[Mesa-dev] [PATCH 3/3] radeonsi/gfx9: implement primitive binning

Marek Olšák maraeo at gmail.com
Fri Sep 1 00:57:51 UTC 2017


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

This increases performance, but it was tuned for Raven, not Vega.
We don't know yet how Vega will perform, hopefully not worse.
---
 src/gallium/drivers/radeon/r600_pipe_common.c   |   2 +
 src/gallium/drivers/radeon/r600_pipe_common.h   |   2 +
 src/gallium/drivers/radeonsi/Makefile.sources   |   1 +
 src/gallium/drivers/radeonsi/si_hw_context.c    |   2 +
 src/gallium/drivers/radeonsi/si_pipe.c          |   5 +
 src/gallium/drivers/radeonsi/si_pipe.h          |   2 +
 src/gallium/drivers/radeonsi/si_state.c         |  26 +-
 src/gallium/drivers/radeonsi/si_state.h         |   6 +-
 src/gallium/drivers/radeonsi/si_state_binning.c | 448 ++++++++++++++++++++++++
 src/gallium/drivers/radeonsi/si_state_shaders.c |   2 +
 10 files changed, 489 insertions(+), 7 deletions(-)
 create mode 100644 src/gallium/drivers/radeonsi/si_state_binning.c

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 6b61901..1799b69 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -816,20 +816,22 @@ static const struct debug_named_value common_debug_options[] = {
 	{ "precompile", DBG_PRECOMPILE, "Compile one shader variant at shader creation." },
 	{ "nowc", DBG_NO_WC, "Disable GTT write combining" },
 	{ "check_vm", DBG_CHECK_VM, "Check VM faults and dump debug info." },
 	{ "nodcc", DBG_NO_DCC, "Disable DCC." },
 	{ "nodccclear", DBG_NO_DCC_CLEAR, "Disable DCC fast clear." },
 	{ "norbplus", DBG_NO_RB_PLUS, "Disable RB+." },
 	{ "sisched", DBG_SI_SCHED, "Enable LLVM SI Machine Instruction Scheduler." },
 	{ "mono", DBG_MONOLITHIC_SHADERS, "Use old-style monolithic shaders compiled on demand" },
 	{ "unsafemath", DBG_UNSAFE_MATH, "Enable unsafe math shader optimizations" },
 	{ "nodccfb", DBG_NO_DCC_FB, "Disable separate DCC on the main framebuffer" },
+	{ "nodpbb", DBG_NO_DPBB, "Disable DPBB." },
+	{ "nodfsm", DBG_NO_DFSM, "Disable DFSM." },
 
 	DEBUG_NAMED_VALUE_END /* must be last */
 };
 
 static const char* r600_get_vendor(struct pipe_screen* pscreen)
 {
 	return "X.Org";
 }
 
 static const char* r600_get_device_vendor(struct pipe_screen* pscreen)
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index 1f3c854..ca921f9 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -108,20 +108,22 @@ struct u_log_context;
 #define DBG_NO_DCC_CLEAR	(1ull << 44)
 #define DBG_NO_RB_PLUS		(1ull << 45)
 #define DBG_SI_SCHED		(1ull << 46)
 #define DBG_MONOLITHIC_SHADERS	(1ull << 47)
 /* gap */
 #define DBG_UNSAFE_MATH		(1ull << 49)
 #define DBG_NO_DCC_FB		(1ull << 50)
 #define DBG_TEST_VMFAULT_CP	(1ull << 51)
 #define DBG_TEST_VMFAULT_SDMA	(1ull << 52)
 #define DBG_TEST_VMFAULT_SHADER	(1ull << 53)
+#define DBG_NO_DPBB		(1ull << 54)
+#define DBG_NO_DFSM		(1ull << 55)
 
