[Mesa-dev] [PATCH 04/11] radeonsi: de-atomize L2 prefetch

Marek Olšák maraeo at gmail.com
Sun Aug 6 22:20:19 UTC 2017


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

I'd like to be able to move the prefetch call site around.
---
 src/gallium/drivers/radeonsi/si_cp_dma.c        | 7 +++----
 src/gallium/drivers/radeonsi/si_descriptors.c   | 2 +-
 src/gallium/drivers/radeonsi/si_hw_context.c    | 2 +-
 src/gallium/drivers/radeonsi/si_pipe.h          | 3 ++-
 src/gallium/drivers/radeonsi/si_state.h         | 1 -
 src/gallium/drivers/radeonsi/si_state_draw.c    | 3 +++
 src/gallium/drivers/radeonsi/si_state_shaders.c | 2 +-
 7 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
index e42f260..9f0e506 100644
--- a/src/gallium/drivers/radeonsi/si_cp_dma.c
+++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
@@ -443,21 +443,21 @@ static void cik_prefetch_shader_async(struct si_context *sctx,
 				      struct si_pm4_state *state)
 {
 	if (state) {
 		struct pipe_resource *bo = &state->bo[0]->b.b;
 		assert(state->nbo == 1);
 
 		cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0);
 	}
 }
 
-static void cik_emit_prefetch_L2(struct si_context *sctx, struct r600_atom *atom)
+void cik_emit_prefetch_L2(struct si_context *sctx)
 {
 	/* Prefetch shaders and VBO descriptors to TC L2. */
 	if (si_pm4_state_changed(sctx, ls))
 		cik_prefetch_shader_async(sctx, sctx->queued.named.ls);
 	if (si_pm4_state_changed(sctx, hs))
 		cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
 	if (si_pm4_state_changed(sctx, es))
 		cik_prefetch_shader_async(sctx, sctx->queued.named.es);
 	if (si_pm4_state_changed(sctx, gs))
 		cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
@@ -466,19 +466,18 @@ static void cik_emit_prefetch_L2(struct si_context *sctx, struct r600_atom *atom
 
 	/* Vertex buffer descriptors are uploaded uncached, so prefetch
 	 * them right after the VS binary. */
 	if (sctx->vertex_buffer_pointer_dirty) {
 		cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b,
 					 sctx->vertex_buffers.buffer_offset,
 					 sctx->vertex_elements->desc_list_byte_size);
 	}
 	if (si_pm4_state_changed(sctx, ps))
 		cik_prefetch_shader_async(sctx, sctx->queued.named.ps);
+
+	sctx->prefetch_L2 = false;
 }
 
 void si_init_cp_dma_functions(struct si_context *sctx)
 {
 	sctx->b.clear_buffer = si_clear_buffer;
-
-	si_init_atom(sctx, &sctx->prefetch_L2, &sctx->atoms.s.prefetch_L2,
-		     cik_emit_prefetch_L2);
 }
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index ea5b89e..917b0e1 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -1170,21 +1170,21 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
 					      RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
 		}
 	}
 
 	/* Don't flush the const cache. It would have a very negative effect
 	 * on performance (confirmed by testing). New descriptors are always
 	 * uploaded to a fresh new buffer, so I don't think flushing the const
 	 * cache is needed. */
 	si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
 	if (sctx->b.chip_class >= CIK)
-		si_mark_atom_dirty(sctx, &sctx->prefetch_L2);
+		sctx->prefetch_L2 = true;
 	sctx->vertex_buffers_dirty = false;
 	sctx->vertex_buffer_pointer_dirty = true;
 	return true;
 }
 
 
 /* CONSTANT BUFFERS */
 
 static unsigned
 si_const_and_shader_buffer_descriptors_idx(unsigned shader)
diff --git a/src/gallium/drivers/radeonsi/si_hw_context.c b/src/gallium/drivers/radeonsi/si_hw_context.c
index f2dfcc7..756b159 100644
--- a/src/gallium/drivers/radeonsi/si_hw_context.c
+++ b/src/gallium/drivers/radeonsi/si_hw_context.c
@@ -210,21 +210,21 @@ void si_begin_new_cs(struct si_context *ctx)
 
 	if (ctx->ce_preamble_ib)
 		si_ce_enable_loads(ctx->ce_preamble_ib);
 	else if (ctx->ce_ib)
 		si_ce_enable_loads(ctx->ce_ib);
 
 	if (ctx->ce_ib)
 		si_ce_restore_all_descriptors_at_ib_start(ctx);
 
 	if (ctx->b.chip_class >= CIK)
-		si_mark_atom_dirty(ctx, &ctx->prefetch_L2);
+		ctx->prefetch_L2 = true;
 
 	/* CLEAR_STATE disables all colorbuffers, so only enable bound ones. */
 	ctx->framebuffer.dirty_cbufs =
 		u_bit_consecutive(0, ctx->framebuffer.state.nr_cbufs);
 	/* CLEAR_STATE disables the zbuffer, so only enable it if it's bound. */
 	ctx->framebuffer.dirty_zsbuf = ctx->framebuffer.state.zsbuf != NULL;
 	/* This should always be marked as dirty to set the framebuffer scissor
 	 * at least. */
 	si_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 1984299..d213886 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -276,31 +276,31 @@ struct si_context {
 	struct radeon_winsys_cs		*ce_ib;
 	struct radeon_winsys_cs		*ce_preamble_ib;
 	struct r600_resource		*ce_ram_saved_buffer;
 	struct u_suballocator		*ce_suballocator;
 	unsigned			ce_ram_saved_offset;
 	uint16_t			total_ce_ram_allocated;
 	bool				ce_need_synchronization:1;
 
 	bool				gfx_flush_in_progress:1;
 	bool				compute_is_busy:1;
+	bool				prefetch_L2:1;
 
