Re: [igt-dev] [PATCH i-g-t v2 15/16] tests/xe_exercise_blt: Check blitter library fast-copy for Xe

[Karolina Stolarek <karolina.stolarek@intel.com>]

On 11.07.2023 13:07, Zbigniew Kempczyński wrote:
> On Fri, Jul 07, 2023 at 01:10:52PM +0200, Karolina Stolarek wrote:
>> On 6.07.2023 08:05, Zbigniew Kempczyński wrote:
>>> Port this test to work on xe. Instead of adding conditional code for
>>> xe code which would decrease readability this is new test for xe.
>>>
>>> Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
>>> ---
>>>    tests/meson.build          |   1 +
>>>    tests/xe/xe_exercise_blt.c | 372 +++++++++++++++++++++++++++++++++++++
>>>    2 files changed, 373 insertions(+)
>>>    create mode 100644 tests/xe/xe_exercise_blt.c
>>>
>>> diff --git a/tests/meson.build b/tests/meson.build
>>> index 9bca57a5e8..137a5cf01f 100644
>>> --- a/tests/meson.build
>>> +++ b/tests/meson.build
>>> @@ -257,6 +257,7 @@ xe_progs = [
>>>    	'xe_exec_reset',
>>>    	'xe_exec_store',
>>>    	'xe_exec_threads',
>>> +	'xe_exercise_blt',
>>>    	'xe_gpgpu_fill',
>>>    	'xe_guc_pc',
>>>    	'xe_huc_copy',
>>> diff --git a/tests/xe/xe_exercise_blt.c b/tests/xe/xe_exercise_blt.c
>>> new file mode 100644
>>> index 0000000000..8340cf7148
>>> --- /dev/null
>>> +++ b/tests/xe/xe_exercise_blt.c
>>> @@ -0,0 +1,372 @@
>>> +// SPDX-License-Identifier: MIT
>>> +/*
>>> + * Copyright © 2023 Intel Corporation
>>> + */
>>> +
>>> +#include "igt.h"
>>> +#include "drm.h"
>>> +#include "lib/intel_chipset.h"
>>> +#include "intel_blt.h"
>>> +#include "intel_mocs.h"
>>> +#include "xe/xe_ioctl.h"
>>> +#include "xe/xe_query.h"
>>> +#include "xe/xe_util.h"
>>> +
>>> +/**
>>> + * TEST: xe exercise blt
>>> + * Description: Exercise blitter commands on Xe
>>> + * Feature: blitter
>>> + * Run type: FULL
>>> + * Test category: GEM_Legacy
>>> + *
>>> + * SUBTEST: fast-copy
>>> + * Description:
>>> + *   Check fast-copy blit
>>> + *   blitter
>>> + *
>>> + * SUBTEST: fast-copy-emit
>>> + * Description:
>>> + *   Check multiple fast-copy in one batch
>>> + *   blitter
>>> + */
>>> +
>>> +IGT_TEST_DESCRIPTION("Exercise blitter commands on Xe");
>>> +
>>> +static struct param {
>>> +	int tiling;
>>> +	bool write_png;
>>> +	bool print_bb;
>>> +	bool print_surface_info;
>>> +	int width;
>>> +	int height;
>>> +} param = {
>>> +	.tiling = -1,
>>> +	.write_png = false,
>>> +	.print_bb = false,
>>> +	.print_surface_info = false,
>>> +	.width = 512,
>>> +	.height = 512,
>>> +};
>>> +
>>> +#define PRINT_SURFACE_INFO(name, obj) do { \
>>> +	if (param.print_surface_info) \
>>> +		blt_surface_info((name), (obj)); } while (0)
>>> +
>>> +#define WRITE_PNG(fd, id, name, obj, w, h) do { \
>>> +	if (param.write_png) \
