From: Rob Clark <robdclark@chromium.org>
Date: Thu, 1 Aug 2019 21:22:46 +0000 (-0700)
Subject: freedreno/ir3: track # of driver params
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=4a188e421522f10af85e5bb8c42c60666d02e778;p=mesa.git

freedreno/ir3: track # of driver params

To avoid emitting unneeded const state.

Signed-off-by: Rob Clark <robdclark@chromium.org>
---

diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 260702ba3c7..f258c46f94b 100644
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -335,6 +335,24 @@ ir3_nir_scan_driver_consts(nir_shader *shader,
 					} else {
 						layout->num_ubos = shader->info.num_ubos;
 					}
+					break;
+				case nir_intrinsic_load_base_vertex:
+				case nir_intrinsic_load_first_vertex:
+					layout->num_driver_params =
+						MAX2(layout->num_driver_params, IR3_DP_VTXID_BASE + 1);
+					break;
+				case nir_intrinsic_load_user_clip_plane:
+					layout->num_driver_params =
+						MAX2(layout->num_driver_params, IR3_DP_UCP7_W + 1);
+					break;
+				case nir_intrinsic_load_num_work_groups:
+					layout->num_driver_params =
+						MAX2(layout->num_driver_params, IR3_DP_NUM_WORK_GROUPS_Z + 1);
+					break;
+				case nir_intrinsic_load_local_group_size:
+					layout->num_driver_params =
+						MAX2(layout->num_driver_params, IR3_DP_LOCAL_GROUP_SIZE_Z + 1);
+					break;
 				default:
 					break;
 				}
@@ -353,8 +371,17 @@ ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir)
 
 	ir3_nir_scan_driver_consts(nir, const_state);
 
+	if ((compiler->gpu_id < 500) &&
+			(shader->stream_output.num_outputs > 0)) {
+		const_state->num_driver_params =
+			MAX2(const_state->num_driver_params, IR3_DP_VTXCNT_MAX + 1);
+	}
+
+	/* num_driver_params is scalar, align to vec4: */
+	const_state->num_driver_params = align(const_state->num_driver_params, 4);
+
 	debug_assert((shader->ubo_state.size % 16) == 0);
-	unsigned constoff = align(shader->ubo_state.size / 16, 4);
+	unsigned constoff = align(shader->ubo_state.size / 16, 8);
 	unsigned ptrsz = ir3_pointer_size(compiler);
 
 	if (const_state->num_ubos > 0) {
@@ -374,15 +401,9 @@ ir3_setup_const_state(struct ir3_shader *shader, nir_shader *nir)
 		constoff += align(cnt, 4) / 4;
 	}
 
-	unsigned num_driver_params = 0;
-	if (shader->type == MESA_SHADER_VERTEX) {
-		num_driver_params = IR3_DP_VS_COUNT;
-	} else if (shader->type == MESA_SHADER_COMPUTE) {
-		num_driver_params = IR3_DP_CS_COUNT;
-	}
-
-	const_state->offsets.driver_param = constoff;
-	constoff += align(num_driver_params, 4) / 4;
+	if (const_state->num_driver_params > 0)
+		const_state->offsets.driver_param = constoff;
+	constoff += const_state->num_driver_params / 4;
 
 	if ((shader->type == MESA_SHADER_VERTEX) &&
 			(compiler->gpu_id < 500) &&
diff --git a/src/freedreno/ir3/ir3_shader.h b/src/freedreno/ir3/ir3_shader.h
index 77e40a71b5c..53889c7f2ed 100644
--- a/src/freedreno/ir3/ir3_shader.h
+++ b/src/freedreno/ir3/ir3_shader.h
@@ -112,6 +112,7 @@ enum ir3_driver_param {
  */
 struct ir3_const_state {
 	unsigned num_ubos;
+	unsigned num_driver_params;   /* scalar */
 
 	struct {
 		/* user const start at zero */
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_gallium.c b/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
index fb83d0e7e63..b72e6c754a6 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_gallium.c
@@ -532,6 +532,8 @@ ir3_emit_vs_driver_params(const struct ir3_shader_variant *v,
 		vertex_params_size = ARRAY_SIZE(vertex_params);
 	}
 
+	vertex_params_size = MAX2(vertex_params_size, const_state->num_driver_params);
+
 	bool needs_vtxid_base =
 		ir3_find_sysval_regid(v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) != regid(63, 0);
 
@@ -656,7 +658,7 @@ ir3_emit_cs_consts(const struct ir3_shader_variant *v, struct fd_ringbuffer *rin
 				[IR3_DP_LOCAL_GROUP_SIZE_Y] = info->block[1],
 				[IR3_DP_LOCAL_GROUP_SIZE_Z] = info->block[2],
 			};
-			uint32_t size = MIN2(ARRAY_SIZE(compute_params),
+			uint32_t size = MIN2(const_state->num_driver_params,
 					v->constlen * 4 - offset * 4);
 
 			emit_const(ctx->screen, ring, v, offset * 4, 0, size,