[Mesa-dev] [PATCH 3/5] radeonsi: enable out-of-order rasterization when possible on VI and GFX9 dGPUs

Nicolai Hähnle nhaehnle at gmail.com
Sat Sep 9 10:43:56 UTC 2017


From: Nicolai Hähnle <nicolai.haehnle at amd.com>

This does not take commutative blending into account yet.

R600_DEBUG=nooutoforder disables it.
---
 src/gallium/drivers/radeon/r600_pipe_common.c   |   1 +
 src/gallium/drivers/radeon/r600_pipe_common.h   |   2 +-
 src/gallium/drivers/radeonsi/si_pipe.c          |   3 +
 src/gallium/drivers/radeonsi/si_pipe.h          |   1 +
 src/gallium/drivers/radeonsi/si_state.c         | 157 +++++++++++++++++++++++-
 src/gallium/drivers/radeonsi/si_state.h         |  28 ++++-
 src/gallium/drivers/radeonsi/si_state_shaders.c |   7 ++
 7 files changed, 193 insertions(+), 6 deletions(-)

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 1302e112c03..64851c615b0 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -818,20 +818,21 @@ static const struct debug_named_value common_debug_options[] = {
 	{ "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." },
+	{ "nooutoforder", DBG_NO_OUT_OF_ORDER, "Disable out-of-order rasterization" },
 
 	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 ed93d99669f..6074f4d440e 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -103,21 +103,21 @@ struct u_log_context;
 #define DBG_FORCE_DMA		(1ull << 38)
 #define DBG_PRECOMPILE		(1ull << 39)
 #define DBG_INFO		(1ull << 40)
 #define DBG_NO_WC		(1ull << 41)
 #define DBG_CHECK_VM		(1ull << 42)
 #define DBG_NO_DCC		(1ull << 43)
 #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_NO_OUT_OF_ORDER	(1ull << 48)
 #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
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index ca2e055a90e..9f3651f2526 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -1038,20 +1038,23 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 		 sscreen->b.info.pfp_fw_version >= 121 &&
 		 sscreen->b.info.me_fw_version >= 87) ||
 		(sscreen->b.chip_class == CIK &&
 		 sscreen->b.info.pfp_fw_version >= 211 &&
 		 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_out_of_order_rast = sscreen->b.chip_class >= VI &&
+					 sscreen->b.info.max_se >= 2 &&
+					 !(sscreen->b.debug_flags & DBG_NO_OUT_OF_ORDER);
 	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
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 8db7028c9a1..b8073ce9c09 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -88,20 +88,21 @@ struct hash_table;
 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_out_of_order_rast;
 	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;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 6978c6ca9a2..06f86aaf92a 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -416,20 +416,21 @@ static void *si_create_blend_state_mode(struct pipe_context *ctx,
 	struct si_pm4_state *pm4 = &blend->pm4;
 	uint32_t sx_mrt_blend_opt[8] = {0};
 	uint32_t color_control = 0;
 
 	if (!blend)
 		return NULL;
 
 	blend->alpha_to_coverage = state->alpha_to_coverage;
 	blend->alpha_to_one = state->alpha_to_one;
 	blend->dual_src_blend = util_blend_state_is_dual(state, 0);
+	blend->logicop_enable = state->logicop_enable;
 
 	if (state->logicop_enable) {
 		color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
 	} else {
 		color_control |= S_028808_ROP3(0xcc);
 	}
 
 	si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
 		       S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
 		       S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
@@ -623,20 +624,27 @@ static void si_bind_blend_state(struct pipe_context *ctx, void *state)
 	    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);
+
+	if (sctx->screen->has_out_of_order_rast &&
+	    (!old_blend ||
+	     (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
+	      old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
+	      old_blend->logicop_enable != blend->logicop_enable)))
+		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 }
 