 #define R600_MAP_BUFFER_ALIGNMENT 64
 #define R600_MAX_VIEWPORTS        16
 
 #define SI_MAX_VARIABLE_THREADS_PER_BLOCK 1024
 
 enum r600_coherency {
 	R600_COHERENCY_NONE, /* no cache flushes needed */
 	R600_COHERENCY_SHADER,
 	R600_COHERENCY_CB_META,
diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
index 74df4d7..3795bed 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -20,14 +20,15 @@ C_SOURCES := \
 	si_perfcounter.c \
 	si_public.h \
 	si_shader.c \
 	si_shader.h \
 	si_shader_internal.h \
 	si_shader_nir.c \
 	si_shader_tgsi_alu.c \
 	si_shader_tgsi_mem.c \
 	si_shader_tgsi_setup.c \
 	si_state.c \
+	si_state_binning.c \
 	si_state_draw.c \
 	si_state_shaders.c \
 	si_state.h \
 	si_uvd.c
diff --git a/src/gallium/drivers/radeonsi/si_hw_context.c b/src/gallium/drivers/radeonsi/si_hw_context.c
index f3af8dc..d090eea 100644
--- a/src/gallium/drivers/radeonsi/si_hw_context.c
+++ b/src/gallium/drivers/radeonsi/si_hw_context.c
@@ -232,20 +232,22 @@ void si_begin_new_cs(struct si_context *ctx)
 	si_mark_atom_dirty(ctx, &ctx->msaa_sample_locs.atom);
 	si_mark_atom_dirty(ctx, &ctx->msaa_config);
 	/* CLEAR_STATE sets 0xffff. */
 	if (!has_clear_state || ctx->sample_mask.sample_mask != 0xffff)
 		si_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
 	si_mark_atom_dirty(ctx, &ctx->cb_render_state);
 	/* CLEAR_STATE sets zeros. */
 	if (!has_clear_state || ctx->blend_color.any_nonzeros)
 		si_mark_atom_dirty(ctx, &ctx->blend_color.atom);
 	si_mark_atom_dirty(ctx, &ctx->db_render_state);
+	if (ctx->b.chip_class >= GFX9)
+		si_mark_atom_dirty(ctx, &ctx->dpbb_state);
 	si_mark_atom_dirty(ctx, &ctx->stencil_ref.atom);
 	si_mark_atom_dirty(ctx, &ctx->spi_map);
 	si_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
 	si_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
 	si_all_descriptors_begin_new_cs(ctx);
 	si_all_resident_buffers_begin_new_cs(ctx);
 
 	ctx->b.scissors.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
 	ctx->b.viewports.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
 	ctx->b.viewports.depth_range_dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index b27159b..12a1eeb 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -1042,20 +1042,25 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 		 sscreen->b.info.me_fw_version >= 173) ||
 		(sscreen->b.chip_class == SI &&
 		 sscreen->b.info.pfp_fw_version >= 79 &&
 		 sscreen->b.info.me_fw_version >= 142);
 
 	sscreen->has_ds_bpermute = sscreen->b.chip_class >= VI;
 	sscreen->has_msaa_sample_loc_bug = (sscreen->b.family >= CHIP_POLARIS10 &&
 					    sscreen->b.family <= CHIP_POLARIS12) ||
 					   sscreen->b.family == CHIP_VEGA10 ||
 					   sscreen->b.family == CHIP_RAVEN;
+	sscreen->dpbb_allowed = sscreen->b.chip_class >= GFX9 &&
+				!(sscreen->b.debug_flags & DBG_NO_DPBB);
+	sscreen->dfsm_allowed = sscreen->dpbb_allowed &&
+				!(sscreen->b.debug_flags & DBG_NO_DFSM);
+
 	/* While it would be nice not to have this flag, we are constrained
 	 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
 	 * on GFX9.
 	 */
 	sscreen->llvm_has_working_vgpr_indexing = sscreen->b.chip_class <= VI;
 
 	sscreen->b.has_cp_dma = true;
 	sscreen->b.has_streamout = true;
 
 	/* Some chips have RB+ registers, but don't support RB+. Those must
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 0228b32..fbf8db9 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -89,20 +89,21 @@ struct u_suballocator;
 
 struct si_screen {
 	struct r600_common_screen	b;
 	unsigned			gs_table_depth;
 	unsigned			tess_offchip_block_dw_size;
 	bool				has_clear_state;
 	bool				has_distributed_tess;
 	bool				has_draw_indirect_multi;
 	bool				has_ds_bpermute;
 	bool				has_msaa_sample_loc_bug;
+	bool				dpbb_allowed;
 	bool				dfsm_allowed;
 	bool				llvm_has_working_vgpr_indexing;
 
