfDisjointPoolParamsSetMinBucketSize(disjoint_params, 64); |
| 807 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 808 | + fprintf(stderr, |
| 809 | + "error: umfDisjointPoolParamsSetMinBucketSize() failed\n"); |
| 810 | + goto err_params_destroy; |
| 811 | + } |
| 812 | + |
| 813 | + umf_pool_create_flags_t flags = UMF_POOL_CREATE_FLAG_NONE; |
| 814 | + umf_memory_pool_handle_t pool; |
| 815 | + umf_result = umfPoolCreate(umfDisjointPoolOps(), provider, disjoint_params, |
| 816 | + flags, &pool); |
| 817 | + if (umf_result != UMF_RESULT_SUCCESS) { |
| 818 | + fprintf(stderr, "error: umfPoolCreate() failed\n"); |
| 819 | + goto err_params_destroy; |
| 820 | + } |
| 821 | + |
| 822 | + for (size_t i = 0; i < N_BUFFERS; ++i) { |
| 823 | + allocs[i].ptr = umfPoolMalloc(pool, BUFFER_SIZE); |
| 824 | + if (allocs[i].ptr == NULL) { |
| 825 | + goto err_buffer_destroy; |
| 826 | + } |
| 827 | + allocs[i].size = BUFFER_SIZE; |
| 828 | + } |
| 829 | + |
| 830 | + do_umf_mem_props_benchmark(context, true, allocs, N_BUFFERS, |
| 831 | + 1); // WARMUP |
| 832 | + UBENCH_DO_BENCHMARK() { |
| 833 | + do_umf_mem_props_benchmark(context, true, allocs, N_BUFFERS, |
| 834 | + N_ITERATIONS); |
| 835 | + } |
| 836 | + |
| 837 | +err_buffer_destroy: |
| 838 | + for (size_t i = 0; i < N_BUFFERS; ++i) { |
| 839 | + umfPoolFree(pool, allocs[i].ptr); |
| 840 | + } |
| 841 | + |
| 842 | + umfPoolDestroy(pool); |
| 843 | + |
| 844 | +err_params_destroy: |
| 845 | + umfDisjointPoolParamsDestroy(disjoint_params); |
| 846 | + |
| 847 | +err_provider_destroy: |
| 848 | + umfMemoryProviderDestroy(provider); |
| 849 | + |
| 850 | +err_free_allocs: |
| 851 | + free(allocs); |
| 852 | + |
| 853 | +err_destroy_params: |
| 854 | + umfLevelZeroMemoryProviderParamsDestroy(level_zero_params); |
| 855 | + |
| 856 | +err_destroy_context: |
| 857 | + //utils_ze_destroy_context(context); |
| 858 | +} |
| 859 | + |
626 | 860 | #endif /* (defined UMF_BUILD_LEVEL_ZERO_PROVIDER && defined UMF_BUILD_GPU_TESTS) */ |
627 | 861 |
|
628 | 862 | // TODO add IPC benchmark for CUDA |
|
0 commit comments