radeonsi: don't use r600_common_context in si_emit_cache_flush
authorMarek Olšák <marek.olsak@amd.com>
Sun, 1 Apr 2018 19:24:07 +0000 (15:24 -0400)
committerMarek Olšák <marek.olsak@amd.com>
Thu, 5 Apr 2018 19:34:58 +0000 (15:34 -0400)
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
src/gallium/drivers/radeonsi/si_state_draw.c

index f571d7fc8185acffe5787db11950be2ad61b24fb..7e86c0f07dc60cc98408d10e7f978b365e58e45d 100644 (file)
@@ -849,12 +849,12 @@ static void si_emit_draw_packets(struct si_context *sctx,
        }
 }
 
-static void si_emit_surface_sync(struct r600_common_context *rctx,
+static void si_emit_surface_sync(struct si_context *sctx,
                                 unsigned cp_coher_cntl)
 {
-       struct radeon_winsys_cs *cs = rctx->gfx_cs;
+       struct radeon_winsys_cs *cs = sctx->b.gfx_cs;
 
-       if (rctx->chip_class >= GFX9) {
+       if (sctx->b.chip_class >= GFX9) {
                /* Flush caches and wait for the caches to assert idle. */
                radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0));
                radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
@@ -875,15 +875,15 @@ static void si_emit_surface_sync(struct r600_common_context *rctx,
 
 void si_emit_cache_flush(struct si_context *sctx)
 {
-       struct r600_common_context *rctx = &sctx->b;
-       struct radeon_winsys_cs *cs = rctx->gfx_cs;
+       struct radeon_winsys_cs *cs = sctx->b.gfx_cs;
+       uint32_t flags = sctx->b.flags;
        uint32_t cp_coher_cntl = 0;
-       uint32_t flush_cb_db = rctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
-                                             SI_CONTEXT_FLUSH_AND_INV_DB);
+       uint32_t flush_cb_db = flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
+                                       SI_CONTEXT_FLUSH_AND_INV_DB);
 
-       if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB)
+       if (flags & SI_CONTEXT_FLUSH_AND_INV_CB)
                sctx->b.num_cb_cache_flushes++;
-       if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB)
+       if (flags & SI_CONTEXT_FLUSH_AND_INV_DB)
                sctx->b.num_db_cache_flushes++;
 
        /* SI has a bug that it always flushes ICACHE and KCACHE if either
@@ -894,13 +894,13 @@ void si_emit_cache_flush(struct si_context *sctx)
         * to add a workaround for it.
         */
 
-       if (rctx->flags & SI_CONTEXT_INV_ICACHE)
+       if (flags & SI_CONTEXT_INV_ICACHE)
                cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
-       if (rctx->flags & SI_CONTEXT_INV_SMEM_L1)
+       if (flags & SI_CONTEXT_INV_SMEM_L1)
                cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
 
-       if (rctx->chip_class <= VI) {
-               if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
+       if (sctx->b.chip_class <= VI) {
+               if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
                        cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
                                         S_0085F0_CB0_DEST_BASE_ENA(1) |
                                         S_0085F0_CB1_DEST_BASE_ENA(1) |
@@ -912,23 +912,23 @@ void si_emit_cache_flush(struct si_context *sctx)
                                         S_0085F0_CB7_DEST_BASE_ENA(1);
 
                        /* Necessary for DCC */
-                       if (rctx->chip_class == VI)
-                               si_gfx_write_event_eop(rctx, V_028A90_FLUSH_AND_INV_CB_DATA_TS,
-                                                        0, EOP_DATA_SEL_DISCARD, NULL,
-                                                        0, 0, SI_NOT_QUERY);
+                       if (sctx->b.chip_class == VI)
+                               si_gfx_write_event_eop(&sctx->b, V_028A90_FLUSH_AND_INV_CB_DATA_TS,
+                                                      0, EOP_DATA_SEL_DISCARD, NULL,
+                                                      0, 0, SI_NOT_QUERY);
                }
-               if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB)
+               if (flags & SI_CONTEXT_FLUSH_AND_INV_DB)
                        cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
                                         S_0085F0_DB_DEST_BASE_ENA(1);
        }
 
