[PATCH] benchmarks: Measure BLT performance

Gurram, Pravalika pravalika.gurram at intel.com
Mon Apr 28 10:40:31 UTC 2025



> -----Original Message-----
> From: Ch, Sai Gowtham <sai.gowtham.ch at intel.com>
> Sent: Friday, April 25, 2025 1:33 PM
> To: Gurram, Pravalika <pravalika.gurram at intel.com>; igt-
> dev at lists.freedesktop.org
> Cc: Ramadeva, Dwarakanath <dwarakanath.ramadeva at intel.com>
> Subject: RE: [PATCH] benchmarks: Measure BLT performance
> 
> Hi Pravalika,
> 
> >-----Original Message-----
> >From: Gurram, Pravalika <pravalika.gurram at intel.com>
> >Sent: Thursday, April 24, 2025 2:40 PM
> >To: igt-dev at lists.freedesktop.org
> >Cc: Ch, Sai Gowtham <sai.gowtham.ch at intel.com>; Ramadeva, Dwarakanath
> ><dwarakanath.ramadeva at intel.com>; Gurram, Pravalika
> ><pravalika.gurram at intel.com>
> >Subject: [PATCH] benchmarks: Measure BLT performance
> >
> >Execute N blits and time how long they complete to measure both GPU
> >limited bandwidth and submission overhead.
> >
> >Signed-off-by: Pravalika Gurram <pravalika.gurram at intel.com>
> >---
> > benchmarks/meson.build |   1 +
> > benchmarks/xe_blt.c    | 330
> +++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 331 insertions(+)
> > create mode 100644 benchmarks/xe_blt.c
> >
> >diff --git a/benchmarks/meson.build b/benchmarks/meson.build index
> >2c9a88fd3..4421ede86 100644
> >--- a/benchmarks/meson.build
> >+++ b/benchmarks/meson.build
> >@@ -21,6 +21,7 @@ benchmark_progs = [
> > 	'kms_vblank',
> > 	'prime_lookup',
> > 	'vgem_mmap',
> >+        'xe_blt',
> > ]
> >
> > benchmarksdir = join_paths(libexecdir, 'benchmarks') diff --git
> >a/benchmarks/xe_blt.c b/benchmarks/xe_blt.c new file mode 100644 index
> >000000000..e95ed6d47
> >--- /dev/null
> >+++ b/benchmarks/xe_blt.c
> >@@ -0,0 +1,330 @@
> >+// SPDX-License-Identifier: MIT
> >+/*
> >+ * Copyright © 2025 Intel Corporation
> >+ */
> >+
> >+#include "drm.h"
> >+#include "igt_syncobj.h"
> >+#include "intel_blt.h"
> >+#include "xe/xe_ioctl.h"
> >+#include "xe/xe_query.h"
> >+#include "xe/xe_util.h"
> >+
> >+
> >+#define COPY_BLT_CMD		(2<<29|0x53<<22|0x6)
> >+#define BLT_WRITE_ALPHA		(1<<21)
> >+#define BLT_WRITE_RGB		(1<<20)
> >+#define BLT_SRC_TILED		(1<<15)
> >+#define BLT_DST_TILED		(1<<11)
> >+
> >+static double
> >+elapsed(const struct timespec *start, const struct timespec *end) {
> >+	return (end->tv_sec - start->tv_sec) + 1e-9*(end->tv_nsec -
> Looks like it's not properly aligned, do checkpatch.
> >+start->tv_nsec); }
> >+
> >+static uint64_t emit_blt_src_copy(int fd,
> >+		uint64_t ahnd,
> Allocator handle can be removed from here.
> >+		const struct blt_copy_data *blt,
> >+		uint64_t bb_pos,
> >+		bool emit_bbe,
> >+		uint64_t dst_offset,
> >+		uint64_t src_offset,
> >+		uint32_t height)
> >+{
> >+	uint32_t b[12];
> >+	uint32_t bbe = MI_BATCH_BUFFER_END;
> >+	uint32_t *bb;
> >+	int i = 0;
> >+
> >+	src_offset += blt->src.plane_offset;
> >+	dst_offset += blt->dst.plane_offset;
> >+
> >+	b[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB;
> It's very fascinating to understand why not use XY_SRC_COPY_BLT_CMD instead
> of COPY_BLT_CMD, As per my understanding this benchmark does basic copy
> operations i think COPY_BLT_CMD will be much faster though, However do we
> have any further plans to include in the future XY_SRC_COPY_BLT_CMD ?