>>> +		blt_surface_to_png((fd), (id), (name), (obj), (w), (h)); } while (0)
>>> +
>>
>> My suggestion with shared functions applies here as well
>>
>>> +struct blt_fast_copy_data {
>>> +	int xe;
>>> +	struct blt_copy_object src;
>>> +	struct blt_copy_object mid;
>>> +	struct blt_copy_object dst;
>>> +
>>> +	struct blt_copy_batch bb;
>>> +	enum blt_color_depth color_depth;
>>> +
>>> +	/* debug stuff */
>>> +	bool print_bb;
>>> +};
>>> +
>>> +static int fast_copy_one_bb(int xe,
>>> +			    const intel_ctx_t *ctx,
>>> +			    uint64_t ahnd,
>>> +			    const struct blt_fast_copy_data *blt)
>>> +{
>>> +	struct blt_copy_data blt_tmp;
>>> +	uint64_t bb_offset, alignment;
>>> +	uint64_t bb_pos = 0;
>>> +	int ret = 0;
>>> +
>>> +	alignment = xe_get_default_alignment(xe);
>>> +
>>> +	get_offset(ahnd, blt->src.handle, blt->src.size, alignment);
>>> +	get_offset(ahnd, blt->mid.handle, blt->mid.size, alignment);
>>> +	get_offset(ahnd, blt->dst.handle, blt->dst.size, alignment);
>>> +	bb_offset = get_offset(ahnd, blt->bb.handle, blt->bb.size, alignment);
>>> +
>>> +	/* First blit */
>>> +	blt_copy_init(xe, &blt_tmp);
>>> +	blt_tmp.src = blt->src;
>>> +	blt_tmp.dst = blt->mid;
>>> +	blt_tmp.bb = blt->bb;
>>> +	blt_tmp.color_depth = blt->color_depth;
>>> +	blt_tmp.print_bb = blt->print_bb;
>>> +	bb_pos = emit_blt_fast_copy(xe, ahnd, &blt_tmp, bb_pos, false);
>>> +
>>> +	/* Second blit */
>>> +	blt_copy_init(xe, &blt_tmp);
>>> +	blt_tmp.src = blt->mid;
>>> +	blt_tmp.dst = blt->dst;
>>> +	blt_tmp.bb = blt->bb;
>>> +	blt_tmp.color_depth = blt->color_depth;
>>> +	blt_tmp.print_bb = blt->print_bb;
>>> +	bb_pos = emit_blt_fast_copy(xe, ahnd, &blt_tmp, bb_pos, true);
>>> +
>>> +	intel_ctx_xe_exec(ctx, ahnd, bb_offset);
>>> +
>>> +	return ret;
>>> +}
>>> +
>>> +static void fast_copy_emit(int xe, const intel_ctx_t *ctx,
>>> +			   uint32_t region1, uint32_t region2,
>>> +			   enum blt_tiling_type mid_tiling)
>>> +{
>>> +	struct blt_copy_data bltinit = {};
>>> +	struct blt_fast_copy_data blt = {};
>>> +	struct blt_copy_object *src, *mid, *dst;
>>> +	const uint32_t bpp = 32;
>>> +	uint64_t bb_size = xe_get_default_alignment(xe);
>>> +	uint64_t ahnd = intel_allocator_open_full(xe, ctx->vm, 0, 0,
>>> +						  INTEL_ALLOCATOR_SIMPLE,
>>> +						  ALLOC_STRATEGY_LOW_TO_HIGH, 0);
>>> +	uint32_t bb, width = param.width, height = param.height;
>>> +	int result;
>>> +
>>> +	bb = xe_bo_create_flags(xe, 0, bb_size, region1);
>>> +
>>> +	blt_copy_init(xe, &bltinit);
>>> +	src = blt_create_object(&bltinit, region1, width, height, bpp, 0,
>>> +				T_LINEAR, COMPRESSION_DISABLED, 0, true);
>>> +	mid = blt_create_object(&bltinit, region2, width, height, bpp, 0,
>>> +				mid_tiling, COMPRESSION_DISABLED, 0, true);
>>> +	dst = blt_create_object(&bltinit, region1, width, height, bpp, 0,
