summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorMarek Olšák <marek.olsak@amd.com>2014-09-05 20:15:16 +0200
committerMarek Olšák <marek.olsak@amd.com>2016-05-10 17:20:09 +0200
commit3af28e558fb9cf71edd83e8cc9c3ddc33373cc93 (patch)
tree769f4c6d59b9a1ec3e6c1cea3b90bfea17d4be88
parentf475c9fb0708c312a4194558da36204becfafd35 (diff)
gallium/radeon: implement randomized SDMA texture copy testing (v2)
v2: - adjustments for exercising all important SDMA code paths - decrease the probability of getting huge sizes (faster testing) - increase the probability of getting power-of-two dimensions - change the memory cap to 128MB (faster testing) - better detect which engine has been used Reviewed-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
-rw-r--r--src/gallium/drivers/r600/r600_pipe.c3
-rw-r--r--src/gallium/drivers/radeon/Makefile.sources1
-rw-r--r--src/gallium/drivers/radeon/r600_pipe_common.c2
-rw-r--r--src/gallium/drivers/radeon/r600_pipe_common.h6
-rw-r--r--src/gallium/drivers/radeon/r600_test_dma.c404
-rw-r--r--src/gallium/drivers/radeonsi/si_pipe.c3
6 files changed, 419 insertions, 0 deletions
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index 52f0986e59..98c5a89f7d 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -707,5 +707,8 @@ struct pipe_screen *r600_screen_create(struct radeon_winsys *ws)
}
#endif
+ if (rscreen->b.debug_flags & DBG_TEST_DMA)
+ r600_test_dma(&rscreen->b);
+
return &rscreen->b.b;
}
diff --git a/src/gallium/drivers/radeon/Makefile.sources b/src/gallium/drivers/radeon/Makefile.sources
index f993d75a6a..6fbed81f6c 100644
--- a/src/gallium/drivers/radeon/Makefile.sources
+++ b/src/gallium/drivers/radeon/Makefile.sources
@@ -10,6 +10,7 @@ C_SOURCES := \
r600_query.c \
r600_query.h \
r600_streamout.c \
+ r600_test_dma.c \
r600_texture.c \
r600_viewport.c \
radeon_uvd.c \
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index b26dd9290d..cea6b7bce8 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -384,6 +384,8 @@ static const struct debug_named_value common_debug_options[] = {
{ "noasm", DBG_NO_ASM, "Don't print disassembled shaders"},
{ "preoptir", DBG_PREOPT_IR, "Print the LLVM IR before initial optimizations" },
+ { "testdma", DBG_TEST_DMA, "Invoke SDMA tests and exit." },
+
/* features */
{ "nodma", DBG_NO_ASYNC_DMA, "Disable asynchronous DMA" },
{ "nohyperz", DBG_NO_HYPERZ, "Disable Hyper-Z" },
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index e0db061c5e..47b73653e9 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -77,6 +77,8 @@
#define DBG_NO_TGSI (1 << 13)
#define DBG_NO_ASM (1 << 14)
#define DBG_PREOPT_IR (1 << 15)
+/* gaps */
+#define DBG_TEST_DMA (1 << 20)
/* Bits 21-31 are reserved for the r600g driver. */
/* features */
#define DBG_NO_ASYNC_DMA (1llu << 32)
@@ -484,6 +486,7 @@ struct r600_common_context {
unsigned max_db; /* for OQ */
/* Misc stats. */
unsigned num_draw_calls;
+ unsigned num_dma_calls;
/* Render condition. */
struct r600_atom render_cond_atom;
@@ -622,6 +625,9 @@ void r600_update_prims_generated_query_state(struct r600_common_context *rctx,
unsigned type, int diff);
void r600_streamout_init(struct r600_common_context *rctx);
+/* r600_test_dma.c */
+void r600_test_dma(struct r600_common_screen *rscreen);
+
/* r600_texture.c */
bool r600_prepare_for_dma_blit(struct r600_common_context *rctx,
struct r600_texture *rdst,
diff --git a/src/gallium/drivers/radeon/r600_test_dma.c b/src/gallium/drivers/radeon/r600_test_dma.c
new file mode 100644
index 0000000000..c203b4d9a2
--- /dev/null
+++ b/src/gallium/drivers/radeon/r600_test_dma.c
@@ -0,0 +1,404 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ */
+
+/* This file implements randomized SDMA texture blit tests. */
+
+#include "r600_pipe_common.h"
+#include "util/u_surface.h"
+
+static uint64_t seed_xorshift128plus[2];
+
+/* Super fast random number generator.