 	/* Whether shaders are monolithic (1-part) or separate (3-part). */
 	bool				use_monolithic_shaders;
 	bool				record_llvm_ir;
 
 	mtx_t			shader_parts_mutex;
 	struct si_shader_part		*vs_prologs;
 	struct si_shader_part		*tcs_epilogs;
@@ -302,20 +303,21 @@ struct si_context {
 	unsigned			dirty_atoms; /* mask */
 	/* PM4 states (precomputed immutable states) */
 	unsigned			dirty_states;
 	union si_state			queued;
 	union si_state			emitted;
 
 	/* Atom declarations. */
 	struct si_framebuffer		framebuffer;
 	struct si_sample_locs		msaa_sample_locs;
 	struct r600_atom		db_render_state;
+	struct r600_atom		dpbb_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;
 	struct si_shader_data		shader_pointers;
 	struct si_stencil_ref		stencil_ref;
 	struct r600_atom		spi_map;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 076b583..5bbfbdd 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -616,20 +616,27 @@ static void si_bind_blend_state(struct pipe_context *ctx, void *state)
 	si_pm4_bind_state(sctx, blend, state);
 
 	if (!old_blend ||
 	    old_blend->cb_target_mask != blend->cb_target_mask ||
 	    old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
 	    old_blend->alpha_to_one != blend->alpha_to_one ||
 	    old_blend->dual_src_blend != blend->dual_src_blend ||
 	    old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
 	    old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
 		sctx->do_update_shaders = true;
+
+	if (sctx->screen->dpbb_allowed &&
+	    (!old_blend ||
+	     old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
+	     old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
+	     old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
+		si_mark_atom_dirty(sctx, &sctx->dpbb_state);
 }
 
 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
 	si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
 }
 
 static void si_set_blend_color(struct pipe_context *ctx,
 			       const struct pipe_blend_color *state)
@@ -1133,20 +1140,27 @@ static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
 	si_pm4_bind_state(sctx, dsa, dsa);
 
 	if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
 		   sizeof(struct si_dsa_stencil_ref_part)) != 0) {
 		sctx->stencil_ref.dsa_part = dsa->stencil_ref;
 		si_mark_atom_dirty(sctx, &sctx->stencil_ref.atom);
 	}
 
 	if (!old_dsa || old_dsa->alpha_func != dsa->alpha_func)
 		sctx->do_update_shaders = true;
+
+	if (sctx->screen->dpbb_allowed &&
+	    (!old_dsa ||
+	     (old_dsa->depth_enabled != dsa->depth_enabled ||
+	      old_dsa->stencil_enabled != dsa->stencil_enabled ||
+	      old_dsa->db_can_write != dsa->db_can_write)))
+		si_mark_atom_dirty(sctx, &sctx->dpbb_state);
 }
 
 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
 	si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
 }
 
 static void *si_create_db_flush_dsa(struct si_context *sctx)
 {
@@ -2684,20 +2698,23 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
 		if (vi_tc_compat_htile_enabled(rtex, surf->base.u.tex.level))
 			sctx->framebuffer.DB_has_shader_readable_metadata = true;
 
 		r600_context_add_resource_size(ctx, surf->base.texture);
 	}
 
 	si_update_poly_offset_state(sctx);
 	si_mark_atom_dirty(sctx, &sctx->cb_render_state);
 	si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
 
+	if (sctx->screen->dpbb_allowed)
+		si_mark_atom_dirty(sctx, &sctx->dpbb_state);
+
 	if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
 		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 
 	if (sctx->framebuffer.nr_samples != old_nr_samples) {
 		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 		si_mark_atom_dirty(sctx, &sctx->db_render_state);
 
 		/* Set sample locations as fragment shader constants. */
 		switch (sctx->framebuffer.nr_samples) {
 		case 1:
@@ -3065,20 +3082,22 @@ static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
 	struct si_context *sctx = (struct si_context *)ctx;
 
 	if (sctx->ps_iter_samples == min_samples)
 		return;
 
 	sctx->ps_iter_samples = min_samples;
 	sctx->do_update_shaders = true;
 
 	if (sctx->framebuffer.nr_samples > 1)
 		si_mark_atom_dirty(sctx, &sctx->msaa_config);
+	if (sctx->screen->dpbb_allowed)
+		si_mark_atom_dirty(sctx, &sctx->dpbb_state);
 }
 
