radeonsi/gfx10: rewrite late alloc computation

- Use conservative late alloc when the number of CUs <= 6.
- Move the late alloc GS register to the GS shader state, so that it can be
  tuned for NGG culling.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

src/gallium/drivers/radeonsi/si_state.c

@@ -5565,46 +5565,46 @@ static void si_init_config(struct si_context *sctx)
 
 		/* Compute LATE_ALLOC_VS.LIMIT. */
 		unsigned num_cu_per_sh = sscreen->info.num_good_cu_per_sh;
-		unsigned late_alloc_limit; /* The limit is per SH. */
+		unsigned late_alloc_wave64 = 0; /* The limit is per SH. */
-
-		if (sctx->family == CHIP_KABINI) {
-			late_alloc_limit = 0; /* Potential hang on Kabini. */
-		} else if (num_cu_per_sh <= 4) {
-			/* Too few available compute units per SH. Disallowing
-			 * VS to run on one CU could hurt us more than late VS
-			 * allocation would help.
-			 *
-			 * 2 is the highest safe number that allows us to keep
-			 * all CUs enabled.
-			 */
-			late_alloc_limit = 2;
-		} else {
-			/* This is a good initial value, allowing 1 late_alloc
-			 * wave per SIMD on num_cu - 2.
-			 */
-			late_alloc_limit = (num_cu_per_sh - 2) * 4;
-		}
-
-		unsigned late_alloc_limit_gs = late_alloc_limit;
 		unsigned cu_mask_vs = 0xffff;
 		unsigned cu_mask_gs = 0xffff;
 
-		if (late_alloc_limit > 2) {
+		if (sctx->chip_class >= GFX10) {
-			if (sctx->chip_class >= GFX10) {
+			/* For Wave32, the hw will launch twice the number of late
-				/* CU2 & CU3 disabled because of the dual CU design */
+			 * alloc waves, so 1 == 2x wave32.
-				cu_mask_vs = 0xfff3;
+			 */
-				cu_mask_gs = 0xfff3; /* NGG only */
+			if (num_cu_per_sh <= 6) {
+				late_alloc_wave64 = num_cu_per_sh - 2;
 			} else {
-				cu_mask_vs = 0xfffe; /* 1 CU disabled */
+				late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
-			}
-		}
 
-		/* Don't use late alloc for NGG on Navi14 due to a hw bug.
+				/* CU2 & CU3 disabled because of the dual CU design */
-		 * If NGG is never used, enable all CUs.
+				/* Late alloc is not used for NGG on Navi14 due to a hw bug. */
-		 */
+				cu_mask_vs = 0xfff3;
-		if (!sscreen->use_ngg || sctx->family == CHIP_NAVI14) {
+				cu_mask_gs = sscreen->use_ngg &&
-			late_alloc_limit_gs = 0;
+					     sctx->family != CHIP_NAVI14 ? 0xfff3 : 0xffff;
-			cu_mask_gs = 0xffff;
+			}
+		} else {
+			if (sctx->family == CHIP_KABINI) {
+				late_alloc_wave64 = 0; /* Potential hang on Kabini. */
+			} else if (num_cu_per_sh <= 4) {
+				/* Too few available compute units per SH. Disallowing
+				 * VS to run on one CU could hurt us more than late VS
+				 * allocation would help.
+				 *
+				 * 2 is the highest safe number that allows us to keep
+				 * all CUs enabled.
+				 */
+				late_alloc_wave64 = 2;
+			} else {
+				/* This is a good initial value, allowing 1 late_alloc
+				 * wave per SIMD on num_cu - 2.
+				 */
+				late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
+			}
+
+			if (late_alloc_wave64 > 2)
+				cu_mask_vs = 0xfffe; /* 1 CU disabled */
 		}
 
 		/* VS can't execute on one CU if the limit is > 2. */
@@ -5612,17 +5612,11 @@ static void si_init_config(struct si_context *sctx)
 			S_00B118_CU_EN(cu_mask_vs) |
 			S_00B118_WAVE_LIMIT(0x3F));
 		si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
-			S_00B11C_LIMIT(late_alloc_limit));
+			S_00B11C_LIMIT(late_alloc_wave64));
 
 		si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
 			       S_00B21C_CU_EN(cu_mask_gs) | S_00B21C_WAVE_LIMIT(0x3F));
 
-		if (sctx->chip_class >= GFX10) {
-			si_pm4_set_reg(pm4, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
-				       S_00B204_CU_EN(0xffff) |
-				       S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_limit_gs));
-		}
-
 		si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
 			       S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
 	}

src/gallium/drivers/radeonsi/si_state_shaders.c

@@ -934,6 +934,12 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 		si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS, rsrc1);
 		si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS, rsrc2);
 
+		if (sscreen->info.chip_class >= GFX10) {
+			si_pm4_set_reg(pm4, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
+				       S_00B204_CU_EN(0xffff) |
+				       S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(0));
+		}
+
 		shader->ctx_reg.gs.vgt_gs_onchip_cntl =
 			S_028A44_ES_VERTS_PER_SUBGRP(shader->gs_info.es_verts_per_subgroup) |
 			S_028A44_GS_PRIMS_PER_SUBGRP(shader->gs_info.gs_prims_per_subgroup) |
@@ -1215,6 +1221,26 @@ static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader
 		       S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
 		       S_00B22C_LDS_SIZE(shader->config.lds_size));
 
+	/* Determine LATE_ALLOC_GS. */
+	unsigned num_cu_per_sh = sscreen->info.num_good_cu_per_sh;
+	unsigned late_alloc_wave64; /* The limit is per SH. */
+
+	/* For Wave32, the hw will launch twice the number of late
+	 * alloc waves, so 1 == 2x wave32.
+	 *
+	 * Don't use late alloc for NGG on Navi14 due to a hw bug.
+	 */
+	if (sscreen->info.family == CHIP_NAVI14)
+		late_alloc_wave64 = 0;
+	else if (num_cu_per_sh <= 6)
+		late_alloc_wave64 = num_cu_per_sh - 2; /* All CUs enabled */
+	else
+		late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
+
+	si_pm4_set_reg(pm4, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
+		       S_00B204_CU_EN(0xffff) |
+		       S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64));
+
 	nparams = MAX2(shader->info.nr_param_exports, 1);
 	shader->ctx_reg.ngg.spi_vs_out_config =
 		S_0286C4_VS_EXPORT_COUNT(nparams - 1) |