 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)
@@ -1052,20 +1060,44 @@ static uint32_t si_translate_stencil_op(int s_op)
 }
 
 static bool si_dsa_writes_stencil(const struct pipe_stencil_state *s)
 {
 	return s->enabled && s->writemask &&
 	       (s->fail_op  != PIPE_STENCIL_OP_KEEP ||
 		s->zfail_op != PIPE_STENCIL_OP_KEEP ||
 		s->zpass_op != PIPE_STENCIL_OP_KEEP);
 }
 
+static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
+{
+	/* REPLACE is normally order invariant, except when the stencil
+	 * reference value is written by the fragment shader. Tracking this
+	 * interaction does not seem worth the effort, so be conservative. */
+	return op != PIPE_STENCIL_OP_INCR &&
+	       op != PIPE_STENCIL_OP_DECR &&
+	       op != PIPE_STENCIL_OP_REPLACE;
+}
+
+/* Compute whether, assuming Z writes are disabled, this stencil state is order
+ * invariant in the sense that the set of passing fragments as well as the
+ * final stencil buffer result does not depend on the order of fragments. */
+static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
+{
+	return !state->enabled || !state->writemask ||
+	       /* The following assumes that Z writes are disabled. */
+	       (state->func == PIPE_FUNC_ALWAYS &&
+	        si_order_invariant_stencil_op(state->zpass_op) &&
+	        si_order_invariant_stencil_op(state->zfail_op)) ||
+	       (state->func == PIPE_FUNC_NEVER &&
+	        si_order_invariant_stencil_op(state->fail_op));
+}
+
 static void *si_create_dsa_state(struct pipe_context *ctx,
 				 const struct pipe_depth_stencil_alpha_state *state)
 {
 	struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
 	struct si_pm4_state *pm4 = &dsa->pm4;
 	unsigned db_depth_control;
 	uint32_t db_stencil_control = 0;
 
 	if (!dsa) {
 		return NULL;
@@ -1118,20 +1150,58 @@ static void *si_create_dsa_state(struct pipe_context *ctx,
 
 	dsa->depth_enabled = state->depth.enabled;
 	dsa->depth_write_enabled = state->depth.enabled &&
 				   state->depth.writemask;
 	dsa->stencil_enabled = state->stencil[0].enabled;
 	dsa->stencil_write_enabled = state->stencil[0].enabled &&
 				     (si_dsa_writes_stencil(&state->stencil[0]) ||
 				      si_dsa_writes_stencil(&state->stencil[1]));
 	dsa->db_can_write = dsa->depth_write_enabled ||
 			    dsa->stencil_write_enabled;
+
+	bool zfunc_is_ordered =
+		state->depth.func == PIPE_FUNC_NEVER ||
+		state->depth.func == PIPE_FUNC_LESS ||
+		state->depth.func == PIPE_FUNC_LEQUAL ||
+		state->depth.func == PIPE_FUNC_GREATER ||
+		state->depth.func == PIPE_FUNC_GEQUAL;
+
+	bool nozwrite_and_order_invariant_stencil =
+		!dsa->db_can_write ||
+		(!dsa->depth_write_enabled &&
+		 si_order_invariant_stencil_state(&state->stencil[0]) &&
+		 si_order_invariant_stencil_state(&state->stencil[1]));
+
+	dsa->order_invariance[1].zs =
+		nozwrite_and_order_invariant_stencil ||
+		(!dsa->stencil_write_enabled && zfunc_is_ordered);
+	dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
+
+	dsa->order_invariance[1].pass_set =
+		nozwrite_and_order_invariant_stencil ||
+		(!dsa->stencil_write_enabled &&
+		 (state->depth.func == PIPE_FUNC_ALWAYS ||
+		  state->depth.func == PIPE_FUNC_NEVER));
+	dsa->order_invariance[0].pass_set =
+		!dsa->depth_write_enabled ||
+		(state->depth.func == PIPE_FUNC_ALWAYS ||
+		 state->depth.func == PIPE_FUNC_NEVER);
+
+	const bool assume_no_z_fights = false;
+
+	dsa->order_invariance[1].pass_last =
+		assume_no_z_fights && !dsa->stencil_write_enabled &&
+		dsa->depth_write_enabled && zfunc_is_ordered;
+	dsa->order_invariance[0].pass_last =
+		assume_no_z_fights &&
+		dsa->depth_write_enabled && zfunc_is_ordered;
+
 	return dsa;
 }
 