 /*
  * Samplers
  */
 
 /**
  * Build the sampler view descriptor for a buffer texture.
  * @param state 256-bit descriptor; only the high 128 bits are filled in
  */
@@ -4124,20 +4143,21 @@ void si_init_state_functions(struct si_context *sctx)
 {
 	si_init_external_atom(sctx, &sctx->b.render_cond_atom, &sctx->atoms.s.render_cond);
 	si_init_external_atom(sctx, &sctx->b.streamout.begin_atom, &sctx->atoms.s.streamout_begin);
 	si_init_external_atom(sctx, &sctx->b.streamout.enable_atom, &sctx->atoms.s.streamout_enable);
 	si_init_external_atom(sctx, &sctx->b.scissors.atom, &sctx->atoms.s.scissors);
 	si_init_external_atom(sctx, &sctx->b.viewports.atom, &sctx->atoms.s.viewports);
 
 	si_init_atom(sctx, &sctx->framebuffer.atom, &sctx->atoms.s.framebuffer, si_emit_framebuffer_state);
 	si_init_atom(sctx, &sctx->msaa_sample_locs.atom, &sctx->atoms.s.msaa_sample_locs, si_emit_msaa_sample_locs);
 	si_init_atom(sctx, &sctx->db_render_state, &sctx->atoms.s.db_render_state, si_emit_db_render_state);
+	si_init_atom(sctx, &sctx->dpbb_state, &sctx->atoms.s.dpbb_state, si_emit_dpbb_state);
 	si_init_atom(sctx, &sctx->msaa_config, &sctx->atoms.s.msaa_config, si_emit_msaa_config);
 	si_init_atom(sctx, &sctx->sample_mask.atom, &sctx->atoms.s.sample_mask, si_emit_sample_mask);
 	si_init_atom(sctx, &sctx->cb_render_state, &sctx->atoms.s.cb_render_state, si_emit_cb_render_state);
 	si_init_atom(sctx, &sctx->blend_color.atom, &sctx->atoms.s.blend_color, si_emit_blend_color);
 	si_init_atom(sctx, &sctx->clip_regs, &sctx->atoms.s.clip_regs, si_emit_clip_regs);
 	si_init_atom(sctx, &sctx->clip_state.atom, &sctx->atoms.s.clip_state, si_emit_clip_state);
 	si_init_atom(sctx, &sctx->stencil_ref.atom, &sctx->atoms.s.stencil_ref, si_emit_stencil_ref);
 
 	sctx->b.b.create_blend_state = si_create_blend_state;
 	sctx->b.b.bind_blend_state = si_bind_blend_state;
@@ -4736,26 +4756,20 @@ static void si_init_config(struct si_context *sctx)
 		case CHIP_VEGA10:
 			pc_lines = 4096;
 			break;
 		case CHIP_RAVEN:
 			pc_lines = 1024;
 			break;
 		default:
 			assert(0);
 		}
 
-		si_pm4_set_reg(pm4, R_028060_DB_DFSM_CONTROL,
-			       S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF));
-		/* TODO: Enable the binner: */
-		si_pm4_set_reg(pm4, R_028C44_PA_SC_BINNER_CNTL_0,
-			       S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
-			       S_028C44_DISABLE_START_OF_PRIM(1));
 		si_pm4_set_reg(pm4, R_028C48_PA_SC_BINNER_CNTL_1,
 			       S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
 			       S_028C48_MAX_PRIM_PER_BATCH(1023));
 		si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
 			       S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
 		si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
 	}
 
 	si_pm4_upload_indirect_buffer(sctx, pm4);
 	sctx->init_config = pm4;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 9770960..17d210a 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -146,20 +146,21 @@ union si_state {
 