-       if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
+       if (flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
                /* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
        }
-       if (rctx->flags & (SI_CONTEXT_FLUSH_AND_INV_DB |
-                          SI_CONTEXT_FLUSH_AND_INV_DB_META)) {
+       if (flags & (SI_CONTEXT_FLUSH_AND_INV_DB |
+                    SI_CONTEXT_FLUSH_AND_INV_DB_META)) {
                /* Flush HTILE. SURFACE_SYNC will wait for idle. */
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
@@ -939,35 +939,35 @@ void si_emit_cache_flush(struct si_context *sctx)
         * for everything including CB/DB cache flushes.
         */
        if (!flush_cb_db) {
-               if (rctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
+               if (flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
                        radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                        radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
                        /* Only count explicit shader flushes, not implicit ones
                         * done by SURFACE_SYNC.
                         */
-                       rctx->num_vs_flushes++;
-                       rctx->num_ps_flushes++;
-               } else if (rctx->flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
+                       sctx->b.num_vs_flushes++;
+                       sctx->b.num_ps_flushes++;
+               } else if (flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
                        radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                        radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
-                       rctx->num_vs_flushes++;
+                       sctx->b.num_vs_flushes++;
                }
        }
 
-       if (rctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH &&
+       if (flags & SI_CONTEXT_CS_PARTIAL_FLUSH &&
            sctx->compute_is_busy) {
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
-               rctx->num_cs_flushes++;
+               sctx->b.num_cs_flushes++;
                sctx->compute_is_busy = false;
        }
 
        /* VGT state synchronization. */
-       if (rctx->flags & SI_CONTEXT_VGT_FLUSH) {
+       if (flags & SI_CONTEXT_VGT_FLUSH) {
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
        }
-       if (rctx->flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
+       if (flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
        }
@@ -1006,21 +1006,21 @@ void si_emit_cache_flush(struct si_context *sctx)
                 */
                tc_flags = 0;
 
-               if (rctx->flags & SI_CONTEXT_INV_L2_METADATA) {
+               if (flags & SI_CONTEXT_INV_L2_METADATA) {
                        tc_flags = EVENT_TC_ACTION_ENA |
                                   EVENT_TC_MD_ACTION_ENA;
                }
 
                /* Ideally flush TC together with CB/DB. */
-               if (rctx->flags & SI_CONTEXT_INV_GLOBAL_L2) {
+               if (flags & SI_CONTEXT_INV_GLOBAL_L2) {
                        /* Writeback and invalidate everything in L2 & L1. */
                        tc_flags = EVENT_TC_ACTION_ENA |
                                   EVENT_TC_WB_ACTION_ENA;
 
                        /* Clear the flags. */
-                       rctx->flags &= ~(SI_CONTEXT_INV_GLOBAL_L2 |
-                                        SI_CONTEXT_WRITEBACK_GLOBAL_L2 |
-                                        SI_CONTEXT_INV_VMEM_L1);
+                       flags &= ~(SI_CONTEXT_INV_GLOBAL_L2 |
+                                  SI_CONTEXT_WRITEBACK_GLOBAL_L2 |
+                                  SI_CONTEXT_INV_VMEM_L1);
                        sctx->b.num_L2_invalidates++;
                }
 
@@ -1028,18 +1028,18 @@ void si_emit_cache_flush(struct si_context *sctx)
                va = sctx->wait_mem_scratch->gpu_address;
                sctx->wait_mem_number++;
 
-               si_gfx_write_event_eop(rctx, cb_db_event, tc_flags,
-                                        EOP_DATA_SEL_VALUE_32BIT,
-                                        sctx->wait_mem_scratch, va,
-                                        sctx->wait_mem_number, SI_NOT_QUERY);
-               si_gfx_wait_fence(rctx, va, sctx->wait_mem_number, 0xffffffff);
+               si_gfx_write_event_eop(&sctx->b, cb_db_event, tc_flags,
+                                      EOP_DATA_SEL_VALUE_32BIT,
+                                      sctx->wait_mem_scratch, va,
+                                      sctx->wait_mem_number, SI_NOT_QUERY);
+               si_gfx_wait_fence(&sctx->b, va, sctx->wait_mem_number, 0xffffffff);
        }
 