 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
 {
         struct si_context *sctx = (struct si_context *)ctx;
 	struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
         struct si_state_dsa *dsa = state;
 
         if (!state)
@@ -1147,20 +1217,26 @@ static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
 
 	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);
+
+	if (sctx->screen->has_out_of_order_rast &&
+	    (!old_dsa ||
+	     memcmp(old_dsa->order_invariance, dsa->order_invariance,
+		    sizeof(old_dsa->order_invariance))))
+		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 }
 
 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)
 {
@@ -1191,20 +1267,25 @@ static void si_set_active_query_state(struct pipe_context *ctx, boolean enable)
 	}
 }
 
 static void si_set_occlusion_query_state(struct pipe_context *ctx,
 					 bool old_enable,
 					 bool old_perfect_enable)
 {
 	struct si_context *sctx = (struct si_context*)ctx;
 
 	si_mark_atom_dirty(sctx, &sctx->db_render_state);
+
+	bool perfect_enable = sctx->b.num_perfect_occlusion_queries != 0;
+
+	if (perfect_enable != old_perfect_enable)
+		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 }
 
 static void si_save_qbo_state(struct pipe_context *ctx, struct r600_qbo_state *st)
 {
 	struct si_context *sctx = (struct si_context*)ctx;
 
 	st->saved_compute = sctx->cs_shader_state.program;
 
 	si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
 	si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
@@ -2534,20 +2615,25 @@ static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *sta
 
 static void si_set_framebuffer_state(struct pipe_context *ctx,
 				     const struct pipe_framebuffer_state *state)
 {
 	struct si_context *sctx = (struct si_context *)ctx;
 	struct pipe_constant_buffer constbuf = {0};
 	struct r600_surface *surf = NULL;
 	struct r600_texture *rtex;
 	bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
 	unsigned old_nr_samples = sctx->framebuffer.nr_samples;
+	unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
+	bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
+	bool old_has_stencil =
+		old_has_zsbuf &&
+		((struct r600_texture*)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
 	bool unbound = false;
 	int i;
 
 	si_update_fb_dirtiness_after_rendering(sctx);
 
 	for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
 		if (!sctx->framebuffer.state.cbufs[i])
 			continue;
 
 		rtex = (struct r600_texture*)sctx->framebuffer.state.cbufs[i]->texture;
@@ -2691,44 +2777,52 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
 
 		p_atomic_inc(&rtex->framebuffers_bound);
 
 		if (rtex->dcc_gather_statistics) {
 			/* Dirty tracking must be enabled for DCC usage analysis. */
 			sctx->framebuffer.compressed_cb_mask |= 1 << i;
 			vi_separate_dcc_start_query(ctx, rtex);
 		}
 	}
 
+	struct r600_texture *zstex = NULL;
+
 	if (state->zsbuf) {
 		surf = (struct r600_surface*)state->zsbuf;
-		rtex = (struct r600_texture*)surf->base.texture;
+		zstex = (struct r600_texture*)surf->base.texture;
 
