[igt-dev] [PATCH i-g-t 3/3] tests/gem_ccs: Verify uncompressed and compressed blits
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Thu Mar 3 13:11:08 UTC 2022
On Wed, Mar 02, 2022 at 11:43:09AM +0100, Kamil Konieczny wrote:
> Hi Zbigniew,
>
> Dnia 2022-02-25 at 12:06:26 +0100, Zbigniew Kempczyński napisał(a):
> > DG2 and above supports flat-ccs what means object compression lays in
> > dedicated part or memory and is directly correlated with object location
> > in physical memory. As no mapping is possible from CPU side to this area
> > dedicated blitter command (XY_CTRL_SURF_COPY_BLT) was created to copy
> > from/to ccs data.
> >
> > Test exercises scenarios:
> > 1. block-copy without compression (TGL/DG1)
> > 2. block-copy with flat-ccs compression (DG2+) + inplace decompression
> > 3. ctrl-surf-copy which verifies copying ccs data to/from flat-ccs area
>
> This is good starting point, what about one more, maybe for next
> patch edition, with check for compressed data, e.g. check if
> compression was really done ?
We can compare both src and mid after blit, it should differ. I'm not sure
but likely we would be able to write compression algorithm on cpu. But
I don't know is it worth to do this (at least now). If src->mid->dst
works fine and I assume mid is compressed algorithm is not much important,
as it is reversible.
>
> >
> > v2: set mip tail lod to 0xf to avoid overlap with lod
> >
> > Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> > ---
> > tests/i915/gem_ccs.c | 488 +++++++++++++++++++++++++++++++++++++++++++
> > tests/meson.build | 1 +
> > 2 files changed, 489 insertions(+)
> > create mode 100644 tests/i915/gem_ccs.c
> >
> > diff --git a/tests/i915/gem_ccs.c b/tests/i915/gem_ccs.c
> > new file mode 100644
> > index 000000000..9770d2840
> > --- /dev/null
> > +++ b/tests/i915/gem_ccs.c
> > @@ -0,0 +1,488 @@
> > +// SPDX-License-Identifier: MIT
> > +/*
> > + * Copyright © 2022 Intel Corporation
> > + */
> > +
> > +#include <errno.h>
> > +#include <sys/ioctl.h>
> > +#include <sys/time.h>
> > +#include <malloc.h>
> > +#include "drm.h"
> > +#include "igt.h"
> > +#include "i915/gem.h"
> > +#include "i915/gem_create.h"
> > +#include "lib/intel_chipset.h"
> > +#include "i915/i915_blt.h"
> > +
> > +IGT_TEST_DESCRIPTION("Exercise gen12 blitter with and without flat ccs");
>
> s/ccs/ccs compression/
>
> Please add "compression" word at the end of description, to
> make it different from compute ccs.
Flat ccs means flat compressed control surface, and definitely is not
the same as compute. CCS there means Compute Command Streamer.
>
> > +
> > +#define BPP 32
This is also very common and I think no BITS_PER_PIXEL defines is necessary,
all knows what that means.
>
> Please add description of BPP here, I guess it is BITS_PER_PIXEL, or maybe
> just use full name ?