>>> +				T_LINEAR, COMPRESSION_DISABLED, 0, true);
>>> +	igt_assert(src->size == dst->size);
>>> +
>>> +	PRINT_SURFACE_INFO("src", src);
>>> +	PRINT_SURFACE_INFO("mid", mid);
>>> +	PRINT_SURFACE_INFO("dst", dst);
>>> +
>>> +	blt_surface_fill_rect(xe, src, width, height);
>>> +	WRITE_PNG(xe, mid_tiling, "src", src, width, height);
>>> +
>>> +	memset(&blt, 0, sizeof(blt));
>>> +	blt.color_depth = CD_32bit;
>>> +	blt.print_bb = param.print_bb;
>>> +	blt_set_copy_object(&blt.src, src);
>>> +	blt_set_copy_object(&blt.mid, mid);
>>> +	blt_set_copy_object(&blt.dst, dst);
>>> +	blt_set_batch(&blt.bb, bb, bb_size, region1);
>>> +
>>> +	fast_copy_one_bb(xe, ctx, ahnd, &blt);
>>> +
>>> +	WRITE_PNG(xe, mid_tiling, "mid", &blt.mid, width, height);
>>> +	WRITE_PNG(xe, mid_tiling, "dst", &blt.dst, width, height);
>>> +
>>> +	result = memcmp(src->ptr, blt.dst.ptr, src->size);
>>> +
>>> +	blt_destroy_object(xe, src);
>>> +	blt_destroy_object(xe, mid);
>>> +	blt_destroy_object(xe, dst);
>>
>> I see that in block_copy tests we also call put_offset() for all copy
>> objects' handles. Although we have a different allocator, I think that we
>> should free the ranges on object destruction.
>>
> 
> I was a little bit surprised we put_offset() in fast-copy and surf-copy.
> I mean this will rebind offsets unnecessary. Finally we should put them
> before releasing ahnd, but for vm-bind (xe) destroying vm just does
> the job (and intel_allocator_init() is freeing the dangling memory
> which wasn't released explicitly).

I see, good to know! Thanks for the explaination.

All the best,
Karolina

> 
>>> +	gem_close(xe, bb);
>>> +	put_ahnd(ahnd);
>>> +
>>> +	munmap(&bb, bb_size);
>>> +
>>> +	igt_assert_f(!result, "source and destination surfaces differs!\n");
>>> +}
>>> +
>>> +static void fast_copy(int xe, const intel_ctx_t *ctx,
>>> +		      uint32_t region1, uint32_t region2,
>>> +		      enum blt_tiling_type mid_tiling)
>>> +{
>>> +	struct blt_copy_data blt = {};
>>> +	struct blt_copy_object *src, *mid, *dst;
>>> +	const uint32_t bpp = 32;
>>> +	uint64_t bb_size = xe_get_default_alignment(xe);
>>> +	uint64_t ahnd = intel_allocator_open_full(xe, ctx->vm, 0, 0,
>>> +						  INTEL_ALLOCATOR_SIMPLE,
>>> +						  ALLOC_STRATEGY_LOW_TO_HIGH, 0);
>>> +	uint32_t bb;
>>> +	uint32_t width = param.width, height = param.height;
>>> +	int result;
>>> +
>>> +	bb = xe_bo_create_flags(xe, 0, bb_size, region1);
>>> +
>>> +	blt_copy_init(xe, &blt);
>>> +	src = blt_create_object(&blt, region1, width, height, bpp, 0,
>>> +				T_LINEAR, COMPRESSION_DISABLED, 0, true);
>>> +	mid = blt_create_object(&blt, region2, width, height, bpp, 0,
>>> +				mid_tiling, COMPRESSION_DISABLED, 0, true);
