[Mesa-dev] [PATCH 08/17] gallium/radeon: merge dirty_fb_counter and dirty_tex_descriptor_counter

Marek Olšák maraeo at gmail.com
Thu Jan 26 16:04:24 UTC 2017


From: Marek Olšák <marek.olsak at amd.com>

to simplify things in draw_vbo a little
---
 src/gallium/drivers/r600/r600_state_common.c  |  8 ++++----
 src/gallium/drivers/radeon/r600_pipe_common.h | 21 ++++++++-------------
 src/gallium/drivers/radeon/r600_texture.c     | 18 +++++-------------
 src/gallium/drivers/radeonsi/si_state_draw.c  | 16 +++++-----------
 4 files changed, 22 insertions(+), 41 deletions(-)

diff --git a/src/gallium/drivers/r600/r600_state_common.c b/src/gallium/drivers/r600/r600_state_common.c
index a9dbc8b..c8502e9 100644
--- a/src/gallium/drivers/r600/r600_state_common.c
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -1662,40 +1662,40 @@ void r600_emit_clip_misc_state(struct r600_context *rctx, struct r600_atom *atom
 }
 
 static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *dinfo)
 {
 	struct r600_context *rctx = (struct r600_context *)ctx;
 	struct pipe_draw_info info = *dinfo;
 	struct pipe_index_buffer ib = {};
 	struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
 	bool render_cond_bit = rctx->b.render_cond && !rctx->b.render_cond_force_off;
 	uint64_t mask;
-	unsigned num_patches, dirty_fb_counter;
+	unsigned num_patches, dirty_tex_counter;
 
 	if (!info.indirect && !info.count && (info.indexed || !info.count_from_stream_output)) {
 		return;
 	}
 
 	if (!rctx->vs_shader || !rctx->ps_shader) {
 		assert(0);
 		return;
 	}
 
 	/* make sure that the gfx ring is only one active */
 	if (radeon_emitted(rctx->b.dma.cs, 0)) {
 		rctx->b.dma.flush(rctx, RADEON_FLUSH_ASYNC, NULL);
 	}
 
 	/* Re-emit the framebuffer state if needed. */
-	dirty_fb_counter = p_atomic_read(&rctx->b.screen->dirty_fb_counter);
-	if (dirty_fb_counter != rctx->b.last_dirty_fb_counter) {
-		rctx->b.last_dirty_fb_counter = dirty_fb_counter;
+	dirty_tex_counter = p_atomic_read(&rctx->b.screen->dirty_tex_counter);
+	if (dirty_tex_counter != rctx->b.last_dirty_tex_counter) {
+		rctx->b.last_dirty_tex_counter = dirty_tex_counter;
 		r600_mark_atom_dirty(rctx, &rctx->framebuffer.atom);
 	}
 
 	if (!r600_update_derived_state(rctx)) {
 		/* useless to render because current rendering command
 		 * can't be achieved
 		 */
 		return;
 	}
 
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index afb1385..f3e42ee 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -411,39 +411,35 @@ struct r600_common_screen {
 	pipe_mutex			gpu_load_mutex;
 	pipe_thread			gpu_load_thread;
 	union r600_grbm_counters	grbm_counters;
 	volatile unsigned		gpu_load_stop_thread; /* bool */
 
 	char				renderer_string[100];
 
 	/* Performance counters. */
 	struct r600_perfcounters	*perfcounters;
 
-	/* If pipe_screen wants to re-emit the framebuffer state of all
-	 * contexts, it should atomically increment this. Each context will
-	 * compare this with its own last known value of the counter before
-	 * drawing and re-emit the framebuffer state accordingly.
+	/* If pipe_screen wants to recompute and re-emit the framebuffer,
+	 * sampler, and image states of all contexts, it should atomically
+	 * increment this.
+	 *
+	 * Each context will compare this with its own last known value of
+	 * the counter before drawing and re-emit the states accordingly.
 	 */
-	unsigned			dirty_fb_counter;
+	unsigned			dirty_tex_counter;
 
 	/* Atomically increment this counter when an existing texture's
 	 * metadata is enabled or disabled in a way that requires changing
 	 * contexts' compressed texture binding masks.
 	 */
 	unsigned			compressed_colortex_counter;
 
-	/* Atomically increment this counter when an existing texture's
-	 * backing buffer or tile mode parameters have changed that requires
-	 * recomputation of shader descriptors.
-	 */
-	unsigned			dirty_tex_descriptor_counter;
-
 	struct {
 		/* Context flags to set so that all writes from earlier jobs
 		 * in the CP are seen by L2 clients.
 		 */
 		unsigned cp_to_L2;
 