Both the instructions are same but in COPT_BTL_CMD  we are using the platform info that is the reason did not use XY_SRC_COPY_BLT_CMD
No in future am not going to change 
> >+	b[i-1] += 2;
> >+	b[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | (16*1024);
> >+	b[i++] = 0;
> >+	b[i++] = height << 16 | (4*1024);
> >+	b[i++] = dst_offset;
> >+	b[i++] = dst_offset >> 32; /* FIXME */
> >+	b[i++] = 0;
> >+	b[i++] = 16*1024;
> >+	b[i++] = src_offset;
> >+	b[i++] = src_offset >> 32; /* FIXME */
> >+
> >+	bb = xe_bo_map(fd, blt->bb.handle, blt->bb.size);
> >+
> >+	igt_assert(bb_pos + sizeof(b) < blt->bb.size);
> >+	memcpy(bb + bb_pos, &b, sizeof(b));
> >+	bb_pos += sizeof(b);
> >+
> >+	if (emit_bbe) {
> >+		igt_assert(bb_pos + sizeof(uint32_t) < blt->bb.size);
> >+		memcpy(bb + bb_pos, &bbe, sizeof(bbe));
> >+		bb_pos += sizeof(uint32_t);
> >+	}
> >+
> >+	munmap(bb, blt->bb.size);
> >+
> >+	return bb_pos;
> >+
> >+}
> >+
> >+static int count;
> >+/*
> >+ * val = True To check BB is working fine or not
> >+ * count_val = To get the counter value how many buffers we can send
> >+in
> >+0.1 sec
> >+ * loop_count = count_val time we have to submit the buffers  */
> >+static int blt_src_copy(int fd,
> >+		const intel_ctx_t *ctx,
> >+		const struct intel_execution_engine2 *e,
> >+		uint64_t ahnd,
> >+		const struct blt_copy_data *blt, uint32_t height,
> >+		bool val, bool count_val, bool loop_count) {
> >+	uint64_t dst_offset = 0, src_offset = 0, bb_offset = 0;
> >+	int ret = 0;
> >+	uint64_t bb_pos = 0;
> >+	struct timespec start, end;
> >+
> >+	igt_assert_f(ahnd, "src-copy supports softpin only\n");
> >+	igt_assert_f(blt, "src-copy requires data to do src-copy blit\n");
> >+	igt_assert_neq(blt->driver, 0);
> >+
> >+	if (!val) {
> >+		src_offset = get_offset_pat_index(ahnd, blt->src.handle, blt-
> >>src.size,
> >+				0, blt->src.pat_index);
> >+		dst_offset = get_offset_pat_index(ahnd, blt->dst.handle, blt-
> >>dst.size,
> >+				0, blt->dst.pat_index);
> >+		bb_offset = get_offset(ahnd, blt->bb.handle, blt->bb.size, 0);
> >+	}
> >+
> >+	bb_pos = emit_blt_src_copy(fd, ahnd, blt, 0, true, src_offset,
> Same here ahnd can be removed.
> >+				  dst_offset, height);
> >+	if (count_val) {
> >+		clock_gettime(CLOCK_MONOTONIC, &start);
> >+		do {
> >+			if (blt->driver == INTEL_DRIVER_XE)
> >+				intel_ctx_xe_exec(ctx, ahnd,
> >CANONICAL(bb_offset));
> >+			count++;
> >+			clock_gettime(CLOCK_MONOTONIC, &end);
> >+			if (elapsed(&start, &end) > (100 / 1000.))
