radeonsi: rewrite late alloc VS limit computation

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

Loosely ported from Vulkan.

diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 4772df25d1f2dc4b96b79775b8a9008dc382a68a..41b08f8de4f068030168c44d209fda8be4eb60be 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -4639,26 +4639,39 @@ static void si_init_config(struct si_context *sctx)
                }
                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, allowing 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));
        }