[Mesa-dev] [PATCH 2/2] radeonsi: rewrite late alloc VS limit computation

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

This is still very simple, but it's better than before.

Loosely ported from Vulkan.
---
 src/gallium/drivers/radeonsi/si_state.c | 38 ++++++++++++++++++++++-----------
 1 file changed, 26 insertions(+), 12 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 24e509c..393f960 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -4649,40 +4649,54 @@ static void si_init_config(struct si_context *sctx)
 			/* If this is 0, Bonaire can hang even if GS isn't being used.
 			 * Other chips are unaffected. These are suboptimal values,
 			 * but we don't use on-chip GS.
 			 */
 			si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
 				       S_028A44_ES_VERTS_PER_SUBGRP(64) |
 				       S_028A44_GS_PRIMS_PER_SUBGRP(4));
 		}
 		si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS, S_00B21C_CU_EN(0xffff));
 
-		if (sscreen->b.info.num_good_compute_units /
-		    (sscreen->b.info.max_se * sscreen->b.info.max_sh_per_se) <= 4) {
+		/* Compute LATE_ALLOC_VS.LIMIT. */
+		unsigned num_cu_per_sh = sscreen->b.info.num_good_compute_units /
+					 (sscreen->b.info.max_se *
+					  sscreen->b.info.max_sh_per_se);
+		unsigned late_alloc_limit; /* The limit is per SH. */
+
+		if (sctx->b.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 CU0 could hurt us more than late VS
+			 * VS to run on one CU could hurt us more than late VS
 			 * allocation would help.
 			 *
-			 * LATE_ALLOC_VS = 2 is the highest safe number.
+			 * 2 is the highest safe number that allows us to keep
+			 * all CUs enabled.
 			 */
-			si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xffff));
-			si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
+			late_alloc_limit = 2;
 		} else {
-			/* Set LATE_ALLOC_VS == 31. It should be less than
-			 * the number of scratch waves. Limitations:
-			 * - VS can't execute on CU0.
-			 * - If HS writes outputs to LDS, LS can't execute on CU0.
+			/* This is a good initial value.
+			 * We shouldn't run into a VS-PS deadlock, because it
+			 * only allows 1 late_alloc wave per SIMD on num_cu - 2.
 			 */
-			si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xfffe));
-			si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
+			late_alloc_limit = (num_cu_per_sh - 2) * 4;
+
+			/* The limit is 0-based, so 0 means 1. */
+			assert(late_alloc_limit > 0 && late_alloc_limit <= 64);
+			late_alloc_limit -= 1;
 		}
 
+		/* VS can't execute on one CU if the limit is > 2. */
+		si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
+			       S_00B118_CU_EN(late_alloc_limit > 2 ? 0xfffe : 0xffff));
+		si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
+			       S_00B11C_LIMIT(late_alloc_limit));
 		si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS, S_00B01C_CU_EN(0xffff));
 	}
 
 	if (sctx->b.chip_class >= VI) {
 		unsigned vgt_tess_distribution;
 
 		si_pm4_set_reg(pm4, R_028424_CB_DCC_CONTROL,
 			       S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
 			       S_028424_OVERWRITE_COMBINER_WATERMARK(4));
 
-- 
2.7.4