> >+				break;
> >+		} while (1);
> >+	} else if (loop_count) {
> >+		for (int loop = 0; loop < count; loop++) {
> >+			if (blt->driver == INTEL_DRIVER_XE)
> >+				intel_ctx_xe_exec(ctx, ahnd,
> >CANONICAL(bb_offset));
> >+		}
> >+
> >+	} else {
> >+		if (blt->driver == INTEL_DRIVER_XE)
> >+			intel_ctx_xe_exec(ctx, ahnd, CANONICAL(bb_offset));
> >+	}
> >+	return ret;
> >+}
> >+
> >+static int src_copy(int xe, const intel_ctx_t *ctx,
> >+		uint32_t width, uint32_t height,
> >+		uint32_t region1, uint32_t region2, bool val, bool count, bool
> loop)
> >+{
> >+	struct blt_copy_data blt = {};
> >+	struct blt_copy_object *src, *dst;
> >+	const uint32_t bpp = 32;
> >+	uint64_t bb_size = xe_bb_size(xe, SZ_4K);
> >+	uint64_t ahnd = intel_allocator_open_full(xe, ctx->vm, 0, 0,
> >+			INTEL_ALLOCATOR_SIMPLE,
> >+			ALLOC_STRATEGY_LOW_TO_HIGH, 0);
> >+	uint32_t bb;
> >+	int ret = 0;
> >+
> >+	bb = xe_bo_create(xe, 0, bb_size, region1, 0);
> >+	blt_copy_init(xe, &blt);
> >+	src = blt_create_object(&blt, region1, width, height, bpp, 0,
> >+			T_LINEAR, 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_set_copy_object(&blt.src, src);
> >+	blt_set_copy_object(&blt.dst, dst);
> >+	blt_set_batch(&blt.bb, bb, bb_size, region1);
> >+
> >+	ret = blt_src_copy(xe, ctx, NULL, ahnd, &blt, height, val, count,
> >+loop);
> >+
> >+	put_offset(ahnd, src->handle);
> >+	put_offset(ahnd, dst->handle);
> >+	put_offset(ahnd, bb);
> >+	intel_allocator_bind(ahnd, 0, 0);
> >+	blt_destroy_object(xe, src);
> >+	blt_destroy_object(xe, dst);
> >+	gem_close(xe, bb);
> >+	put_ahnd(ahnd);
> >+	return ret;
> >+
> >+}
> >+
> >+#define SYNC 0x1
> >+
> >+static int run(int width, int batch, int time, int reps, int ncpus,
> >+unsigned int flags) {
> >+	struct igt_collection *set;
> >+	int xe;
> >+	int height = width / (16 * 1024);
> >+	struct drm_xe_engine_class_instance inst = {
> >+		.engine_class = DRM_XE_ENGINE_CLASS_COPY,
> >+	};
> >+	intel_ctx_t *ctx;
> >+	double *shared;
> >+	uint32_t region1, region2;
> >+	uint32_t vm, exec_queue;
> >+	int ret;
> >+
> >+	shared = mmap(0, 4096, PROT_WRITE, MAP_SHARED | MAP_ANON, -1,
> >0);
> >+
> >+	xe = drm_open_driver(DRIVER_XE);
> >+	xe_device_get(xe);
> >+	set = xe_get_memory_region_set(xe,
> >+			DRM_XE_MEM_REGION_CLASS_SYSMEM,
> >+			DRM_XE_MEM_REGION_CLASS_VRAM);
> >+	intel_allocator_multiprocess_start();
> >+
> >+	region1 = 1;
> >+	region2 = 2;
> >+
> >+	vm = xe_vm_create(xe, 0, 0);
> >+	exec_queue = xe_exec_queue_create(xe, vm, &inst, 0);
> >+	ctx = intel_ctx_xe(xe, vm, exec_queue, 0, 0, 0);
> >+	ret = src_copy(xe, ctx, width, height, region1, region2, true,
> >+								false, false);
> >+	xe_exec_queue_destroy(xe, exec_queue);
> >+	xe_vm_destroy(xe, vm);
> >+	free(ctx);
> >+	if (!ret) {
> >+		vm = xe_vm_create(xe, 0, 0);
> >+		exec_queue = xe_exec_queue_create(xe, vm, &inst, 0);
> >+		ctx = intel_ctx_xe(xe, vm, exec_queue, 0, 0, 0);
> >+		ret = src_copy(xe, ctx, width, height, region1, region2,
> >+							false, false, false);
> >+		xe_exec_queue_destroy(xe, exec_queue);
> >+		xe_vm_destroy(xe, vm);
> >+		free(ctx);
> >+	}
> >+	if (batch > 1) {
> >+		for (int i = 1; i < batch; i++) {
> >+			vm = xe_vm_create(xe, 0, 0);
> >+			exec_queue = xe_exec_queue_create(xe, vm, &inst, 0);
> >+			ctx = intel_ctx_xe(xe, vm, exec_queue, 0, 0, 0);
> >+			ret = src_copy(xe, ctx, width, height, region1, region2,
> >+							false, false, false);
> >+			xe_exec_queue_destroy(xe, exec_queue);
> >+			xe_vm_destroy(xe, vm);
> >+			free(ctx);
> >+		}
> >+	}
> >+	vm = xe_vm_create(xe, 0, 0);
> >+	exec_queue = xe_exec_queue_create(xe, vm, &inst, 0);
> >+	ctx = intel_ctx_xe(xe, vm, exec_queue, 0, 0, 0);
> >+	ret = src_copy(xe, ctx, width, height, region1, region2,
> >+						false, true, false);
> >+	xe_exec_queue_destroy(xe, exec_queue);
> >+	xe_vm_destroy(xe, vm);
> Can't we use same vm_id's and exec_queue_id's for all the different conditions ?
