radeonsi/gfx9: rework the gfx9 scissor bug workaround (v2)

Needed to track context rolls caused by streamout and ACQUIRE_MEM.
ACQUIRE_MEM can occur outside of draw calls.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110355

v2: squashed patches and done more rework

Cc: 19.0 <[email protected]>

8 files changed, 68 insertions, 48 deletions

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 07642246ab6..aaf5138a3a2 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -1097,6 +1097,8 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 #include "si_debug_options.h"
 	}
 
+	sscreen->has_gfx9_scissor_bug = sscreen->info.family == CHIP_VEGA10 ||
+					sscreen->info.family == CHIP_RAVEN;
 	sscreen->has_msaa_sample_loc_bug = (sscreen->info.family >= CHIP_POLARIS10 &&
 					    sscreen->info.family <= CHIP_POLARIS12) ||
 					   sscreen->info.family == CHIP_VEGA10 ||
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 7fc0319973b..1d26ca90219 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -468,6 +468,7 @@ struct si_screen {
 	bool				has_out_of_order_rast;
 	bool				assume_no_z_fights;
 	bool				commutative_blend_add;
+	bool				has_gfx9_scissor_bug;
 	bool				has_msaa_sample_loc_bug;
 	bool				has_ls_vgpr_init_bug;
 	bool				has_dcc_constant_encode;
@@ -1075,7 +1076,7 @@ struct si_context {
 	unsigned			num_resident_handles;
 	uint64_t			num_alloc_tex_transfer_bytes;
 	unsigned			last_tex_ps_draw_ratio; /* for query */
-	unsigned			context_roll_counter;
+	unsigned			context_roll;
 
 	/* Queries. */
 	/* Maintain the list of active queries for pausing between IBs. */
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 757c17f7df8..bc7e777ad73 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -256,7 +256,7 @@ static void si_emit_cb_render_state(struct si_context *sctx)
 					    sx_blend_opt_control);
 	}
 	if (initial_cdw != cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 /*
@@ -793,7 +793,7 @@ static void si_emit_clip_regs(struct si_context *sctx)
 		S_028810_CLIP_DISABLE(window_space));
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 /*
@@ -1455,7 +1455,7 @@ static void si_emit_db_render_state(struct si_context *sctx)
 				   SI_TRACKED_DB_SHADER_CONTROL, db_shader_control);
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 /*
@@ -3544,7 +3544,7 @@ static void si_emit_msaa_config(struct si_context *sctx)
 				   SI_TRACKED_PA_SC_MODE_CNTL_1, sc_mode_cntl_1);
 
 	if (initial_cdw != cs->current.cdw) {
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 
 		/* GFX9: Flush DFSM when the AA mode changes. */
 		if (sctx->screen->dfsm_allowed) {
diff --git a/src/gallium/drivers/radeonsi/si_state_binning.c b/src/gallium/drivers/radeonsi/si_state_binning.c
index 3516e561282..5c6c2e69b90 100644
--- a/src/gallium/drivers/radeonsi/si_state_binning.c
+++ b/src/gallium/drivers/radeonsi/si_state_binning.c
@@ -321,7 +321,7 @@ static void si_emit_dpbb_disable(struct si_context *sctx)
 				   S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF) |
 				   S_028060_POPS_DRAIN_PS_ON_OVERLAP(1));
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 void si_emit_dpbb_state(struct si_context *sctx)
@@ -443,5 +443,5 @@ void si_emit_dpbb_state(struct si_context *sctx)
 				   S_028060_PUNCHOUT_MODE(punchout_mode) |
 				   S_028060_POPS_DRAIN_PS_ON_OVERLAP(1));
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 80e1bc4b475..4b60679484f 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -66,7 +66,7 @@ static unsigned si_conv_pipe_prim(unsigned mode)
  * The information about LDS and other non-compile-time parameters is then
  * written to userdata SGPRs.
  */
-static bool si_emit_derived_tess_state(struct si_context *sctx,
+static void si_emit_derived_tess_state(struct si_context *sctx,
 				       const struct pipe_draw_info *info,
 				       unsigned *num_patches)
 {
@@ -110,7 +110,7 @@ static bool si_emit_derived_tess_state(struct si_context *sctx,
 	    (!has_primid_instancing_bug ||
 	     (sctx->last_tess_uses_primid == tess_uses_primid))) {
 		*num_patches = sctx->last_num_patches;
-		return false;
+		return;
 	}
 
 	sctx->last_ls = ls_current;
@@ -305,9 +305,8 @@ static bool si_emit_derived_tess_state(struct si_context *sctx,
 					       ls_hs_config);
 		}
 		sctx->last_ls_hs_config = ls_hs_config;
-		return true; /* true if the context rolls */
+		sctx->context_roll = true;
 	}
-	return false;
 }
 
 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
@@ -541,7 +540,7 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
 }
 