>
> > +
> > +static struct param {
> > + int compression_format;
> > + int tiling;
> > + bool write_png;
> > + bool print_bb;
> > + bool print_surface_info;
> > + int width;
> > + int height;
> > + uint32_t increase_mid_surf_size;
> > +} param = {
> > + .compression_format = 0,
> > + .tiling = -1,
> > + .write_png = false,
> > + .print_bb = false,
> > + .print_surface_info = false,
> > + .width = 512,
> > + .height = 512,
> > + .increase_mid_surf_size = 0,
> > +};
> > +
> > +struct test_config {
> > + bool compression;
> > + bool inplace;
> > + bool surfcopy;
> > +};
> > +
> > +static void set_object(struct blt_copy_object *obj,
> > + uint32_t handle, uint64_t size, uint32_t region,
> > + uint8_t mocs, enum blt_tiling tiling,
> > + enum blt_compression compression,
> > + enum blt_compression_type compression_type)
> > +{
> > + obj->handle = handle;
> > + obj->size = size;
> > + obj->region = region;
> > + obj->mocs = mocs;
> > + obj->tiling = tiling;
> > + obj->compression = compression;
> > + obj->compression_type = compression_type;
> > +}
> > +
> > +static void set_geom(struct blt_copy_object *obj, uint32_t pitch,
> > + int16_t x1, int16_t y1, int16_t x2, int16_t y2,
> > + uint16_t x_offset, uint16_t y_offset)
> > +{
> > + obj->pitch = pitch;
> > + obj->x1 = x1;
> > + obj->y1 = y1;
> > + obj->x2 = x2;
> > + obj->y2 = y2;
> > + obj->x_offset = x_offset;
> > + obj->y_offset = y_offset;
> > +}
> > +
> > +static void set_batch(struct blt_copy_batch *batch,
> > + uint32_t handle, uint64_t size, uint32_t region)
> > +{
> > + batch->handle = handle;
> > + batch->size = size;
> > + batch->region = region;
> > +}
> > +
> > +static void set_object_ext(struct blt_block_copy_object_ext *obj,
> > + uint8_t compression_format,
> > + uint16_t surface_width, uint16_t surface_height,
> > + enum blt_surface_type surface_type)
> > +{
> > + obj->compression_format = compression_format;
> > + obj->surface_width = surface_width;
> > + obj->surface_height = surface_height;
> > + obj->surface_type = surface_type;
> > +
> > + /* Ensure mip tail won't overlap lod */
> > + obj->mip_tail_start_lod = 0xf;
> > +}
> > +
> > +static void set_surf_object(struct blt_ctrl_surf_copy_object *obj,
> > + uint32_t handle, uint32_t region, uint64_t size,
> > + uint8_t mocs, enum blt_access_type access_type)
> > +{
> > + obj->handle = handle;
> > + obj->region = region;
> > + obj->size = size;
> > + obj->mocs = mocs;
> > + obj->access_type = access_type;
> > +}
> > +
> > +static struct blt_copy_object *
> > +create_object(int i915, uint32_t region,
> > + uint32_t width, uint32_t height, uint32_t bpp, uint8_t mocs,
> > + enum blt_tiling tiling,
> > + enum blt_compression compression,
> > + enum blt_compression_type compression_type,
> > + bool create_mapping,
> > + uint64_t increase_size)
> > +{
> > + struct blt_copy_object *obj;
> > + uint64_t size = width * height * bpp / 8 + increase_size;
> > + uint32_t stride = tiling == T_LINEAR ? width * 4 : width;
> > + uint32_t handle;
> > +
> > + obj = calloc(1, sizeof(*obj));
> > +
> > + obj->size = size;
> > + igt_assert(__gem_create_in_memory_regions(i915, &handle,
> > + &size, region) == 0);
> > +
> > + set_object(obj, handle, size, region, mocs, tiling,
> > + compression, compression_type);
> > + set_geom(obj, stride, 0, 0, width, height, 0, 0);
> > +
> > + if (create_mapping)
> > + obj->ptr = gem_mmap__device_coherent(i915, handle, 0, size,
> > + PROT_READ | PROT_WRITE);
> > +
> > + return obj;
> > +}
> > +
> > +static void destroy_object(int i915, struct blt_copy_object *obj)
> > +{
> > + if (obj->ptr)
> > + munmap(obj->ptr, obj->size);
> > +
> > + gem_close(i915, obj->handle);
> > +}
> > +