+ *
+ * This rand_xorshift128plus function by Sebastiano Vigna belongs
+ * to the public domain.
+ */
+static uint64_t rand_xorshift128plus(void)
+{
+ uint64_t *s = seed_xorshift128plus;
+
+ uint64_t s1 = s[0];
+ const uint64_t s0 = s[1];
+ s[0] = s0;
+ s1 ^= s1 << 23;
+ s[1] = s1 ^ s0 ^ (s1 >> 18) ^ (s0 >> 5);
+ return s[1] + s0;
+}
+
+#define RAND_NUM_SIZE 8
+
+/* The GPU blits are emulated on the CPU using these CPU textures. */
+
+struct cpu_texture {
+ uint8_t *ptr;
+ uint64_t size;
+ uint64_t layer_stride;
+ unsigned stride;
+};
+
+static void alloc_cpu_texture(struct cpu_texture *tex,
+ struct pipe_resource *templ, int bpp)
+{
+ tex->stride = align(templ->width0 * bpp, RAND_NUM_SIZE);
+ tex->layer_stride = (uint64_t)tex->stride * templ->height0;
+ tex->size = tex->layer_stride * templ->array_size;
+ tex->ptr = malloc(tex->size);
+ assert(tex->ptr);
+}
+
+static void set_random_pixels(struct pipe_context *ctx,
+ struct pipe_resource *tex,
+ struct cpu_texture *cpu)
+{
+ struct pipe_transfer *t;
+ uint8_t *map;
+ int x,y,z;
+
+ map = pipe_transfer_map_3d(ctx, tex, 0, PIPE_TRANSFER_WRITE,
+ 0, 0, 0, tex->width0, tex->height0,
+ tex->array_size, &t);
+ assert(map);
+
+ for (z = 0; z < tex->array_size; z++) {
+ for (y = 0; y < tex->height0; y++) {
+ uint64_t *ptr = (uint64_t*)
+ (map + t->layer_stride*z + t->stride*y);
+ uint64_t *ptr_cpu = (uint64_t*)
+ (cpu->ptr + cpu->layer_stride*z + cpu->stride*y);
+ unsigned size = cpu->stride / RAND_NUM_SIZE;
+
+ assert(t->stride % RAND_NUM_SIZE == 0);
+ assert(cpu->stride % RAND_NUM_SIZE == 0);
+
+ for (x = 0; x < size; x++)
+ *ptr++ = *ptr_cpu++ = rand_xorshift128plus();
+ }
+ }
+
+ pipe_transfer_unmap(ctx, t);
+}
+
+static bool compare_textures(struct pipe_context *ctx,
+ struct pipe_resource *tex,
+ struct cpu_texture *cpu, int bpp)
+{
+ struct pipe_transfer *t;
+ uint8_t *map;
+ int y,z;
+ bool pass = true;
+
+ map = pipe_transfer_map_3d(ctx, tex, 0, PIPE_TRANSFER_READ,
+ 0, 0, 0, tex->width0, tex->height0,
+ tex->array_size, &t);
+ assert(map);
+
+ for (z = 0; z < tex->array_size; z++) {
+ for (y = 0; y < tex->height0; y++) {
+ uint8_t *ptr = map + t->layer_stride*z + t->stride*y;
+ uint8_t *cpu_ptr = cpu->ptr +
+ cpu->layer_stride*z + cpu->stride*y;
+
+ if (memcmp(ptr, cpu_ptr, tex->width0 * bpp)) {
+ pass = false;
+ goto done;
+ }
+ }
+ }
+done:
+ pipe_transfer_unmap(ctx, t);
+ return pass;
+}
+
+static enum pipe_format get_format_from_bpp(int bpp)
+{
+ switch (bpp) {
+ case 1:
+ return PIPE_FORMAT_R8_UINT;
+ case 2:
+ return PIPE_FORMAT_R16_UINT;
+ case 4:
+ return PIPE_FORMAT_R32_UINT;
+ case 8:
+ return PIPE_FORMAT_R32G32_UINT;
+ case 16:
+ return PIPE_FORMAT_R32G32B32A32_UINT;
+ default:
+ assert(0);
+ return PIPE_FORMAT_NONE;
+ }
+}
+