 union si_state_atoms {
 	struct {
 		/* The order matters. */
 		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 *dpbb_state;
 		struct r600_atom *msaa_config;
 		struct r600_atom *sample_mask;
 		struct r600_atom *cb_render_state;
 		struct r600_atom *blend_color;
 		struct r600_atom *clip_regs;
 		struct r600_atom *clip_state;
 		struct r600_atom *shader_pointers;
 		struct r600_atom *scissors;
 		struct r600_atom *viewports;
 		struct r600_atom *stencil_ref;
@@ -366,21 +367,24 @@ si_make_texture_descriptor(struct si_screen *screen,
 			   uint32_t *state,
 			   uint32_t *fmask_state);
 struct pipe_sampler_view *
 si_create_sampler_view_custom(struct pipe_context *ctx,
 			      struct pipe_resource *texture,
 			      const struct pipe_sampler_view *state,
 			      unsigned width0, unsigned height0,
 			      unsigned force_level);
 void si_update_fb_dirtiness_after_rendering(struct si_context *sctx);
 
-/* si_state_shader.c */
+/* si_state_binning.c */
+void si_emit_dpbb_state(struct si_context *sctx, struct r600_atom *state);
+
+/* si_state_shaders.c */
 bool si_update_shaders(struct si_context *sctx);
 void si_init_shader_functions(struct si_context *sctx);
 bool si_init_shader_cache(struct si_screen *sscreen);
 void si_destroy_shader_cache(struct si_screen *sscreen);
 void si_init_shader_selector_async(void *job, int thread_index);
 void si_get_active_slot_masks(const struct tgsi_shader_info *info,
 			      uint32_t *const_and_shader_buffers,
 			      uint64_t *samplers_and_images);
 
