[Mesa-dev] [PATCH 13/13] radeonsi: emit SURFACE_SYNC last

[Marek Olšák]

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

This fixes a case where a transform feedback buffer is fed back as an index
buffer, because SURFACE_SYNC must be after VS_PARTIAL_FLUSH.
---
 src/gallium/drivers/radeonsi/si_state_draw.c | 58 +++++++++++++++++-----------
 1 file changed, 35 insertions(+), 23 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 3703e5f..cd4880b 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -369,6 +369,7 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 {
 	struct radeon_winsys_cs *cs = sctx->rings.gfx.cs;
 	uint32_t cp_coher_cntl = 0;
+	uint32_t sqc_caches = 0;
 	uint32_t compute =
 		PKT3_SHADER_TYPE_S(!!(sctx->flags & SI_CONTEXT_FLAG_COMPUTE));
 
@@ -377,10 +378,9 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 	if (sctx->chip_class == SI &&
 	    sctx->flags & BOTH_ICACHE_KCACHE &&
 	    (sctx->flags & BOTH_ICACHE_KCACHE) != BOTH_ICACHE_KCACHE) {
-		r600_write_config_reg(cs, R_008C08_SQC_CACHES,
+		sqc_caches =
 			S_008C08_INST_INVALIDATE(!!(sctx->flags & SI_CONTEXT_INV_ICACHE)) |
-			S_008C08_DATA_INVALIDATE(!!(sctx->flags & SI_CONTEXT_INV_KCACHE)));
-		cs->buf[cs->cdw-3] |= compute; /* set the compute bit in the header */
+			S_008C08_DATA_INVALIDATE(!!(sctx->flags & SI_CONTEXT_INV_KCACHE));
 	} else {
 		if (sctx->flags & SI_CONTEXT_INV_ICACHE)
 			cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
@@ -409,24 +409,6 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 				 S_0085F0_DB_DEST_BASE_ENA(1);
 	}
 
-	if (cp_coher_cntl) {
-		if (sctx->chip_class >= CIK) {
-			radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0) | compute);
-			radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
-			radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
-			radeon_emit(cs, 0xff);            /* CP_COHER_SIZE_HI */
-			radeon_emit(cs, 0);               /* CP_COHER_BASE */
-			radeon_emit(cs, 0);               /* CP_COHER_BASE_HI */
-			radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
-		} else {
-			radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0) | compute);
-			radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
-			radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
-			radeon_emit(cs, 0);               /* CP_COHER_BASE */
-			radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
-		}
-	}
-
 	if (sctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB_META) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
 		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
@@ -441,6 +423,12 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 				EVENT_WRITE_INV_L2);
         }
 
+	/* FLUSH_AND_INV events must be emitted before PS_PARTIAL_FLUSH.
+	 * Otherwise, clearing CMASK (CB meta) with CP DMA isn't reliable.
+	 *
+	 * I think the reason is that FLUSH_AND_INV is only added to a queue
+	 * and it is PS_PARTIAL_FLUSH that waits for it to complete.
+	 */
 	if (sctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
 		radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
@@ -448,12 +436,10 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
 		radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
 	}
-
 	if (sctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
 		radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
 	}
-
 	if (sctx->flags & SI_CONTEXT_VGT_FLUSH) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0) | compute);
 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
@@ -463,6 +449,32 @@ void si_emit_cache_flush(struct r600_common_context *sctx, struct r600_atom *ato
 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
 	}
 
+	/* SURFACE_SYNC must be emitted after partial flushes.
+	 * It looks like SURFACE_SYNC flushes caches immediately and doesn't
+	 * wait for any engines. This should be last.
+	 */
+	if (sqc_caches) {
+		r600_write_config_reg(cs, R_008C08_SQC_CACHES, sqc_caches);
+		cs->buf[cs->cdw-3] |= compute; /* set the compute bit in the header */
+	}
+	if (cp_coher_cntl) {
+		if (sctx->chip_class >= CIK) {
+			radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0) | compute);
+			radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
+			radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
+			radeon_emit(cs, 0xff);            /* CP_COHER_SIZE_HI */
+			radeon_emit(cs, 0);               /* CP_COHER_BASE */
+			radeon_emit(cs, 0);               /* CP_COHER_BASE_HI */
+			radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
+		} else {
+			radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0) | compute);
+			radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
+			radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
+			radeon_emit(cs, 0);               /* CP_COHER_BASE */
+			radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
+		}
+	}
+
 	sctx->flags = 0;
 }
 
-- 
2.1.0