+static const char *array_mode_to_string(unsigned mode)
+{
+ switch (mode) {
+ case RADEON_SURF_MODE_LINEAR_ALIGNED:
+ return "LINEAR_ALIGNED";
+ case RADEON_SURF_MODE_1D:
+ return "1D_TILED_THIN1";
+ case RADEON_SURF_MODE_2D:
+ return "2D_TILED_THIN1";
+ default:
+ assert(0);
+ return " UNKNOWN";
+ }
+}
+
+static unsigned generate_max_tex_side(unsigned max_tex_side)
+{
+ switch (rand() % 4) {
+ case 0:
+ /* Try to hit large sizes in 1/4 of the cases. */
+ return max_tex_side;
+ case 1:
+ /* Try to hit 1D tiling in 1/4 of the cases. */
+ return 128;
+ default:
+ /* Try to hit common sizes in 2/4 of the cases. */
+ return 2048;
+ }
+}
+
+void r600_test_dma(struct r600_common_screen *rscreen)
+{
+ struct pipe_screen *screen = &rscreen->b;
+ struct pipe_context *ctx = screen->context_create(screen, NULL, 0);
+ struct r600_common_context *rctx = (struct r600_common_context*)ctx;
+ uint64_t max_alloc_size;
+ unsigned i, iterations, num_partial_copies, max_levels, max_tex_side;
+ unsigned num_pass = 0, num_fail = 0;
+
+ max_levels = screen->get_param(screen, PIPE_CAP_MAX_TEXTURE_2D_LEVELS);
+ max_tex_side = 1 << (max_levels - 1);
+
+ /* Max 128 MB allowed for both textures. */
+ max_alloc_size = 128 * 1024 * 1024;
+
+ /* the seed for random test parameters */
+ srand(0x9b47d95b);
+ /* the seed for random pixel data */
+ seed_xorshift128plus[0] = 0x3bffb83978e24f88;
+ seed_xorshift128plus[1] = 0x9238d5d56c71cd35;
+
+ iterations = 1000000000; /* just kill it when you are bored */
+ num_partial_copies = 30;
+
+ /* These parameters are randomly generated per test:
+ * - whether to do one whole-surface copy or N partial copies per test
+ * - which tiling modes to use (LINEAR_ALIGNED, 1D, 2D)
+ * - which texture dimensions to use
+ * - whether to use VRAM (all tiling modes) and GTT (staging, linear
+ * only) allocations
+ * - random initial pixels in src
+ * - generate random subrectangle copies for partial blits
+ */
+ for (i = 0; i < iterations; i++) {
+ struct pipe_resource tsrc = {}, tdst = {}, *src, *dst;
+ struct r600_texture *rdst;
+ struct r600_texture *rsrc;
+ struct cpu_texture src_cpu, dst_cpu;
+ unsigned bpp, max_width, max_height, max_depth, j, num;
+ unsigned gfx_blits = 0, dma_blits = 0, max_tex_side_gen;
+ unsigned max_tex_layers;
+ bool pass;
+ bool do_partial_copies = rand() & 1;
+
+ /* generate a random test case */
+ tsrc.target = tdst.target = PIPE_TEXTURE_2D_ARRAY;
+ tsrc.depth0 = tdst.depth0 = 1;
+
+ bpp = 1 << (rand() % 5);
+ tsrc.format = tdst.format = get_format_from_bpp(bpp);
+
+ max_tex_side_gen = generate_max_tex_side(max_tex_side);
+ max_tex_layers = rand() % 4 ? 1 : 5;
+
+ tsrc.width0 = (rand() % max_tex_side_gen) + 1;
+ tsrc.height0 = (rand() % max_tex_side_gen) + 1;
+ tsrc.array_size = (rand() % max_tex_layers) + 1;
+