 /* rast_prim is the primitive type after GS. */
-static bool si_emit_rasterizer_prim_state(struct si_context *sctx)
+static void si_emit_rasterizer_prim_state(struct si_context *sctx)
 {
 	struct radeon_cmdbuf *cs = sctx->gfx_cs;
 	enum pipe_prim_type rast_prim = sctx->current_rast_prim;
@@ -549,11 +548,11 @@ static bool si_emit_rasterizer_prim_state(struct si_context *sctx)
 
 	/* Skip this if not rendering lines. */
 	if (!util_prim_is_lines(rast_prim))
-		return false;
+		return;
 
 	if (rast_prim == sctx->last_rast_prim &&
 	    rs->pa_sc_line_stipple == sctx->last_sc_line_stipple)
-		return false;
+		return;
 
 	/* For lines, reset the stipple pattern at each primitive. Otherwise,
 	 * reset the stipple pattern at each packet (line strips, line loops).
@@ -564,7 +563,7 @@ static bool si_emit_rasterizer_prim_state(struct si_context *sctx)
 
 	sctx->last_rast_prim = rast_prim;
 	sctx->last_sc_line_stipple = rs->pa_sc_line_stipple;
-	return true; /* true if the context rolls */
+	sctx->context_roll = true;
 }
 
 static void si_emit_vs_state(struct si_context *sctx,
@@ -659,6 +658,7 @@ static void si_emit_draw_registers(struct si_context *sctx,
 		radeon_set_context_reg(cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
 				       info->restart_index);
 		sctx->last_restart_index = info->restart_index;
+		sctx->context_roll = true;
 	}
 }
 
@@ -886,6 +886,11 @@ static void si_emit_surface_sync(struct si_context *sctx,
 		radeon_emit(cs, 0);               /* CP_COHER_BASE */
 		radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
 	}
+
+	/* ACQUIRE_MEM has an implicit context roll if the current context
+	 * is busy. */
+	if (sctx->has_graphics)
+		sctx->context_roll = true;
 }
 
 void si_emit_cache_flush(struct si_context *sctx)
@@ -1213,26 +1218,10 @@ static void si_emit_all_states(struct si_context *sctx, const struct pipe_draw_i
 			       unsigned skip_atom_mask)
 {
 	unsigned num_patches = 0;
-	/* Vega10/Raven scissor bug workaround. When any context register is
-	 * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
-	 * registers must be written too.
-	 */
-	bool handle_scissor_bug = (sctx->family == CHIP_VEGA10 || sctx->family == CHIP_RAVEN) &&
-				  !si_is_atom_dirty(sctx, &sctx->atoms.s.scissors);
-	bool context_roll = false; /* set correctly for GFX9 only */
 
-	context_roll |= si_emit_rasterizer_prim_state(sctx);
+	si_emit_rasterizer_prim_state(sctx);
 	if (sctx->tes_shader.cso)
-		context_roll |= si_emit_derived_tess_state(sctx, info, &num_patches);
-
-	if (handle_scissor_bug &&
-	    (info->count_from_stream_output ||
-	     sctx->dirty_atoms & si_atoms_that_always_roll_context() ||
-	     sctx->dirty_states & si_states_that_always_roll_context() ||
-	     si_prim_restart_index_changed(sctx, info)))
-		context_roll = true;
-
-	sctx->context_roll_counter = 0;
+		si_emit_derived_tess_state(sctx, info, &num_patches);
 
 	/* Emit state atoms. */
 	unsigned mask = sctx->dirty_atoms & ~skip_atom_mask;
@@ -1255,12 +1244,6 @@ static void si_emit_all_states(struct si_context *sctx, const struct pipe_draw_i
 	}
 	sctx->dirty_states = 0;
 
-	if (handle_scissor_bug &&
-	    (context_roll || sctx->context_roll_counter)) {
-		sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
-		sctx->atoms.s.scissors.emit(sctx);
-	}
-
 	/* Emit draw states. */
 	si_emit_vs_state(sctx, info);
 	si_emit_draw_registers(sctx, info, num_patches);
@@ -1462,6 +1445,22 @@ static void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *i
 	if (!si_upload_vertex_buffer_descriptors(sctx))
 		goto return_cleanup;
 
+	/* Vega10/Raven scissor bug workaround. When any context register is
+	 * written (i.e. the GPU rolls the context), PA_SC_VPORT_SCISSOR
+	 * registers must be written too.
+	 */
+	bool has_gfx9_scissor_bug = sctx->screen->has_gfx9_scissor_bug;
+	unsigned masked_atoms = 0;
+
+	if (has_gfx9_scissor_bug) {
+		masked_atoms |= si_get_atom_bit(sctx, &sctx->atoms.s.scissors);
+
+		if (info->count_from_stream_output ||
+		    sctx->dirty_atoms & si_atoms_that_always_roll_context() ||
+		    sctx->dirty_states & si_states_that_always_roll_context())
+			sctx->context_roll = true;
+	}
+
 	/* Use optimal packet order based on whether we need to sync the pipeline. */
 	if (unlikely(sctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
 				      SI_CONTEXT_FLUSH_AND_INV_DB |
@@ -1472,8 +1471,6 @@ static void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *i
 		 * Then draw and prefetch at the end. This ensures that the time
 		 * the CUs are idle is very short.
 		 */
-		unsigned masked_atoms = 0;
-
 		if (unlikely(sctx->flags & SI_CONTEXT_FLUSH_FOR_RENDER_COND))
 			masked_atoms |= si_get_atom_bit(sctx, &sctx->atoms.s.render_cond);
 