 /* si_state_draw.c */
diff --git a/src/gallium/drivers/radeonsi/si_state_binning.c b/src/gallium/drivers/radeonsi/si_state_binning.c
new file mode 100644
index 0000000..56bcdc8
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/si_state_binning.c
@@ -0,0 +1,448 @@
+/*
+ * Copyright 2017 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
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, 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 NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS 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 handles register programming of primitive binning. */
+
+#include "si_pipe.h"
+#include "sid.h"
+#include "gfx9d.h"
+#include "radeon/r600_cs.h"
+
+struct uvec2 {
+	unsigned x, y;
+};
+
+struct si_bin_size_map {
+	unsigned start;
+	unsigned bin_size_x;
+	unsigned bin_size_y;
+};
+
+typedef struct si_bin_size_map si_bin_size_subtable[3][9];
+
+/* Find the bin size where sum is >= table[i].start and < table[i + 1].start. */
+static struct uvec2 si_find_bin_size(struct si_screen *sscreen,
+				     const si_bin_size_subtable table[],
+				     unsigned sum)
+{
+	unsigned log_num_rb_per_se =
+		util_logbase2_ceil(sscreen->b.info.num_render_backends /
+				   sscreen->b.info.max_se);
+	unsigned log_num_se = util_logbase2_ceil(sscreen->b.info.max_se);
+	unsigned i;
+
+	/* Get the chip-specific subtable. */
+	const struct si_bin_size_map *subtable =
+		&table[log_num_rb_per_se][log_num_se][0];
+
+	for (i = 0; subtable[i].bin_size_x != 0; i++) {
+		if (sum >= subtable[i].start && sum < subtable[i + 1].start)
+			break;
+	}
+
+	struct uvec2 size = {subtable[i].bin_size_x, subtable[i].bin_size_y};
+	return size;
+}
+
+static struct uvec2 si_get_color_bin_size(struct si_context *sctx,
+					  unsigned cb_target_enabled_4bit)
+{
+	unsigned nr_samples = sctx->framebuffer.nr_samples;
+	unsigned sum = 0;
+
+	/* Compute the sum of all Bpp. */
+	for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
+		if (!(cb_target_enabled_4bit & (0xf << (i * 4))))
+			continue;
+
+		struct r600_texture *rtex =
+			(struct r600_texture*)sctx->framebuffer.state.cbufs[i]->texture;
+		sum += rtex->surface.bpe;
+	}
+
+	/* Multiply the sum by some function of the number of samples. */
+	if (nr_samples >= 2) {
+		if (sctx->ps_iter_samples >= 2)
+			sum *= nr_samples;
+		else
+			sum *= 2;
+	}
+
+	static const si_bin_size_subtable table[] = {
+		{
+			/* One RB / SE */
+			{
+				/* One shader engine */
+				{        0,  128,  128 },
+				{        1,   64,  128 },
+				{        2,   32,  128 },
+				{        3,   16,  128 },
+				{       17,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Two shader engines */
+				{        0,  128,  128 },
+				{        2,   64,  128 },
+				{        3,   32,  128 },
+				{        5,   16,  128 },
+				{       17,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Four shader engines */
+				{        0,  128,  128 },
+				{        3,   64,  128 },
+				{        5,   16,  128 },
+				{       17,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+		{
+			/* Two RB / SE */
+			{
+				/* One shader engine */
+				{        0,  128,  128 },
+				{        2,   64,  128 },
+				{        3,   32,  128 },
+				{        5,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Two shader engines */
+				{        0,  128,  128 },
+				{        3,   64,  128 },
+				{        5,   32,  128 },
+				{        9,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Four shader engines */
+				{        0,  256,  256 },
+				{        2,  128,  256 },
+				{        3,  128,  128 },
+				{        5,   64,  128 },
+				{        9,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+		{
+			/* Four RB / SE */
+			{
+				/* One shader engine */
+				{        0,  128,  256 },
+				{        2,  128,  128 },
+				{        3,   64,  128 },
+				{        5,   32,  128 },
+				{        9,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Two shader engines */
+				{        0,  256,  256 },
+				{        2,  128,  256 },
+				{        3,  128,  128 },
+				{        5,   64,  128 },
+				{        9,   32,  128 },
+				{       17,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				/* Four shader engines */
+				{        0,  256,  512 },
+				{        2,  256,  256 },
+				{        3,  128,  256 },
+				{        5,  128,  128 },
+				{        9,   64,  128 },
+				{       17,   16,  128 },
+				{       33,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+	};
+
+	return si_find_bin_size(sctx->screen, table, sum);
+}
+