        /* Make sure ME is idle (it executes most packets) before continuing.
         * This prevents read-after-write hazards between PFP and ME.
         */
        if (cp_coher_cntl ||
-           (rctx->flags & (SI_CONTEXT_CS_PARTIAL_FLUSH |
+           (flags & (SI_CONTEXT_CS_PARTIAL_FLUSH |
                            SI_CONTEXT_INV_VMEM_L1 |
                            SI_CONTEXT_INV_GLOBAL_L2 |
                            SI_CONTEXT_WRITEBACK_GLOBAL_L2))) {
@@ -1056,38 +1056,38 @@ void si_emit_cache_flush(struct si_context *sctx)
         *
         * SI-CIK don't support L2 write-back.
         */
-       if (rctx->flags & SI_CONTEXT_INV_GLOBAL_L2 ||
-           (rctx->chip_class <= CIK &&
-            (rctx->flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2))) {
+       if (flags & SI_CONTEXT_INV_GLOBAL_L2 ||
+           (sctx->b.chip_class <= CIK &&
+            (flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2))) {
                /* Invalidate L1 & L2. (L1 is always invalidated on SI)
                 * WB must be set on VI+ when TC_ACTION is set.
                 */
-               si_emit_surface_sync(rctx, cp_coher_cntl |
+               si_emit_surface_sync(sctx, cp_coher_cntl |
                                     S_0085F0_TC_ACTION_ENA(1) |
                                     S_0085F0_TCL1_ACTION_ENA(1) |
-                                    S_0301F0_TC_WB_ACTION_ENA(rctx->chip_class >= VI));
+                                    S_0301F0_TC_WB_ACTION_ENA(sctx->b.chip_class >= VI));
                cp_coher_cntl = 0;
                sctx->b.num_L2_invalidates++;
        } else {
                /* L1 invalidation and L2 writeback must be done separately,
                 * because both operations can't be done together.
                 */
-               if (rctx->flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2) {
+               if (flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2) {
                        /* WB = write-back
                         * NC = apply to non-coherent MTYPEs
                         *      (i.e. MTYPE <= 1, which is what we use everywhere)
                         *
                         * WB doesn't work without NC.
                         */
-                       si_emit_surface_sync(rctx, cp_coher_cntl |
+                       si_emit_surface_sync(sctx, cp_coher_cntl |
                                             S_0301F0_TC_WB_ACTION_ENA(1) |
                                             S_0301F0_TC_NC_ACTION_ENA(1));
                        cp_coher_cntl = 0;
                        sctx->b.num_L2_writebacks++;
                }
-               if (rctx->flags & SI_CONTEXT_INV_VMEM_L1) {
+               if (flags & SI_CONTEXT_INV_VMEM_L1) {
                        /* Invalidate per-CU VMEM L1. */
-                       si_emit_surface_sync(rctx, cp_coher_cntl |
+                       si_emit_surface_sync(sctx, cp_coher_cntl |
                                             S_0085F0_TCL1_ACTION_ENA(1));
                        cp_coher_cntl = 0;
                }
@@ -1095,19 +1095,19 @@ void si_emit_cache_flush(struct si_context *sctx)
 
        /* If TC flushes haven't cleared this... */
        if (cp_coher_cntl)
-               si_emit_surface_sync(rctx, cp_coher_cntl);
+               si_emit_surface_sync(sctx, cp_coher_cntl);
 
-       if (rctx->flags & SI_CONTEXT_START_PIPELINE_STATS) {
+       if (flags & SI_CONTEXT_START_PIPELINE_STATS) {
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
                                EVENT_INDEX(0));
-       } else if (rctx->flags & SI_CONTEXT_STOP_PIPELINE_STATS) {
+       } else if (flags & SI_CONTEXT_STOP_PIPELINE_STATS) {
                radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
                radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
                                EVENT_INDEX(0));
        }
 
-       rctx->flags = 0;
+       sctx->b.flags = 0;
 }
 
 static void si_get_draw_start_count(struct si_context *sctx,