@@ -1487,6 +1484,13 @@ static void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *i
 
 		if (si_is_atom_dirty(sctx, &sctx->atoms.s.render_cond))
 			sctx->atoms.s.render_cond.emit(sctx);
+
+		if (has_gfx9_scissor_bug &&
+		    (sctx->context_roll ||
+		     si_is_atom_dirty(sctx, &sctx->atoms.s.scissors))) {
+			sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+			sctx->atoms.s.scissors.emit(sctx);
+		}
 		sctx->dirty_atoms = 0;
 
 		si_emit_draw_packets(sctx, info, indexbuf, index_size, index_offset);
@@ -1511,7 +1515,16 @@ static void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *i
 		if (!si_upload_graphics_shader_descriptors(sctx))
 			return;
 
-		si_emit_all_states(sctx, info, 0);
+		si_emit_all_states(sctx, info, masked_atoms);
+
+		if (has_gfx9_scissor_bug &&
+		    (sctx->context_roll ||
+		     si_is_atom_dirty(sctx, &sctx->atoms.s.scissors))) {
+			sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+			sctx->atoms.s.scissors.emit(sctx);
+		}
+		sctx->dirty_atoms = 0;
+
 		si_emit_draw_packets(sctx, info, indexbuf, index_size, index_offset);
 
 		/* Prefetch the remaining shaders after the draw has been
@@ -1520,6 +1533,9 @@ static void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *i
 			cik_emit_prefetch_L2(sctx, false);
 	}
 
+	/* Clear the context roll flag after the draw call. */
+	sctx->context_roll = false;
+
 	if (unlikely(sctx->current_saved_cs)) {
 		si_trace_emit(sctx);
 		si_log_draw_state(sctx, sctx->log);
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index f57e7730905..55df95477d3 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -576,7 +576,7 @@ static void si_emit_shader_es(struct si_context *sctx)
 					   shader->vgt_vertex_reuse_block_cntl);
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
@@ -825,7 +825,7 @@ static void si_emit_shader_gs(struct si_context *sctx)
 	}
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
@@ -1002,7 +1002,7 @@ static void si_emit_shader_vs(struct si_context *sctx)
 					   shader->vgt_vertex_reuse_block_cntl);
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 /**
@@ -1194,7 +1194,7 @@ static void si_emit_shader_ps(struct si_context *sctx)
 				   shader->ctx_reg.ps.cb_shader_mask);
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 static void si_shader_ps(struct si_shader *shader)
@@ -2877,7 +2877,7 @@ static void si_emit_spi_map(struct si_context *sctx)
 				    sctx->tracked_regs.spi_ps_input_cntl, num_interp);
 
 	if (initial_cdw != sctx->gfx_cs->current.cdw)
-		sctx->context_roll_counter++;
+		sctx->context_roll = true;
 }
 
 /**
diff --git a/src/gallium/drivers/radeonsi/si_state_streamout.c b/src/gallium/drivers/radeonsi/si_state_streamout.c
index 2bf6862c89b..2a0a4bef9a2 100644
--- a/src/gallium/drivers/radeonsi/si_state_streamout.c
+++ b/src/gallium/drivers/radeonsi/si_state_streamout.c
@@ -303,6 +303,7 @@ void si_emit_streamout_end(struct si_context *sctx)
 		 * buffer bound. This ensures that the primitives-emitted query
 		 * won't increment. */
 		radeon_set_context_reg(cs, R_028AD0_VGT_STRMOUT_BUFFER_SIZE_0 + 16*i, 0);
+		sctx->context_roll = true;
 
 		t[i]->buf_filled_size_valid = true;
 	}
diff --git a/src/gallium/drivers/radeonsi/si_state_viewport.c b/src/gallium/drivers/radeonsi/si_state_viewport.c
index f988da4520b..6f348a9b58d 100644
--- a/src/gallium/drivers/radeonsi/si_state_viewport.c
+++ b/src/gallium/drivers/radeonsi/si_state_viewport.c
@@ -283,7 +283,7 @@ static void si_emit_guardband(struct si_context *ctx)
 				   S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH +
 						       vp_as_scissor.quant_mode));
 	if (initial_cdw != ctx->gfx_cs->current.cdw)
-		ctx->context_roll_counter++;
+		ctx->context_roll = true;
 }
 
 static void si_emit_scissors(struct si_context *ctx)