> why to create and destroy multiple times ?
> Idea behind this benchmark looks good.
> 
> ---
> Gowtham
> >+	free(ctx);
> >+	if (flags & SYNC) {
> >+		time *= count / 2;
> >+		count = 1;
> >+	}
> >+
> >+	while (reps--) {
> >+		memset(shared, 0, 4096);
> >+
> >+		igt_fork(child, ncpus) {
> >+			double min = HUGE_VAL;
> >+
> >+			for (int s = 0; s <= time / 100; s++) {
> >+				struct timespec start, end;
> >+				double t;
> >+
> >+				clock_gettime(CLOCK_MONOTONIC, &start);
> >+				vm = xe_vm_create(xe, 0, 0);
> >+				exec_queue = xe_exec_queue_create(xe, vm,
> >&inst, 0);
> >+				ctx = intel_ctx_xe(xe, vm, exec_queue, 0, 0, 0);
> >+				ret = src_copy(xe, ctx, width, height, region1,
> >+						region2, false, false, true);
> >+				free(ctx);
> >+				xe_exec_queue_destroy(xe, exec_queue);
> >+				xe_vm_destroy(xe, vm);
> >+				clock_gettime(CLOCK_MONOTONIC, &end);
> >+
> >+				t = elapsed(&start, &end);
> >+				if (t < min)
> >+					min = t;
> >+			}
> >+
> >+			shared[child] = width/(1024*1024.)*batch*count/min;
> >+		}
> >+		igt_waitchildren();
> >+
> >+		for (int child = 0; child < ncpus; child++)
> >+			shared[ncpus] += shared[child];
> >+		printf("%7.3f\n", shared[ncpus] / ncpus);
> >+	}
> >+	intel_allocator_multiprocess_stop();
> >+
> >+	close(xe);
> >+	return 0;
> >+}
> >+
> >+int main(int argc, char **argv)
> >+{
> >+	int size = 1024*1024;
> >+	int reps = 13;
> >+	int time = 2000;
> >+	int ncpus = 1;
> >+	int batch = 1;
> >+	unsigned int flags = 0;
> >+	int c;
> >+
> >+	while ((c = getopt(argc, argv, "s:S:t:r:b:f")) != -1) {
> >+		switch (c) {
> >+		case 's':
> >+			size = atoi(optarg);
> >+			size = ALIGN(size, 4);
> >+			if (size < 4)
> >+				size = 4;
> >+			break;
> >+
> >+		case 'S':
> >+			flags |= SYNC;
> >+			break;
> >+
> >+		case 't':
> >+			time = atoi(optarg);
> >+			if (time < 1)
> >+				time = 1;
> >+			break;
> >+
> >+		case 'r':
> >+			reps = atoi(optarg);
> >+			if (reps < 1)
> >+				reps = 1;
> >+			break;
> >+
> >+		case 'b':
> >+			batch = atoi(optarg);
> >+			if (batch < 1)
> >+				batch = 1;
> >+			break;
> >+
> >+		case 'f':
> >+			ncpus = sysconf(_SC_NPROCESSORS_ONLN);
> >+			break;
> >+
> >+		default:
> >+			break;
> >+		}
> >+	}
> >+
> >+	return run(size, batch, time, reps, ncpus, flags); }
> >--
> >2.34.1



More information about the igt-dev mailing list