>>> +	dst = blt_create_object(&blt, region1, width, height, bpp, 0,
>>> +				T_LINEAR, COMPRESSION_DISABLED, 0, true);
>>> +	igt_assert(src->size == dst->size);
>>> +
>>> +	blt_surface_fill_rect(xe, src, width, height);
>>> +
>>> +	blt.color_depth = CD_32bit;
>>> +	blt.print_bb = param.print_bb;
>>> +	blt_set_copy_object(&blt.src, src);
>>> +	blt_set_copy_object(&blt.dst, mid);
>>> +	blt_set_batch(&blt.bb, bb, bb_size, region1);
>>> +
>>> +	blt_fast_copy(xe, ctx, NULL, ahnd, &blt);
>>> +
>>> +	WRITE_PNG(xe, mid_tiling, "src", &blt.src, width, height);
>>> +	WRITE_PNG(xe, mid_tiling, "mid", &blt.dst, width, height);
>>> +
>>> +	blt_copy_init(xe, &blt);
>>> +	blt.color_depth = CD_32bit;
>>> +	blt.print_bb = param.print_bb;
>>> +	blt_set_copy_object(&blt.src, mid);
>>> +	blt_set_copy_object(&blt.dst, dst);
>>> +	blt_set_batch(&blt.bb, bb, bb_size, region1);
>>> +
>>> +	blt_fast_copy(xe, ctx, NULL, ahnd, &blt);
>>> +
>>> +	WRITE_PNG(xe, mid_tiling, "dst", &blt.dst, width, height);
>>> +
>>> +	result = memcmp(src->ptr, blt.dst.ptr, src->size);
>>> +
>>> +	blt_destroy_object(xe, src);
>>> +	blt_destroy_object(xe, mid);
>>> +	blt_destroy_object(xe, dst);
>>> +	gem_close(xe, bb);
>>> +	put_ahnd(ahnd);
>>> +
>>> +	igt_assert_f(!result, "source and destination surfaces differs!\n");
>>> +}
>>> +
>>> +enum fast_copy_func {
>>> +	FAST_COPY,
>>> +	FAST_COPY_EMIT
>>> +};
>>> +
>>> +static char
>>> +	*full_subtest_str(char *regtxt, enum blt_tiling_type tiling,
>>> +			  enum fast_copy_func func)
>>> +{
>>> +	char *name;
>>> +	uint32_t len;
>>> +
>>> +	len = asprintf(&name, "%s-%s%s", blt_tiling_name(tiling), regtxt,
>>> +		       func == FAST_COPY_EMIT ? "-emit" : "");
>>> +
>>> +	igt_assert_f(len >= 0, "asprintf failed!\n");
>>> +
>>> +	return name;
>>> +}
>>> +
>>> +static void fast_copy_test(int xe,
>>> +			   struct igt_collection *set,
>>> +			   enum fast_copy_func func)
>>> +{
>>> +	struct drm_xe_engine_class_instance inst = {
>>> +		.engine_class = DRM_XE_ENGINE_CLASS_COPY,
>>> +	};
>>> +	struct igt_collection *regions;
>>> +	void (*copy_func)(int xe, const intel_ctx_t *ctx,
>>> +			  uint32_t r1, uint32_t r2, enum blt_tiling_type tiling);
>>> +	intel_ctx_t *ctx;
>>> +	int tiling;
>>> +
>>> +	for_each_tiling(tiling) {
>>> +		if (!blt_fast_copy_supports_tiling(xe, tiling))
>>> +			continue;
>>> +
>>> +		for_each_variation_r(regions, 2, set) {
>>> +			uint32_t region1, region2;
>>> +			uint32_t vm, engine;
>>> +			char *regtxt, *test_name;
>>> +
>>> +			region1 = igt_collection_get_value(regions, 0);
>>> +			region2 = igt_collection_get_value(regions, 1);
>>> +
>>> +			vm = xe_vm_create(xe, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
>>> +			engine = xe_engine_create(xe, vm, &inst, 0);