+static struct uvec2 si_get_depth_bin_size(struct si_context *sctx)
+{
+	struct si_state_dsa *dsa = sctx->queued.named.dsa;
+
+	if (!sctx->framebuffer.state.zsbuf ||
+	    (!dsa->depth_enabled && !dsa->stencil_enabled)) {
+		/* Return the max size. */
+		struct uvec2 size = {512, 512};
+		return size;
+	}
+
+	struct r600_texture *rtex =
+		(struct r600_texture*)sctx->framebuffer.state.zsbuf->texture;
+	unsigned depth_coeff = dsa->depth_enabled ? 5 : 0;
+	unsigned stencil_coeff = rtex->surface.flags & RADEON_SURF_SBUFFER &&
+				 dsa->stencil_enabled ? 1 : 0;
+	unsigned sum = 4 * (depth_coeff + stencil_coeff) *
+		       sctx->framebuffer.nr_samples;
+
+	static const si_bin_size_subtable table[] = {
+		{
+			// One RB / SE
+			{
+				// One shader engine
+				{        0,  128,  256 },
+				{        2,  128,  128 },
+				{        4,   64,  128 },
+				{        7,   32,  128 },
+				{       13,   16,  128 },
+				{       49,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Two shader engines
+				{        0,  256,  256 },
+				{        2,  128,  256 },
+				{        4,  128,  128 },
+				{        7,   64,  128 },
+				{       13,   32,  128 },
+				{       25,   16,  128 },
+				{       49,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Four shader engines
+				{        0,  256,  512 },
+				{        2,  256,  256 },
+				{        4,  128,  256 },
+				{        7,  128,  128 },
+				{       13,   64,  128 },
+				{       25,   16,  128 },
+				{       49,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+		{
+			// Two RB / SE
+			{
+				// One shader engine
+				{        0,  256,  256 },
+				{        2,  128,  256 },
+				{        4,  128,  128 },
+				{        7,   64,  128 },
+				{       13,   32,  128 },
+				{       25,   16,  128 },
+				{       97,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Two shader engines
+				{        0,  256,  512 },
+				{        2,  256,  256 },
+				{        4,  128,  256 },
+				{        7,  128,  128 },
+				{       13,   64,  128 },
+				{       25,   32,  128 },
+				{       49,   16,  128 },
+				{       97,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Four shader engines
+				{        0,  512,  512 },
+				{        2,  256,  512 },
+				{        4,  256,  256 },
+				{        7,  128,  256 },
+				{       13,  128,  128 },
+				{       25,   64,  128 },
+				{       49,   16,  128 },
+				{       97,    0,    0 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+		{
+			// Four RB / SE
+			{
+				// One shader engine
+				{        0,  256,  512 },
+				{        2,  256,  256 },
+				{        4,  128,  256 },
+				{        7,  128,  128 },
+				{       13,   64,  128 },
+				{       25,   32,  128 },
+				{       49,   16,  128 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Two shader engines
+				{        0,  512,  512 },
+				{        2,  256,  512 },
+				{        4,  256,  256 },
+				{        7,  128,  256 },
+				{       13,  128,  128 },
+				{       25,   64,  128 },
+				{       49,   32,  128 },
+				{       97,   16,  128 },
+				{ UINT_MAX,    0,    0 },
+			},
+			{
+				// Four shader engines
+				{        0,  512,  512 },
+				{        4,  256,  512 },
+				{        7,  256,  256 },
+				{       13,  128,  256 },
+				{       25,  128,  128 },
+				{       49,   64,  128 },
+				{       97,   16,  128 },
+				{ UINT_MAX,    0,    0 },
+			},
+		},
+	};
+
+	return si_find_bin_size(sctx->screen, table, sum);
+}
+
+static void si_emit_dpbb_disable(struct si_context *sctx)
+{
+	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+
+	radeon_set_context_reg(cs, R_028C44_PA_SC_BINNER_CNTL_0,
+			       S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
+			       S_028C44_DISABLE_START_OF_PRIM(1));
+	radeon_set_context_reg(cs, R_028060_DB_DFSM_CONTROL,
+			       S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF));
+}
+
+void si_emit_dpbb_state(struct si_context *sctx, struct r600_atom *state)
+{
+	struct si_screen *sscreen = sctx->screen;
+	struct si_state_blend *blend = sctx->queued.named.blend;
+	struct si_state_dsa *dsa = sctx->queued.named.dsa;
+	unsigned db_shader_control = sctx->ps_db_shader_control;
+
+	assert(sctx->b.chip_class >= GFX9);
+
+	if (!sscreen->dpbb_allowed || !blend || !dsa) {
+		si_emit_dpbb_disable(sctx);
+		return;
+	}
+
+	bool ps_can_kill = G_02880C_KILL_ENABLE(db_shader_control) ||
+			   G_02880C_MASK_EXPORT_ENABLE(db_shader_control) ||
+			   G_02880C_COVERAGE_TO_MASK_ENABLE(db_shader_control) ||
+			   blend->alpha_to_coverage;
+
+	/* This is ported from Vulkan, but it doesn't make much sense to me.