> > +static void set_blt_object(struct blt_copy_object *obj,
> > + const struct blt_copy_object *orig)
> > +{
> > + memcpy(obj, orig, sizeof(*obj));
> > +}
> > +
> > +#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)
> > +
> > +static void surf_copy(int i915,
> > + const intel_ctx_t *ctx,
> > + const struct intel_execution_engine2 *e,
> > + uint64_t ahnd,
> > + const struct blt_copy_object *src,
> > + const struct blt_copy_object *mid,
> > + const struct blt_copy_object *dst,
> > + int run_id)
> > +{
> > + struct blt_copy_data blt = {};
> > + struct blt_block_copy_data_ext ext = {};
> > + struct blt_ctrl_surf_copy_data surf = {};
> > + uint32_t bb, bb2, ccs, *ccsmap, bb_size = 4096;
> > + uint64_t ccssize = mid->size / CCS_RATIO;
> > + uint32_t *ccscopy;
> > + int result;
> > +
> > + igt_assert(mid->compression);
> > + ccscopy = (uint32_t *) malloc(ccssize);
> > + bb = gem_create(i915, bb_size);
> > + bb2 = gem_create(i915, bb_size);
> > + ccs = gem_create(i915, ccssize);
> > +
> > + surf.i915 = i915;
> > + surf.print_bb = param.print_bb;
> > + set_surf_object(&surf.src, mid->handle, mid->region, mid->size,
> > + 0, INDIRECT_ACCESS);
> > + set_surf_object(&surf.dst, ccs, REGION_SMEM, ccssize,
> > + 0, DIRECT_ACCESS);
> > + set_batch(&surf.bb, bb, bb_size, REGION_SMEM);
> > + blt_ctrl_surf_copy(i915, ctx, e, ahnd, &surf);
> > + gem_sync(i915, surf.dst.handle);
> > +
> > + ccsmap = gem_mmap__device_coherent(i915, ccs, 0, surf.dst.size,
> > + PROT_READ | PROT_WRITE);
> > + memcpy(ccscopy, ccsmap, ccssize);
> > +
> > + /* corrupt ccs */
> > + for (int i = 0; i < surf.dst.size / sizeof(uint32_t); i++)
> > + ccsmap[i] = i;
> > + set_surf_object(&surf.src, ccs, REGION_SMEM, ccssize,
> > + 0, DIRECT_ACCESS);
> > + set_surf_object(&surf.dst, mid->handle, mid->region, mid->size,
> > + 0, INDIRECT_ACCESS);
> > + blt_ctrl_surf_copy(i915, ctx, e, ahnd, &surf);
> > +
> > + memset(&blt, 0, sizeof(blt));
> > + blt.color_depth = CD_32bit;
> > + blt.print_bb = param.print_bb;
> > + set_blt_object(&blt.src, mid);
> > + set_blt_object(&blt.dst, dst);
> > + set_object_ext(&ext.src, mid->compression_type, mid->x2, mid->y2, SURFACE_TYPE_2D);
> > + set_object_ext(&ext.dst, 0, dst->x2, dst->y2, SURFACE_TYPE_2D);
> > + set_batch(&blt.bb, bb2, bb_size, REGION_SMEM);
> > + blt_block_copy(i915, ctx, e, ahnd, &blt, &ext);
> > + gem_sync(i915, blt.dst.handle);
> > + WRITE_PNG(i915, run_id, "corrupted", &blt.dst, dst->x2, dst->y2);
> > + result = memcmp(src->ptr, dst->ptr, src->size);
> > + igt_assert(result != 0);
> > +
> > + /* retrieve back ccs */
> > + memcpy(ccsmap, ccscopy, ccssize);
> > + blt_ctrl_surf_copy(i915, ctx, e, ahnd, &surf);
> > +
> > + blt_block_copy(i915, ctx, e, ahnd, &blt, &ext);
> > + gem_sync(i915, blt.dst.handle);
> > + WRITE_PNG(i915, run_id, "corrected", &blt.dst, dst->x2, dst->y2);
> > + result = memcmp(src->ptr, dst->ptr, src->size);
> > + igt_assert(result == 0);
> > +
> > + munmap(ccsmap, ccssize);
> > + gem_close(i915, bb);
> > + gem_close(i915, bb2);
> > + gem_close(i915, ccs);
> > +}
> > +
> > +static void block_copy(int i915,
> > + const intel_ctx_t *ctx,
> > + const struct intel_execution_engine2 *e,
> > + uint32_t region1, uint32_t region2,
> > + enum blt_tiling mid_tiling, bool compression,
> > + bool inplace,
> > + bool surfcopy)
> > +{
> > + struct blt_copy_data blt = {};
> > + struct blt_block_copy_data_ext ext = {}, *pext = &ext;
> > + struct blt_copy_object *src, *mid, *dst;
> > + const uint32_t bpp = BPP;
>
> This BPP is used only once here, maybe change to full name ?
I like bpp - it is short and I don't know another acronym which
is used in graphics programming. But I wondered to initalize bpp to 32
here, without BPP macro on the top.