 		if (!surf->depth_initialized) {
 			si_init_depth_surface(sctx, surf);
 		}
 
-		if (vi_tc_compat_htile_enabled(rtex, surf->base.u.tex.level))
+		if (vi_tc_compat_htile_enabled(zstex, 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->screen->has_out_of_order_rast &&
+	    (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
+	     !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
+	     (zstex && zstex->surface.has_stencil != old_has_stencil)))
+		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:
 			constbuf.user_buffer = sctx->b.sample_locations_1x;
 			break;
 		case 2:
@@ -3051,30 +3145,89 @@ static void si_emit_msaa_sample_locs(struct si_context *sctx,
 		if (has_msaa_sample_loc_bug &&
 		    sctx->framebuffer.nr_samples > 1 &&
 		    rs && !rs->multisample_enable)
 			small_prim_filter_cntl &= C_028830_SMALL_PRIM_FILTER_ENABLE;
 
 		radeon_set_context_reg(cs, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
 				       small_prim_filter_cntl);
 	}
 }
 
+static bool si_out_of_order_rasterization(struct si_context *sctx)
+{
+	struct si_state_blend *blend = sctx->queued.named.blend;
+	struct si_state_dsa *dsa = sctx->queued.named.dsa;
+
+	if (!sctx->screen->has_out_of_order_rast)
+		return false;
+
+	unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
+
+	if (blend) {
+		colormask &= blend->cb_target_enabled_4bit;
+	} else {
+		colormask = 0;
+	}
+
+	/* Conservative: No logic op. */
+	if (colormask && blend->logicop_enable)
+		return false;
+
+	struct si_dsa_order_invariance dsa_order_invariant = {
+		.zs = true, .pass_set = true, .pass_last = false
+	};
+
+	if (sctx->framebuffer.state.zsbuf) {
+		struct r600_texture *zstex =
+			(struct r600_texture*)sctx->framebuffer.state.zsbuf->texture;
+		bool has_stencil = zstex->surface.has_stencil;
+		dsa_order_invariant = dsa->order_invariance[has_stencil];
+		if (!dsa_order_invariant.zs)
+			return false;
+
+		/* The set of PS invocations is always order invariant,
+		 * except when early Z/S tests are requested. */
+		if (sctx->ps_shader.cso &&
+		    sctx->ps_shader.cso->info.writes_memory &&
+		    sctx->ps_shader.cso->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] &&
+		    !dsa_order_invariant.pass_set)
+			return false;
+
+		if (sctx->b.num_perfect_occlusion_queries != 0 &&
+		    !dsa_order_invariant.pass_set)
+			return false;
+	}
+
+	if (!colormask)
+		return true;
+
+	bool blend_enabled = (colormask & blend->blend_enable_4bit) != 0;
+
+	if (blend_enabled)
+		return false; /* TODO */
+
+	return dsa_order_invariant.pass_last;
+}
+
 static void si_emit_msaa_config(struct si_context *sctx, struct r600_atom *atom)
 {
 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
 	unsigned num_tile_pipes = sctx->screen->b.info.num_tile_pipes;
 	/* 33% faster rendering to linear color buffers */
 	bool dst_is_linear = sctx->framebuffer.any_dst_linear;
+	bool out_of_order_rast = si_out_of_order_rasterization(sctx);
 	unsigned sc_mode_cntl_1 =
 		S_028A4C_WALK_SIZE(dst_is_linear) |
 		S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
 		S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
+		S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
+		S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
 		/* always 1: */
 		S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
 		S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
 		S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
 		S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
 		S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
 		S_028A4C_FORCE_EOV_REZ_ENABLE(1);
 
 	cayman_emit_msaa_config(cs, sctx->framebuffer.nr_samples,
 				sctx->ps_iter_samples,
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 17d210a3af2..56e597a5813 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -42,29 +42,30 @@
 #define SI_NUM_IMAGES			16
 #define SI_NUM_SHADER_BUFFERS		16
 
 struct si_screen;
 struct si_shader;
 struct si_shader_selector;
 