+	 * Maybe it's for RE-Z? But Vulkan doesn't use RE-Z. TODO: Clarify this.
+	 */
+	bool ps_can_reject_z_trivially =
+		!G_02880C_Z_EXPORT_ENABLE(db_shader_control) ||
+		G_02880C_CONSERVATIVE_Z_EXPORT(db_shader_control);
+
+	/* Disable binning if PS can kill trivially with DB writes.
+	 * Ported from Vulkan. (heuristic?)
+	 */
+	if (ps_can_kill &&
+	    ps_can_reject_z_trivially &&
+	    sctx->framebuffer.state.zsbuf &&
+	    dsa->db_can_write) {
+		si_emit_dpbb_disable(sctx);
+		return;
+	}
+
+	/* Compute the bin size. */
+	/* TODO: We could also look at enabled pixel shader outputs. */
+	unsigned cb_target_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit &
+					  blend->cb_target_enabled_4bit;
+	struct uvec2 color_bin_size =
+		si_get_color_bin_size(sctx, cb_target_enabled_4bit);
+	struct uvec2 depth_bin_size = si_get_depth_bin_size(sctx);
+
+	unsigned color_area = color_bin_size.x * color_bin_size.y;
+	unsigned depth_area = depth_bin_size.x * depth_bin_size.y;
+
+	struct uvec2 bin_size = color_area < depth_area ? color_bin_size
+							: depth_bin_size;
+
+	if (!bin_size.x || !bin_size.y) {
+		si_emit_dpbb_disable(sctx);
+		return;
+	}
+
+	/* Enable DFSM if it's preferred. */
+	unsigned punchout_mode = V_028060_FORCE_OFF;
+	bool disable_start_of_prim = true;
+
+	if (sscreen->dfsm_allowed &&
+	    cb_target_enabled_4bit &&
+	    !G_02880C_KILL_ENABLE(db_shader_control) &&
+	    /* These two also imply that DFSM is disabled when PS writes to memory. */
+	    !G_02880C_EXEC_ON_HIER_FAIL(db_shader_control) &&
+	    !G_02880C_EXEC_ON_NOOP(db_shader_control) &&
+	    G_02880C_Z_ORDER(db_shader_control) == V_02880C_EARLY_Z_THEN_LATE_Z) {
+		punchout_mode = V_028060_AUTO;
+		disable_start_of_prim = (cb_target_enabled_4bit &
+					 blend->blend_enable_4bit) != 0;
+	}
+
+	/* Tunable parameters. Also test with DFSM enabled/disabled. */
+	unsigned context_states_per_bin; /* allowed range: [0, 5] */
+	unsigned persistent_states_per_bin; /* allowed range: [0, 31] */
+	unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
+
+	switch (sctx->b.family) {
+	case CHIP_VEGA10:
+	case CHIP_RAVEN:
+		/* Tuned for Raven. Vega might need different values. */
+		context_states_per_bin = 5;
+		persistent_states_per_bin = 31;
+		fpovs_per_batch = 63;
+		break;
+	default:
+		assert(0);
+	}
+
+	/* Emit registers. */
+	struct uvec2 bin_size_extend = {};
+	if (bin_size.x >= 32)
+		bin_size_extend.x = util_logbase2(bin_size.x) - 5;
+	if (bin_size.y >= 32)
+		bin_size_extend.y = util_logbase2(bin_size.y) - 5;
+
+	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+	radeon_set_context_reg(cs, R_028C44_PA_SC_BINNER_CNTL_0,
+			       S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
+			       S_028C44_BIN_SIZE_X(bin_size.x == 16) |
+			       S_028C44_BIN_SIZE_Y(bin_size.y == 16) |
+			       S_028C44_BIN_SIZE_X_EXTEND(bin_size_extend.x) |
+			       S_028C44_BIN_SIZE_Y_EXTEND(bin_size_extend.y) |
+			       S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin) |
+			       S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin) |
+			       S_028C44_DISABLE_START_OF_PRIM(disable_start_of_prim) |
+			       S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) |
+			       S_028C44_OPTIMAL_BIN_SELECTION(1));
+	radeon_set_context_reg(cs, R_028060_DB_DFSM_CONTROL,
+			       S_028060_PUNCHOUT_MODE(punchout_mode));
+}
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 061b3d2..4f04bbd 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -3278,20 +3278,22 @@ bool si_update_shaders(struct si_context *sctx)
 		if (sctx->screen->b.rbplus_allowed &&
 		    si_pm4_state_changed(sctx, ps) &&
 		    (!old_ps ||
 		     old_spi_shader_col_format !=
 		     sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format))
 			si_mark_atom_dirty(sctx, &sctx->cb_render_state);
 
 		if (sctx->ps_db_shader_control != db_shader_control) {
 			sctx->ps_db_shader_control = db_shader_control;
 			si_mark_atom_dirty(sctx, &sctx->db_render_state);
+			if (sctx->screen->dpbb_allowed)
+				si_mark_atom_dirty(sctx, &sctx->dpbb_state);
 		}
 
 		if (sctx->smoothing_enabled != sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing) {
 			sctx->smoothing_enabled = sctx->ps_shader.current->key.part.ps.epilog.poly_line_smoothing;
 			si_mark_atom_dirty(sctx, &sctx->msaa_config);
 
 			if (sctx->b.chip_class == SI)
 				si_mark_atom_dirty(sctx, &sctx->db_render_state);
 
 			if (sctx->framebuffer.nr_samples <= 1)
-- 
2.7.4



More information about the mesa-dev mailing list