+ /* Have a 1/4 chance of getting power-of-two dimensions. */
+ if (rand() % 4 == 0) {
+ tsrc.width0 = util_next_power_of_two(tsrc.width0);
+ tsrc.height0 = util_next_power_of_two(tsrc.height0);
+ }
+
+ if (!do_partial_copies) {
+ /* whole-surface copies only, same dimensions */
+ tdst = tsrc;
+ } else {
+ max_tex_side_gen = generate_max_tex_side(max_tex_side);
+ max_tex_layers = rand() % 4 ? 1 : 5;
+
+ /* many partial copies, dimensions can be different */
+ tdst.width0 = (rand() % max_tex_side_gen) + 1;
+ tdst.height0 = (rand() % max_tex_side_gen) + 1;
+ tdst.array_size = (rand() % max_tex_layers) + 1;
+
+ /* Have a 1/4 chance of getting power-of-two dimensions. */
+ if (rand() % 4 == 0) {
+ tdst.width0 = util_next_power_of_two(tdst.width0);
+ tdst.height0 = util_next_power_of_two(tdst.height0);
+ }
+ }
+
+ /* check texture sizes */
+ if ((uint64_t)tsrc.width0 * tsrc.height0 * tsrc.array_size * bpp +
+ (uint64_t)tdst.width0 * tdst.height0 * tdst.array_size * bpp >
+ max_alloc_size) {
+ /* too large, try again */
+ i--;
+ continue;
+ }
+
+ /* VRAM + the tiling mode depends on dimensions (3/4 of cases),
+ * or GTT + linear only (1/4 of cases)
+ */
+ tsrc.usage = rand() % 4 ? PIPE_USAGE_DEFAULT : PIPE_USAGE_STAGING;
+ tdst.usage = rand() % 4 ? PIPE_USAGE_DEFAULT : PIPE_USAGE_STAGING;
+
+ /* Allocate textures (both the GPU and CPU copies).
+ * The CPU will emulate what the GPU should be doing.
+ */
+ src = screen->resource_create(screen, &tsrc);
+ dst = screen->resource_create(screen, &tdst);
+ assert(src);
+ assert(dst);
+ rdst = (struct r600_texture*)dst;
+ rsrc = (struct r600_texture*)src;
+ alloc_cpu_texture(&src_cpu, &tsrc, bpp);
+ alloc_cpu_texture(&dst_cpu, &tdst, bpp);
+
+ printf("%4u: dst = (%5u x %5u x %u, %s), "
+ " src = (%5u x %5u x %u, %s), bpp = %2u, ",
+ i, tdst.width0, tdst.height0, tdst.array_size,
+ array_mode_to_string(rdst->surface.level[0].mode),
+ tsrc.width0, tsrc.height0, tsrc.array_size,
+ array_mode_to_string(rsrc->surface.level[0].mode), bpp);
+ fflush(stdout);
+
+ /* set src pixels */
+ set_random_pixels(ctx, src, &src_cpu);
+
+ /* clear dst pixels */
+ rctx->clear_buffer(ctx, dst, 0, rdst->surface.bo_size, 0, true);
+ memset(dst_cpu.ptr, 0, dst_cpu.layer_stride * tdst.array_size);
+
+ /* preparation */
+ max_width = MIN2(tsrc.width0, tdst.width0);
+ max_height = MIN2(tsrc.height0, tdst.height0);
+ max_depth = MIN2(tsrc.array_size, tdst.array_size);
+
+ num = do_partial_copies ? num_partial_copies : 1;
+ for (j = 0; j < num; j++) {
+ int width, height, depth;
+ int srcx, srcy, srcz, dstx, dsty, dstz;
+ struct pipe_box box;
+ unsigned old_num_draw_calls = rctx->num_draw_calls;
+ unsigned old_num_dma_calls = rctx->num_dma_calls;
+
+ if (!do_partial_copies) {
+ /* copy whole src to dst */
+ width = max_width;