 		/* Context flags to set so that all writes from earlier
 		 * compute jobs are seen by L2 clients.
 		 */
 		unsigned compute_to_L2;
@@ -548,23 +544,22 @@ struct r600_common_context {
 	struct radeon_winsys_ctx	*ctx;
 	enum radeon_family		family;
 	enum chip_class			chip_class;
 	struct r600_ring		gfx;
 	struct r600_ring		dma;
 	struct pipe_fence_handle	*last_gfx_fence;
 	struct pipe_fence_handle	*last_sdma_fence;
 	unsigned			num_gfx_cs_flushes;
 	unsigned			initial_gfx_cs_size;
 	unsigned			gpu_reset_counter;
-	unsigned			last_dirty_fb_counter;
+	unsigned			last_dirty_tex_counter;
 	unsigned			last_compressed_colortex_counter;
-	unsigned			last_dirty_tex_descriptor_counter;
 
 	struct u_upload_mgr		*uploader;
 	struct u_suballocator		*allocator_zeroed_memory;
 	struct slab_child_pool		pool_transfers;
 
 	/* Current unaccounted memory usage. */
 	uint64_t			vram;
 	uint64_t			gtt;
 
 	/* States. */
diff --git a/src/gallium/drivers/radeon/r600_texture.c b/src/gallium/drivers/radeon/r600_texture.c
index 0b77c82..55c8e6f 100644
--- a/src/gallium/drivers/radeon/r600_texture.c
+++ b/src/gallium/drivers/radeon/r600_texture.c
@@ -291,25 +291,20 @@ static void r600_texture_init_metadata(struct r600_texture *rtex,
 	metadata->pipe_config = surface->pipe_config;
 	metadata->bankw = surface->bankw;
 	metadata->bankh = surface->bankh;
 	metadata->tile_split = surface->tile_split;
 	metadata->mtilea = surface->mtilea;
 	metadata->num_banks = surface->num_banks;
 	metadata->stride = surface->level[0].nblk_x * surface->bpe;
 	metadata->scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
 }
 
-static void r600_dirty_all_framebuffer_states(struct r600_common_screen *rscreen)
-{
-	p_atomic_inc(&rscreen->dirty_fb_counter);
-}
-
 static void r600_eliminate_fast_color_clear(struct r600_common_context *rctx,
 					    struct r600_texture *rtex)
 {
 	struct r600_common_screen *rscreen = rctx->screen;
 	struct pipe_context *ctx = &rctx->b;
 
 	if (ctx == rscreen->aux_context)
 		pipe_mutex_lock(rscreen->aux_context_lock);
 
 	ctx->flush_resource(ctx, &rtex->resource.b.b);
@@ -334,21 +329,21 @@ static void r600_texture_discard_cmask(struct r600_common_screen *rscreen,
 
 	if (rscreen->chip_class >= SI)
 		rtex->cb_color_info &= ~SI_S_028C70_FAST_CLEAR(1);
 	else
 		rtex->cb_color_info &= ~EG_S_028C70_FAST_CLEAR(1);
 
 	if (rtex->cmask_buffer != &rtex->resource)
 	    r600_resource_reference(&rtex->cmask_buffer, NULL);
 
 	/* Notify all contexts about the change. */
-	r600_dirty_all_framebuffer_states(rscreen);
+	p_atomic_inc(&rscreen->dirty_tex_counter);
 	p_atomic_inc(&rscreen->compressed_colortex_counter);
 }
 
 static bool r600_can_disable_dcc(struct r600_texture *rtex)
 {
 	/* We can't disable DCC if it can be written by another process. */
 	return rtex->dcc_offset &&
 	       (!rtex->resource.is_shared ||
 		!(rtex->resource.external_usage & PIPE_HANDLE_USAGE_WRITE));
 }
@@ -358,21 +353,21 @@ static bool r600_texture_discard_dcc(struct r600_common_screen *rscreen,
 {
 	if (!r600_can_disable_dcc(rtex))
 		return false;
 
 	assert(rtex->dcc_separate_buffer == NULL);
 
 	/* Disable DCC. */
 	rtex->dcc_offset = 0;
 
 	/* Notify all contexts about the change. */
-	r600_dirty_all_framebuffer_states(rscreen);
+	p_atomic_inc(&rscreen->dirty_tex_counter);
 	return true;
 }
 
 /**
  * Disable DCC for the texture. (first decompress, then discard metadata).
  *
  * There is unresolved multi-context synchronization issue between
  * screen::aux_context and the current context. If applications do this with
  * multiple contexts, it's already undefined behavior for them and we don't
  * have to worry about that. The scenario is:
@@ -473,22 +468,21 @@ static void r600_degrade_tile_mode_to_linear(struct r600_common_context *rctx,
 	rtex->cmask = new_tex->cmask; /* needed even without CMASK */
 
 	assert(!rtex->htile_buffer);
 	assert(!rtex->cmask.size);
 	assert(!rtex->fmask.size);
 	assert(!rtex->dcc_offset);
 	assert(!rtex->is_depth);
 
 	r600_texture_reference(&new_tex, NULL);
 
-	r600_dirty_all_framebuffer_states(rctx->screen);
-	p_atomic_inc(&rctx->screen->dirty_tex_descriptor_counter);
+	p_atomic_inc(&rctx->screen->dirty_tex_counter);
 }
 
 static boolean r600_texture_get_handle(struct pipe_screen* screen,
 				       struct pipe_context *ctx,
 				       struct pipe_resource *resource,
 				       struct winsys_handle *whandle,
                                        unsigned usage)
 {
 	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 	struct r600_common_context *rctx = (struct r600_common_context*)
@@ -1412,22 +1406,21 @@ static void r600_texture_invalidate_storage(struct r600_common_context *rctx,
 	assert(!rtex->is_depth);
 	assert(rtex->surface.is_linear);
 