 	/* Atoms (direct states). */
 	union si_state_atoms		atoms;
 	unsigned			dirty_atoms; /* mask */
 	/* PM4 states (precomputed immutable states) */
 	unsigned			dirty_states;
 	union si_state			queued;
 	union si_state			emitted;
 
 	/* Atom declarations. */
-	struct r600_atom		prefetch_L2;
 	struct si_framebuffer		framebuffer;
 	struct si_sample_locs		msaa_sample_locs;
 	struct r600_atom		db_render_state;
 	struct r600_atom		msaa_config;
 	struct si_sample_mask		sample_mask;
 	struct r600_atom		cb_render_state;
 	unsigned			last_cb_target_mask;
 	struct si_blend_color		blend_color;
 	struct r600_atom		clip_regs;
 	struct si_clip_state		clip_state;
@@ -477,20 +477,21 @@ void si_resource_copy_region(struct pipe_context *ctx,
 			   SI_CPDMA_SKIP_SYNC_BEFORE | \
 			   SI_CPDMA_SKIP_GFX_SYNC | \
 			   SI_CPDMA_SKIP_BO_LIST_UPDATE)
 
 void si_copy_buffer(struct si_context *sctx,
 		    struct pipe_resource *dst, struct pipe_resource *src,
 		    uint64_t dst_offset, uint64_t src_offset, unsigned size,
 		    unsigned user_flags);
 void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
 			      uint64_t offset, unsigned size);
+void cik_emit_prefetch_L2(struct si_context *sctx);
 void si_init_cp_dma_functions(struct si_context *sctx);
 
 /* si_debug.c */
 void si_init_debug_functions(struct si_context *sctx);
 void si_check_vm_faults(struct r600_common_context *ctx,
 			struct radeon_saved_cs *saved, enum ring_type ring);
 bool si_replace_shader(unsigned num, struct ac_shader_binary *binary);
 
 /* si_dma.c */
 void si_init_dma_functions(struct si_context *sctx);
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index acc8fb7..9fbede7 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -134,21 +134,20 @@ union si_state {
 		struct si_pm4_state		*ps;
 	} named;
 	struct si_pm4_state	*array[0];
 };
 
 #define SI_NUM_STATES (sizeof(union si_state) / sizeof(struct si_pm4_state *))
 
 union si_state_atoms {
 	struct {
 		/* The order matters. */
-		struct r600_atom *prefetch_L2;
 		struct r600_atom *render_cond;
 		struct r600_atom *streamout_begin;
 		struct r600_atom *streamout_enable; /* must be after streamout_begin */
 		struct r600_atom *framebuffer;
 		struct r600_atom *msaa_sample_locs;
 		struct r600_atom *db_render_state;
 		struct r600_atom *msaa_config;
 		struct r600_atom *sample_mask;
 		struct r600_atom *cb_render_state;
 		struct r600_atom *blend_color;
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 5254645..3f933fe 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -1339,20 +1339,23 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	/* GFX9 scissor bug workaround. There is also a more efficient but
 	 * more involved alternative workaround. */
 	if (sctx->b.chip_class == GFX9 &&
 	    si_is_atom_dirty(sctx, &sctx->b.scissors.atom))
 		sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
 
 	/* Flush caches before the first state atom, which does L2 prefetches. */
 	if (sctx->b.flags)
 		si_emit_cache_flush(sctx);
 
+	if (sctx->prefetch_L2)
+		cik_emit_prefetch_L2(sctx);
+
 	/* Emit state atoms. */
 	mask = sctx->dirty_atoms;
 	while (mask) {
 		struct r600_atom *atom = sctx->atoms.array[u_bit_scan(&mask)];
 
 		atom->emit(&sctx->b, atom);
 	}
 	sctx->dirty_atoms = 0;
 
 	/* Emit states. */
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index daf4af5..0dd6402 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -3301,21 +3301,21 @@ bool si_update_shaders(struct si_context *sctx)
 	    si_pm4_state_changed(sctx, hs) ||
 	    si_pm4_state_changed(sctx, es) ||
 	    si_pm4_state_changed(sctx, gs) ||
 	    si_pm4_state_changed(sctx, vs) ||
 	    si_pm4_state_changed(sctx, ps)) {
 		if (!si_update_spi_tmpring_size(sctx))
 			return false;
 	}
 
 	if (sctx->b.chip_class >= CIK)
-		si_mark_atom_dirty(sctx, &sctx->prefetch_L2);
+		sctx->prefetch_L2 = true;
 
 	sctx->do_update_shaders = false;
 	return true;
 }
 
 static void si_emit_scratch_state(struct si_context *sctx,
 				  struct r600_atom *atom)
 {
 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
 
-- 
2.7.4



More information about the mesa-dev mailing list