 struct si_state_blend {
 	struct si_pm4_state	pm4;
 	uint32_t		cb_target_mask;
-	bool			alpha_to_coverage;
-	bool			alpha_to_one;
-	bool			dual_src_blend;
 	/* Set 0xf or 0x0 (4 bits) per render target if the following is
 	 * true. ANDed with spi_shader_col_format.
 	 */
 	unsigned		cb_target_enabled_4bit;
 	unsigned		blend_enable_4bit;
 	unsigned		need_src_alpha_4bit;
+	bool			alpha_to_coverage:1;
+	bool			alpha_to_one:1;
+	bool			dual_src_blend:1;
+	bool			logicop_enable:1;
 };
 
 struct si_state_rasterizer {
 	struct si_pm4_state	pm4;
 	/* poly offset states for 16-bit, 24-bit, and 32-bit zbuffers */
 	struct si_pm4_state	*pm4_poly_offset;
 	unsigned		pa_sc_line_stipple;
 	unsigned		pa_cl_clip_cntl;
 	unsigned		sprite_coord_enable:8;
 	unsigned		clip_plane_enable:8;
@@ -82,29 +83,50 @@ struct si_state_rasterizer {
 	unsigned		rasterizer_discard:1;
 	unsigned		scissor_enable:1;
 	unsigned		clip_halfz:1;
 };
 
 struct si_dsa_stencil_ref_part {
 	uint8_t			valuemask[2];
 	uint8_t			writemask[2];
 };
 
+struct si_dsa_order_invariance {
+	/** Whether the final result in Z/S buffers is guaranteed to be
+	 * invariant under changes to the order in which fragments arrive. */
+	bool zs:1;
+
+	/** Whether the set of fragments that pass the combined Z/S test is
+	 * guaranteed to be invariant under changes to the order in which
+	 * fragments arrive. */
+	bool pass_set:1;
+
+	/** Whether the last fragment that passes the combined Z/S test at each
+	 * sample is guaranteed to be invariant under changes to the order in
+	 * which fragments arrive. */
+	bool pass_last:1;
+};
+
 struct si_state_dsa {
 	struct si_pm4_state		pm4;
 	struct si_dsa_stencil_ref_part	stencil_ref;
+
+	/* 0 = without stencil buffer, 1 = when both Z and S buffers are present */
+	struct si_dsa_order_invariance	order_invariance[2];
+
 	ubyte				alpha_func:3;
 	bool				depth_enabled:1;
 	bool				depth_write_enabled:1;
 	bool				stencil_enabled:1;
 	bool				stencil_write_enabled:1;
 	bool				db_can_write:1;
+
 };
 
 struct si_stencil_ref {
 	struct r600_atom		atom;
 	struct pipe_stencil_ref		state;
 	struct si_dsa_stencil_ref_part	dsa_part;
 };
 
 struct si_vertex_elements
 {
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 9f76551cfbb..bed24798d8d 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -2402,20 +2402,27 @@ static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
 	sctx->ps_shader.current = sel ? sel->first_variant : NULL;
 
 	si_update_common_shader_state(sctx);
 	if (sel) {
 		if (sctx->ia_multi_vgt_param_key.u.uses_tess)
 			si_update_tess_uses_prim_id(sctx);
 
 		if (!old_sel ||
 		    old_sel->info.colors_written != sel->info.colors_written)
 			si_mark_atom_dirty(sctx, &sctx->cb_render_state);
+
+		if (sctx->screen->has_out_of_order_rast &&
+		    (!old_sel ||
+		     old_sel->info.writes_memory != sel->info.writes_memory ||
+		     old_sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] !=
+		     sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]))
+			si_mark_atom_dirty(sctx, &sctx->msaa_config);
 	}
 	si_set_active_descriptors_for_shader(sctx, sel);
 }
 
 static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
 {
 	if (shader->is_optimized) {
 		util_queue_drop_job(&sctx->screen->shader_compiler_queue_low_priority,
 				    &shader->optimized_ready);
 		util_queue_fence_destroy(&shader->optimized_ready);
-- 
2.11.0



More information about the mesa-dev mailing list