 	/* Reallocate the buffer in the same pipe_resource. */
 	r600_alloc_resource(rscreen, &rtex->resource);
 
 	/* Initialize the CMASK base address (needed even without CMASK). */
 	rtex->cmask.base_address_reg =
 		(rtex->resource.gpu_address + rtex->cmask.offset) >> 8;
 
-	r600_dirty_all_framebuffer_states(rscreen);
-	p_atomic_inc(&rscreen->dirty_tex_descriptor_counter);
+	p_atomic_inc(&rscreen->dirty_tex_counter);
 
 	rctx->num_alloc_tex_transfer_bytes += rtex->size;
 }
 
 static void *r600_texture_transfer_map(struct pipe_context *ctx,
 				       struct pipe_resource *texture,
 				       unsigned level,
 				       unsigned usage,
 				       const struct pipe_box *box,
 				       struct pipe_transfer **ptransfer)
@@ -2396,22 +2389,21 @@ static void si_set_optimal_micro_tile_mode(struct r600_common_screen *rscreen,
 			}
 			break;
 		default: /* depth, thick */
 			assert(!"unexpected micro mode");
 			return;
 		}
 	}
 
 	rtex->surface.micro_tile_mode = rtex->last_msaa_resolve_target_micro_mode;
 
-	p_atomic_inc(&rscreen->dirty_fb_counter);
-	p_atomic_inc(&rscreen->dirty_tex_descriptor_counter);
+	p_atomic_inc(&rscreen->dirty_tex_counter);
 }
 
 void evergreen_do_fast_color_clear(struct r600_common_context *rctx,
 				   struct pipe_framebuffer_state *fb,
 				   struct r600_atom *fb_state,
 				   unsigned *buffers, unsigned *dirty_cbufs,
 				   const union pipe_color_union *color)
 {
 	int i;
 
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 0374841..c80d4d6 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -938,21 +938,21 @@ void si_ce_post_draw_synchronization(struct si_context *sctx)
 
 		sctx->ce_need_synchronization = false;
 	}
 }
 
 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
 	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
 	struct pipe_index_buffer ib = {};
-	unsigned mask, dirty_fb_counter, dirty_tex_counter, rast_prim;
+	unsigned mask, dirty_tex_counter, rast_prim;
 
 	if (likely(!info->indirect)) {
 		/* SI-CI treat instance_count==0 as instance_count==1. There is
 		 * no workaround for indirect draws, but we can at least skip
 		 * direct draws.
 		 */
 		if (unlikely(!info->instance_count))
 			return;
 
 		/* Handle count == 0. */
@@ -967,35 +967,29 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	}
 	if (unlikely(!sctx->ps_shader.cso && (!rs || !rs->rasterizer_discard))) {
 		assert(0);
 		return;
 	}
 	if (unlikely(!!sctx->tes_shader.cso != (info->mode == PIPE_PRIM_PATCHES))) {
 		assert(0);
 		return;
 	}
 
-	/* Re-emit the framebuffer state if needed. */
-	dirty_fb_counter = p_atomic_read(&sctx->b.screen->dirty_fb_counter);
-	if (unlikely(dirty_fb_counter != sctx->b.last_dirty_fb_counter)) {
-		sctx->b.last_dirty_fb_counter = dirty_fb_counter;
+	/* Recompute and re-emit the texture resource states if needed. */
+	dirty_tex_counter = p_atomic_read(&sctx->b.screen->dirty_tex_counter);
+	if (unlikely(dirty_tex_counter != sctx->b.last_dirty_tex_counter)) {
+		sctx->b.last_dirty_tex_counter = dirty_tex_counter;
 		sctx->framebuffer.dirty_cbufs |=
 			((1 << sctx->framebuffer.state.nr_cbufs) - 1);
 		sctx->framebuffer.dirty_zsbuf = true;
 		sctx->framebuffer.do_update_surf_dirtiness = true;
 		si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
-	}
-
-	/* Invalidate & recompute texture descriptors if needed. */
-	dirty_tex_counter = p_atomic_read(&sctx->b.screen->dirty_tex_descriptor_counter);
-	if (unlikely(dirty_tex_counter != sctx->b.last_dirty_tex_descriptor_counter)) {
-		sctx->b.last_dirty_tex_descriptor_counter = dirty_tex_counter;
 		si_update_all_texture_descriptors(sctx);
 	}
 
 	si_decompress_graphics_textures(sctx);
 
 	/* Set the rasterization primitive type.
 	 *
 	 * This must be done after si_decompress_textures, which can call
 	 * draw_vbo recursively, and before si_update_shaders, which uses
 	 * current_rast_prim for this draw_vbo call. */
-- 
2.7.4



More information about the mesa-dev mailing list