radeonsi: optimizing SET_CONTEXT_REG for shaders ES

Signed-off-by: Sonny Jiang <sonny.jiang@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

src/gallium/drivers/radeonsi/si_gfx_cs.c

@@ -349,6 +349,7 @@ void si_begin_new_gfx_cs(struct si_context *ctx)
 		ctx->tracked_regs.reg_value[SI_TRACKED_PA_CL_GB_HORZ_CLIP_ADJ]	= 0x3f800000;
 		ctx->tracked_regs.reg_value[SI_TRACKED_PA_CL_GB_HORZ_DISC_ADJ]	= 0x3f800000;
 		ctx->tracked_regs.reg_value[SI_TRACKED_PA_SC_CLIPRECT_RULE]	= 0xffff;
+		ctx->tracked_regs.reg_value[SI_TRACKED_VGT_ESGS_RING_ITEMSIZE]  = 0x00000000;
 
 		/* Set all saved registers state to saved. */
 		ctx->tracked_regs.reg_saved = 0xffffffff;

src/gallium/drivers/radeonsi/si_pm4.c

@@ -144,6 +144,9 @@ void si_pm4_emit(struct si_context *sctx, struct si_pm4_state *state)
 		radeon_emit(cs, ib->gpu_address >> 32);
 		radeon_emit(cs, (ib->b.b.width0 >> 2) & 0xfffff);
 	}
+
+	if (state->atom.emit)
+		state->atom.emit(sctx);
 }
 
 void si_pm4_reset_emitted(struct si_context *sctx)

src/gallium/drivers/radeonsi/si_pm4.h

@@ -33,6 +33,13 @@
 // forward defines
 struct si_context;
 
+/* State atoms are callbacks which write a sequence of packets into a GPU
+ * command buffer (AKA indirect buffer, AKA IB, AKA command stream, AKA CS).
+ */
+struct si_atom {
+	void (*emit)(struct si_context *ctx);
+};
+
 struct si_pm4_state
 {
 	/* optional indirect buffer */
@@ -52,6 +59,10 @@ struct si_pm4_state
 	struct r600_resource	*bo[SI_PM4_MAX_BO];
 	enum radeon_bo_usage	bo_usage[SI_PM4_MAX_BO];
 	enum radeon_bo_priority	bo_priority[SI_PM4_MAX_BO];
+
+	/* For shader states only */
+	struct si_shader *shader;
+	struct si_atom atom;
 };
 
 void si_pm4_cmd_begin(struct si_pm4_state *state, unsigned opcode);

src/gallium/drivers/radeonsi/si_state.h

@@ -45,13 +45,6 @@ struct si_shader_selector;
 struct si_texture;
 struct si_qbo_state;
 
-/* State atoms are callbacks which write a sequence of packets into a GPU
- * command buffer (AKA indirect buffer, AKA IB, AKA command stream, AKA CS).
- */
-struct si_atom {
-	void (*emit)(struct si_context *ctx);
-};
-
 struct si_state_blend {
 	struct si_pm4_state	pm4;
 	uint32_t		cb_target_mask;
@@ -284,6 +277,8 @@ enum si_tracked_reg {
 
 	SI_TRACKED_PA_SC_CLIPRECT_RULE,
 
+	SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+
 	SI_NUM_TRACKED_REGS,
 };
 

src/gallium/drivers/radeonsi/si_state_shaders.c

@@ -452,7 +452,13 @@ static struct si_pm4_state *si_get_shader_pm4_state(struct si_shader *shader)
 	else
 		shader->pm4 = CALLOC_STRUCT(si_pm4_state);
 
-	return shader->pm4;
+	if (shader->pm4) {
+		shader->pm4->shader = shader;
+		return shader->pm4;
+	} else {
+		fprintf(stderr, "radeonsi: Failed to create pm4 state.\n");
+		return NULL;
+	}
 }
 
 static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs)
@@ -552,6 +558,18 @@ static void si_shader_hs(struct si_screen *sscreen, struct si_shader *shader)
 	}
 }
 
+static void si_emit_shader_es(struct si_context *sctx)
+{
+	struct si_shader *shader = sctx->queued.named.es->shader;
+
+	if (!shader)
+		return;
+
+	radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+				   SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+				   shader->selector->esgs_itemsize / 4);
+}
+
 static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 {
 	struct si_pm4_state *pm4;
@@ -566,6 +584,7 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 	if (!pm4)
 		return;
 
+	pm4->atom.emit = si_emit_shader_es;
 	va = shader->bo->gpu_address;
 	si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
 
@@ -581,8 +600,6 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
 
 	oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
 
-	si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
-		       shader->selector->esgs_itemsize / 4);
 	si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
 	si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, S_00B324_MEM_BASE(va >> 40));
 	si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,