>>> +			ctx = intel_ctx_xe(xe, vm, engine, 0, 0, 0);
>>> +
>>> +			copy_func = (func == FAST_COPY) ? fast_copy : fast_copy_emit;
>>> +			regtxt = xe_memregion_dynamic_subtest_name(xe, regions);
>>> +			test_name = full_subtest_str(regtxt, tiling, func);
>>> +
>>> +			igt_dynamic_f("%s", test_name) {
>>> +				copy_func(xe, ctx,
>>> +					  region1, region2,
>>> +					  tiling);
>>> +			}
>>> +
>>> +			free(regtxt);
>>> +			free(test_name);
>>> +			xe_engine_destroy(xe, engine);
>>> +			xe_vm_destroy(xe, vm);
>>> +			free(ctx);
>>> +		}
>>> +	}
>>> +}
>>> +
>>> +static int opt_handler(int opt, int opt_index, void *data)
>>> +{
>>> +	switch (opt) {
>>> +	case 'b':
>>> +		param.print_bb = true;
>>> +		igt_debug("Print bb: %d\n", param.print_bb);
>>> +		break;
>>> +	case 'p':
>>> +		param.write_png = true;
>>> +		igt_debug("Write png: %d\n", param.write_png);
>>> +		break;
>>> +	case 's':
>>> +		param.print_surface_info = true;
>>> +		igt_debug("Print surface info: %d\n", param.print_surface_info);
>>> +		break;
>>> +	case 't':
>>> +		param.tiling = atoi(optarg);
>>> +		igt_debug("Tiling: %d\n", param.tiling);
>>> +		break;
>>> +	case 'W':
>>> +		param.width = atoi(optarg);
>>> +		igt_debug("Width: %d\n", param.width);
>>> +		break;
>>> +	case 'H':
>>> +		param.height = atoi(optarg);
>>> +		igt_debug("Height: %d\n", param.height);
>>> +		break;
>>> +	default:
>>> +		return IGT_OPT_HANDLER_ERROR;
>>> +	}
>>> +
>>> +	return IGT_OPT_HANDLER_SUCCESS;
>>> +}
>>> +
>>> +const char *help_str =
>>> +	"  -b\tPrint bb\n"
>>> +	"  -p\tWrite PNG\n"
>>> +	"  -s\tPrint surface info\n"
>>> +	"  -t\tTiling format (0 - linear, 1 - XMAJOR, 2 - YMAJOR, 3 - TILE4, 4 - TILE64, 5 - YFMAJOR)\n"
>>> +	"  -W\tWidth (default 512)\n"
>>> +	"  -H\tHeight (default 512)"
>>> +	;
>>> +
>>> +igt_main_args("b:pst:W:H:", NULL, help_str, opt_handler, NULL)
>>> +{
>>> +	struct igt_collection *set;
>>> +	int xe;
>>> +
>>> +	igt_fixture {
>>> +		xe = drm_open_driver(DRIVER_XE);
>>> +		igt_require(blt_has_block_copy(xe));
>>
>> Should be blt_has_fast_copy(xe)
> 
> Yes, missed that during copying fixture from xe_ccs. Thanks for spotting
> this.
> 
> --
> Zbigniew
> 
>>
>> All the best,
>> Karolina
>>> +
>>> +		xe_device_get(xe);
>>> +
>>> +		set = xe_get_memory_region_set(xe,
>>> +					       XE_MEM_REGION_CLASS_SYSMEM,
>>> +					       XE_MEM_REGION_CLASS_VRAM);
>>> +	}
>>> +
>>> +	igt_describe("Check fast-copy blit");
>>> +	igt_subtest_with_dynamic("fast-copy") {
>>> +		fast_copy_test(xe, set, FAST_COPY);
>>> +	}
>>> +
>>> +	igt_describe("Check multiple fast-copy in one batch");
>>> +	igt_subtest_with_dynamic("fast-copy-emit") {
>>> +		fast_copy_test(xe, set, FAST_COPY_EMIT);
>>> +	}
>>> +
>>> +	igt_fixture {
>>> +		drm_close_driver(xe);
>>> +	}
>>> +}