[Mesa-dev] [PATCH 08/14] ac/gpu_info: add kernel_flushes_tc_l2_after_ib
Marek Olšák
maraeo at gmail.com
Thu May 3 00:19:46 UTC 2018
From: Marek Olšák <marek.olsak at amd.com>
---
src/amd/common/ac_gpu_info.c | 4 ++++
src/amd/common/ac_gpu_info.h | 1 +
src/gallium/drivers/radeonsi/si_gfx_cs.c | 3 +--
src/gallium/winsys/radeon/drm/radeon_drm_winsys.c | 1 +
4 files changed, 7 insertions(+), 2 deletions(-)
diff --git a/src/amd/common/ac_gpu_info.c b/src/amd/common/ac_gpu_info.c
index e0e30a4a572..5c1bab2e9a0 100644
--- a/src/amd/common/ac_gpu_info.c
+++ b/src/amd/common/ac_gpu_info.c
@@ -317,20 +317,23 @@ bool ac_query_gpu_info(int fd, amdgpu_device_handle dev,
info->has_local_buffers = info->drm_minor >= 20 &&
!info->has_dedicated_vram;
info->kernel_flushes_hdp_before_ib = true;
info->htile_cmask_support_1d_tiling = true;
info->si_TA_CS_BC_BASE_ADDR_allowed = true;
info->has_bo_metadata = true;
info->has_gpu_reset_status_query = true;
info->has_gpu_reset_counter_query = false;
info->has_eqaa_surface_allocator = true;
info->has_format_bc1_through_bc7 = true;
+ /* DRM 3.1.0 doesn't flush TC for VI correctly. */
+ info->kernel_flushes_tc_l2_after_ib = info->chip_class != VI ||
+ info->drm_minor >= 2;
info->num_render_backends = amdinfo->rb_pipes;
/* The value returned by the kernel driver was wrong. */
if (info->family == CHIP_KAVERI)
info->num_render_backends = 2;
info->clock_crystal_freq = amdinfo->gpu_counter_freq;
if (!info->clock_crystal_freq) {
fprintf(stderr, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
info->clock_crystal_freq = 1;
@@ -472,20 +475,21 @@ void ac_print_gpu_info(struct radeon_info *info)
printf(" has_ctx_priority = %u\n", info->has_ctx_priority);
printf(" has_local_buffers = %u\n", info->has_local_buffers);
printf(" kernel_flushes_hdp_before_ib = %u\n", info->kernel_flushes_hdp_before_ib);
printf(" htile_cmask_support_1d_tiling = %u\n", info->htile_cmask_support_1d_tiling);
printf(" si_TA_CS_BC_BASE_ADDR_allowed = %u\n", info->si_TA_CS_BC_BASE_ADDR_allowed);
printf(" has_bo_metadata = %u\n", info->has_bo_metadata);
printf(" has_gpu_reset_status_query = %u\n", info->has_gpu_reset_status_query);
printf(" has_gpu_reset_counter_query = %u\n", info->has_gpu_reset_counter_query);
printf(" has_eqaa_surface_allocator = %u\n", info->has_eqaa_surface_allocator);
printf(" has_format_bc1_through_bc7 = %u\n", info->has_format_bc1_through_bc7);
+ printf(" kernel_flushes_tc_l2_after_ib = %u\n", info->kernel_flushes_tc_l2_after_ib);
printf("Shader core info:\n");
printf(" max_shader_clock = %i\n", info->max_shader_clock);
printf(" num_good_compute_units = %i\n", info->num_good_compute_units);
printf(" max_se = %i\n", info->max_se);
printf(" max_sh_per_se = %i\n", info->max_sh_per_se);
printf("Render backend info:\n");
printf(" num_render_backends = %i\n", info->num_render_backends);
printf(" num_tile_pipes = %i\n", info->num_tile_pipes);
diff --git a/src/amd/common/ac_gpu_info.h b/src/amd/common/ac_gpu_info.h
index 9c4c6cb11f0..5e404714db6 100644
--- a/src/amd/common/ac_gpu_info.h
+++ b/src/amd/common/ac_gpu_info.h
@@ -97,20 +97,21 @@ struct radeon_info {
bool has_ctx_priority;
bool has_local_buffers;
bool kernel_flushes_hdp_before_ib;
bool htile_cmask_support_1d_tiling;
bool si_TA_CS_BC_BASE_ADDR_allowed;
bool has_bo_metadata;
bool has_gpu_reset_status_query;
bool has_gpu_reset_counter_query;
bool has_eqaa_surface_allocator;
bool has_format_bc1_through_bc7;
+ bool kernel_flushes_tc_l2_after_ib;
/* Shader cores. */
uint32_t r600_max_quad_pipes; /* wave size / 16 */
uint32_t max_shader_clock;
uint32_t num_good_compute_units;
uint32_t max_se; /* shader engines */
uint32_t max_sh_per_se; /* shader arrays per shader engine */
/* Render backends (color + depth blocks). */
uint32_t r300_num_gb_pipes;
diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index 0af16dd3474..ec74c1bc703 100644
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -67,22 +67,21 @@ void si_need_gfx_cs_space(struct si_context *ctx)
void si_flush_gfx_cs(struct si_context *ctx, unsigned flags,
struct pipe_fence_handle **fence)
{
struct radeon_winsys_cs *cs = ctx->gfx_cs;
struct radeon_winsys *ws = ctx->ws;
unsigned wait_flags = 0;
if (ctx->gfx_flush_in_progress)
return;
- if (ctx->chip_class == VI && ctx->screen->info.drm_minor <= 1) {
- /* DRM 3.1.0 doesn't flush TC for VI correctly. */
+ if (!ctx->screen->info.kernel_flushes_tc_l2_after_ib) {
wait_flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
SI_CONTEXT_CS_PARTIAL_FLUSH |
SI_CONTEXT_INV_GLOBAL_L2;
} else if (ctx->chip_class == SI) {
/* The kernel flushes L2 before shaders are finished. */
wait_flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
SI_CONTEXT_CS_PARTIAL_FLUSH;
} else if (!(flags & RADEON_FLUSH_START_NEXT_GFX_IB_NOW)) {
wait_flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
SI_CONTEXT_CS_PARTIAL_FLUSH;
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
index 7d20b7333b7..2f79d6691c3 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
@@ -535,20 +535,21 @@ static bool do_winsys_init(struct radeon_drm_winsys *ws)
ws->info.kernel_flushes_hdp_before_ib = ws->info.drm_minor >= 40;
/* HTILE is broken with 1D tiling on old kernels and CIK. */
ws->info.htile_cmask_support_1d_tiling = ws->info.chip_class != CIK ||
ws->info.drm_minor >= 38;
ws->info.si_TA_CS_BC_BASE_ADDR_allowed = ws->info.drm_minor >= 48;
ws->info.has_bo_metadata = false;
ws->info.has_gpu_reset_status_query = false;
ws->info.has_gpu_reset_counter_query = ws->info.drm_minor >= 43;
ws->info.has_eqaa_surface_allocator = false;
ws->info.has_format_bc1_through_bc7 = ws->info.drm_minor >= 31;
+ ws->info.kernel_flushes_tc_l2_after_ib = true;
ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
return true;
}
static void radeon_winsys_destroy(struct radeon_winsys *rws)
{
struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
--
2.17.0
More information about the mesa-dev
mailing list