+ height = max_height;
+ depth = max_depth;
+
+ srcx = srcy = srcz = dstx = dsty = dstz = 0;
+ } else {
+ /* random sub-rectangle copies from src to dst */
+ depth = (rand() % max_depth) + 1;
+ srcz = rand() % (tsrc.array_size - depth + 1);
+ dstz = rand() % (tdst.array_size - depth + 1);
+
+ /* special code path to hit the tiled partial copies */
+ if (rsrc->surface.level[0].mode >= RADEON_SURF_MODE_1D &&
+ rdst->surface.level[0].mode >= RADEON_SURF_MODE_1D &&
+ rand() & 1) {
+ if (max_width < 8 || max_height < 8)
+ continue;
+ width = ((rand() % (max_width / 8)) + 1) * 8;
+ height = ((rand() % (max_height / 8)) + 1) * 8;
+
+ srcx = rand() % (tsrc.width0 - width + 1) & ~0x7;
+ srcy = rand() % (tsrc.height0 - height + 1) & ~0x7;
+
+ dstx = rand() % (tdst.width0 - width + 1) & ~0x7;
+ dsty = rand() % (tdst.height0 - height + 1) & ~0x7;
+ } else {
+ width = (rand() % max_width) + 1;
+ height = (rand() % max_height) + 1;
+
+ srcx = rand() % (tsrc.width0 - width + 1);
+ srcy = rand() % (tsrc.height0 - height + 1);
+
+ dstx = rand() % (tdst.width0 - width + 1);
+ dsty = rand() % (tdst.height0 - height + 1);
+ }
+
+ /* special code path to hit out-of-bounds reads in L2T */
+ if (rsrc->surface.level[0].mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
+ rdst->surface.level[0].mode >= RADEON_SURF_MODE_1D &&
+ rand() % 4 == 0) {
+ srcx = 0;
+ srcy = 0;
+ srcz = 0;
+ }
+ }
+
+ /* GPU copy */
+ u_box_3d(srcx, srcy, srcz, width, height, depth, &box);
+ rctx->dma_copy(ctx, dst, 0, dstx, dsty, dstz, src, 0, &box);
+
+ /* See which engine was used. */
+ gfx_blits += rctx->num_draw_calls > old_num_draw_calls;
+ dma_blits += rctx->num_dma_calls > old_num_dma_calls;
+
+ /* CPU copy */
+ util_copy_box(dst_cpu.ptr, tdst.format, dst_cpu.stride,
+ dst_cpu.layer_stride,
+ dstx, dsty, dstz, width, height, depth,
+ src_cpu.ptr, src_cpu.stride,
+ src_cpu.layer_stride,
+ srcx, srcy, srcz);
+ }
+
+ pass = compare_textures(ctx, dst, &dst_cpu, bpp);
+ if (pass)
+ num_pass++;
+ else
+ num_fail++;
+
+ printf("BLITs: GFX = %2u, DMA = %2u, %s [%u/%u]\n",
+ gfx_blits, dma_blits, pass ? "pass" : "fail",
+ num_pass, num_pass+num_fail);
+
+ /* cleanup */
+ pipe_resource_reference(&src, NULL);
+ pipe_resource_reference(&dst, NULL);
+ free(src_cpu.ptr);
+ free(dst_cpu.ptr);
+ }
+
+ ctx->destroy(ctx);
+ exit(0);
+}
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 61d5578116..1a69f43c8c 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -690,5 +690,8 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
/* Create the auxiliary context. This must be done last. */
sscreen->b.aux_context = sscreen->b.b.context_create(&sscreen->b.b, NULL, 0);
+ if (sscreen->b.debug_flags & DBG_TEST_DMA)
+ r600_test_dma(&sscreen->b);
+
return &sscreen->b.b;
}