>
> > + uint64_t add_size = param.increase_mid_surf_size;
> > + uint64_t bb_size = 4096;
> > + uint64_t ahnd = intel_allocator_open_full(i915, ctx->id, 0, 0,
> > + INTEL_ALLOCATOR_SIMPLE,
> > + ALLOC_STRATEGY_LOW_TO_HIGH);
> > + uint32_t run_id = mid_tiling;
> > + uint32_t mid_region = region2, bb;
> > + uint32_t width = param.width, height = param.height;
> > + enum blt_compression mid_compression = compression;
> > + int mid_compression_format = param.compression_format;
> > + enum blt_compression_type comp_type = COMPRESSION_TYPE_3D;
> > + int result;
> > +
> > + igt_assert(__gem_create_in_memory_regions(i915, &bb, &bb_size, region1) == 0);
> > +
> > + if (!blt_supports_compression(i915))
> > + pext = NULL;
> > +
> > + src = create_object(i915, region1, width, height, bpp, 0,
> > + T_LINEAR, COMPRESSION_DISABLED, comp_type, true, 0);
> > + mid = create_object(i915, mid_region, width, height, bpp, 0,
> > + mid_tiling, mid_compression, comp_type, true, add_size);
> > + dst = create_object(i915, region1, width, height, bpp, 0,
> > + T_LINEAR, COMPRESSION_DISABLED, comp_type, true, 0);
> > + 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(i915, src, width, height);
> > + WRITE_PNG(i915, run_id, "src", src, width, height);
> > +
> > + memset(&blt, 0, sizeof(blt));
> > + blt.color_depth = CD_32bit;
> > + blt.print_bb = param.print_bb;
> > + set_blt_object(&blt.src, src);
> > + set_blt_object(&blt.dst, mid);
> > + set_object_ext(&ext.src, 0, width, height, SURFACE_TYPE_2D);
> > + set_object_ext(&ext.dst, mid_compression_format, width, height, SURFACE_TYPE_2D);
> > + set_batch(&blt.bb, bb, bb_size, region1);
> > +
> > + blt_block_copy(i915, ctx, e, ahnd, &blt, pext);
> > + gem_sync(i915, mid->handle);
> > +
> > + WRITE_PNG(i915, run_id, "src", &blt.src, width, height);
> > + WRITE_PNG(i915, run_id, "mid", &blt.dst, width, height);
> > +
> > + if (surfcopy && pext)
> > + surf_copy(i915, ctx, e, ahnd, src, mid, dst, run_id);
> > +
> > + memset(&blt, 0, sizeof(blt));
> > + blt.color_depth = CD_32bit;
> > + blt.print_bb = param.print_bb;
> > + set_blt_object(&blt.src, mid);
> > + set_blt_object(&blt.dst, dst);
> > + set_object_ext(&ext.src, mid_compression_format, width, height, SURFACE_TYPE_2D);
> > + set_object_ext(&ext.dst, 0, width, height, SURFACE_TYPE_2D);
> > + if (inplace) {
> > + set_object(&blt.dst, mid->handle, dst->size, mid->region, 0,
> > + T_LINEAR, COMPRESSION_DISABLED, comp_type);
> > + blt.dst.ptr = mid->ptr;
> > + }
> > +
> > + set_batch(&blt.bb, bb, bb_size, region1);
> > + blt_block_copy(i915, ctx, e, ahnd, &blt, pext);
> > + gem_sync(i915, blt.dst.handle);
> > + WRITE_PNG(i915, run_id, "dst", &blt.dst, width, height);
> > +
> > + result = memcmp(src->ptr, blt.dst.ptr, src->size);
> > +
> > + destroy_object(i915, src);
> > + destroy_object(i915, mid);
> > + destroy_object(i915, dst);
> > + gem_close(i915, bb);
> > + put_ahnd(ahnd);
> > +
> > + igt_assert_f(!result, "source and destination surfaces differs!\n");
> > +}
> > +
> > +static void block_copy_test(int i915,
> > + const struct test_config *config,
> > + const intel_ctx_t *ctx,
> > + struct igt_collection *set)
> > +{
> > + struct igt_collection *regions;
> > + const struct intel_execution_engine2 *e;
> > +
> > + if (config->compression && !blt_supports_compression(i915))
> > + return;
> > +
> > + if (config->inplace && !config->compression)
> > + return;
> > +
> > + for (int tiling = T_LINEAR; tiling <= T_TILE64; tiling++) {
> > + if (!blt_supports_tiling(i915, tiling) ||
> > + (param.tiling >= 0 && param.tiling != tiling))
> > + continue;
> > +
> > + for_each_ctx_engine(i915, ctx, e) {
> > + if (!gem_engine_can_block_copy(i915, e))
> > + continue;
> > +
> > + for_each_variation_r(regions, 2, set) {
> > + uint32_t region1, region2;
> > + char *regtxt;
> > +
> > + region1 = igt_collection_get_value(regions, 0);
> > + region2 = igt_collection_get_value(regions, 1);
> > +
> > + /* Compressed surface must be in device memory */
> > + if (config->compression && !IS_DEVICE_MEMORY_REGION(region2))
> > + continue;
> > +
> > + regtxt = memregion_dynamic_subtest_name(regions);
> > + igt_dynamic_f("%s-%s-compfmt%d-%s",
> > + blt_tiling_name(tiling),
> > + config->compression ?
> > + "compressed" : "uncompressed",
> > + param.compression_format, regtxt) {
> > + block_copy(i915, ctx, e,
> > + region1, region2,
> > + tiling,
> > + config->compression,
> > + config->inplace,
> > + config->surfcopy);
> > + }
> > + free(regtxt);
> > + }
> > + }
> > + }
> > +}
> > +
> > +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 'f':
> > + param.compression_format = atoi(optarg);
> > + igt_debug("Compression format: %d\n", param.compression_format);
> > + igt_assert((param.compression_format & ~0x1f) == 0);
> > + 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;
> > + case 'I':
> > + param.increase_mid_surf_size = atoi(optarg);
> > + igt_debug("Additional mid size: %x\n", param.increase_mid_surf_size);
> > + break;
> > + default:
> > + return IGT_OPT_HANDLER_ERROR;
> > + }
> > +
> > + return IGT_OPT_HANDLER_SUCCESS;
> > +}
> > +
> > +const char *help_str =
> > + " -b\tPrint bb\n"
> > + " -f\tCompression format (0-31)"
> > + " -p\tWrite PNG\n"
> > + " -s\tPrint surface info\n"
> > + " -t\tTiling format (0 - linear, 1 - XMAJOR, 2 - YMAJOR, 3 - TILE4, 4 - TILE64)\n"
> > + " -W\tWidth (default 512)\n"
> > + " -H\tHeight (default 512)\n"
> > + " -I\tIncrease Tile64 surface size (value)"
> > + ;
> > +
> > +igt_main_args("bf:pst:W:H:I:", NULL, help_str, opt_handler, NULL)
> > +{
> > + struct drm_i915_query_memory_regions *query_info;
> > + struct igt_collection *set;
> > + const intel_ctx_t *ctx;
> > + int i915;
> > + igt_hang_t hang;
> > +
> > + igt_fixture {
> > + i915 = drm_open_driver(DRIVER_INTEL);
> > + igt_require_gem(i915);
> > + igt_require(AT_LEAST_GEN(intel_get_drm_devid(i915), 12) > 0);
> > +
> > + query_info = gem_get_query_memory_regions(i915);
> > + igt_require(query_info);
> > +
> > + set = get_memory_region_set(query_info,
> > + I915_SYSTEM_MEMORY,
> > + I915_DEVICE_MEMORY);
> > + ctx = intel_ctx_create_all_physical(i915);
> > + hang = igt_allow_hang(i915, ctx->id, 0);
> > + }
> > +
> > + igt_subtest_with_dynamic("block-copy-uncompressed") {
> > + struct test_config config = {};
> > +
> > + block_copy_test(i915, &config, ctx, set);
> > + }
> > +
> > + igt_subtest_with_dynamic("block-copy-compressed") {
> > + struct test_config config = { .compression = true };
> > +
> > + block_copy_test(i915, &config, ctx, set);
> > + }
> > +
> > + igt_subtest_with_dynamic("block-copy-inplace") {
> > + struct test_config config = { .compression = true,
> > + .inplace = true };
> > +
> > + block_copy_test(i915, &config, ctx, set);
> > + }
> > +
> > + igt_subtest_with_dynamic("ctrl-surf-copy") {
> > + struct test_config config = { .compression = true,
> > + .surfcopy = true };
> > +
> > + block_copy_test(i915, &config, ctx, set);
> > + }
>
> imho we need descriptions of subtests here, before each one,
> use --describe to see where it is needed.
Agree, I'm going to add this.
--
Zbigniew
>
> > +
> > + igt_fixture {
> > + igt_disallow_hang(i915, hang);
> > + close(i915);
> > + }
> > +}
> > diff --git a/tests/meson.build b/tests/meson.build
> > index 3dbba7a1b..3af4576f4 100644
> > --- a/tests/meson.build
> > +++ b/tests/meson.build
> > @@ -109,6 +109,7 @@ i915_progs = [
> > 'gem_blits',
> > 'gem_busy',
> > 'gem_caching',
> > + 'gem_ccs',
> > 'gem_close',
> > 'gem_close_race',
> > 'gem_concurrent_blit',
> > --
> > 2.32.0
> >
>
> Rest looks good to me,
> regards,
> Kamil Konieczny
More information about the igt-dev
mailing list