i965/miptree: Allow get_aux_isl_surf when there is no aux surface
[mesa.git] / src / mesa / drivers / dri / i965 / brw_pipe_control.c
index bd45a114f2f6e231a2694197df6b412acd267a07..05e8c05f3ecdf20ff4bceab3f3a2f58b7ae073df 100644 (file)
@@ -36,6 +36,7 @@
  *  - Stall at Pixel Scoreboard
  *  - Post-Sync Operation
  *  - Depth Stall
+ *  - DC Flush Enable
  *
  * I chose "Stall at Pixel Scoreboard" since we've used it effectively
  * in the past, but the choice is fairly arbitrary.
@@ -49,7 +50,8 @@ gen8_add_cs_stall_workaround_bits(uint32_t *flags)
                       PIPE_CONTROL_WRITE_DEPTH_COUNT |
                       PIPE_CONTROL_WRITE_TIMESTAMP |
                       PIPE_CONTROL_STALL_AT_SCOREBOARD |
-                      PIPE_CONTROL_DEPTH_STALL;
+                      PIPE_CONTROL_DEPTH_STALL |
+                      PIPE_CONTROL_DATA_CACHE_FLUSH;
 
    /* If we're doing a CS stall, and don't already have one of the
     * workaround bits set, add "Stall at Pixel Scoreboard."
@@ -72,13 +74,13 @@ gen7_cs_stall_every_four_pipe_controls(struct brw_context *brw, uint32_t flags)
    if (brw->gen == 7 && !brw->is_haswell) {
       if (flags & PIPE_CONTROL_CS_STALL) {
          /* If we're doing a CS stall, reset the counter and carry on. */
-         brw->batch.pipe_controls_since_last_cs_stall = 0;
+         brw->pipe_controls_since_last_cs_stall = 0;
          return 0;
       }
 
       /* If this is the fourth pipe control without a CS stall, do one now. */
-      if (++brw->batch.pipe_controls_since_last_cs_stall == 4) {
-         brw->batch.pipe_controls_since_last_cs_stall = 0;
+      if (++brw->pipe_controls_since_last_cs_stall == 4) {
+         brw->pipe_controls_since_last_cs_stall = 0;
          return PIPE_CONTROL_CS_STALL;
       }
    }
@@ -94,8 +96,37 @@ gen7_cs_stall_every_four_pipe_controls(struct brw_context *brw, uint32_t flags)
 void
 brw_emit_pipe_control_flush(struct brw_context *brw, uint32_t flags)
 {
+   if (brw->gen >= 6 &&
+       (flags & PIPE_CONTROL_CACHE_FLUSH_BITS) &&
+       (flags & PIPE_CONTROL_CACHE_INVALIDATE_BITS)) {
+      /* A pipe control command with flush and invalidate bits set
+       * simultaneously is an inherently racy operation on Gen6+ if the
+       * contents of the flushed caches were intended to become visible from
+       * any of the invalidated caches.  Split it in two PIPE_CONTROLs, the
+       * first one should stall the pipeline to make sure that the flushed R/W
+       * caches are coherent with memory once the specified R/O caches are
+       * invalidated.  On pre-Gen6 hardware the (implicit) R/O cache
+       * invalidation seems to happen at the bottom of the pipeline together
+       * with any write cache flush, so this shouldn't be a concern.
+       */
+      brw_emit_pipe_control_flush(brw, (flags & PIPE_CONTROL_CACHE_FLUSH_BITS) |
+                                       PIPE_CONTROL_CS_STALL);
+      flags &= ~(PIPE_CONTROL_CACHE_FLUSH_BITS | PIPE_CONTROL_CS_STALL);
+   }
+
    if (brw->gen >= 8) {
-      gen8_add_cs_stall_workaround_bits(&flags);
+      if (brw->gen == 8)
+         gen8_add_cs_stall_workaround_bits(&flags);
+
+      if (brw->gen == 9 &&
+          (flags & PIPE_CONTROL_VF_CACHE_INVALIDATE)) {
+         /* Hardware workaround: SKL
+          *
+          * Emit Pipe Control with all bits set to zero before emitting
+          * a Pipe Control with VF Cache Invalidate set.
+          */
+         brw_emit_pipe_control_flush(brw, 0);
+      }
 
       BEGIN_BATCH(6);
       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (6 - 2));
@@ -106,6 +137,17 @@ brw_emit_pipe_control_flush(struct brw_context *brw, uint32_t flags)
       OUT_BATCH(0);
       ADVANCE_BATCH();
    } else if (brw->gen >= 6) {
+      if (brw->gen == 6 &&
+          (flags & PIPE_CONTROL_RENDER_TARGET_FLUSH)) {
+         /* Hardware workaround: SNB B-Spec says:
+          *
+          *   [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush
+          *   Enable = 1, a PIPE_CONTROL with any non-zero post-sync-op is
+          *   required.
+          */
+         brw_emit_post_sync_nonzero_flush(brw);
+      }
+
       flags |= gen7_cs_stall_every_four_pipe_controls(brw, flags);
 
       BEGIN_BATCH(5);
@@ -139,7 +181,8 @@ brw_emit_pipe_control_write(struct brw_context *brw, uint32_t flags,
                             uint32_t imm_lower, uint32_t imm_upper)
 {
    if (brw->gen >= 8) {
-      gen8_add_cs_stall_workaround_bits(&flags);
+      if (brw->gen == 8)
+         gen8_add_cs_stall_workaround_bits(&flags);
 
       BEGIN_BATCH(6);
       OUT_BATCH(_3DSTATE_PIPE_CONTROL | (6 - 2));
@@ -189,10 +232,18 @@ brw_emit_pipe_control_write(struct brw_context *brw, uint32_t flags,
  * already flushed (e.g., via a preceding MI_FLUSH).
  */
 void
-intel_emit_depth_stall_flushes(struct brw_context *brw)
+brw_emit_depth_stall_flushes(struct brw_context *brw)
 {
    assert(brw->gen >= 6 && brw->gen <= 9);
 
+   /* Starting on BDW, these pipe controls are unnecessary.
+    *
+    *   WM HW will internally manage the draining pipe and flushing of the caches
+    *   when this command is issued. The PIPE_CONTROL restrictions are removed.
+    */
+   if (brw->gen >= 8)
+      return;
+
    brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
    brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_CACHE_FLUSH);
    brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
@@ -213,7 +264,7 @@ gen7_emit_vs_workaround_flush(struct brw_context *brw)
    brw_emit_pipe_control_write(brw,
                                PIPE_CONTROL_WRITE_IMMEDIATE
                                | PIPE_CONTROL_DEPTH_STALL,
-                               brw->batch.workaround_bo, 0,
+                               brw->workaround_bo, 0,
                                0, 0);
 }
 
@@ -227,7 +278,7 @@ gen7_emit_cs_stall_flush(struct brw_context *brw)
    brw_emit_pipe_control_write(brw,
                                PIPE_CONTROL_CS_STALL
                                | PIPE_CONTROL_WRITE_IMMEDIATE,
-                               brw->batch.workaround_bo, 0,
+                               brw->workaround_bo, 0,
                                0, 0);
 }
 
@@ -270,14 +321,14 @@ gen7_emit_cs_stall_flush(struct brw_context *brw)
  * really our business.  That leaves only stall at scoreboard.
  */
 void
-intel_emit_post_sync_nonzero_flush(struct brw_context *brw)
+brw_emit_post_sync_nonzero_flush(struct brw_context *brw)
 {
    brw_emit_pipe_control_flush(brw,
                                PIPE_CONTROL_CS_STALL |
                                PIPE_CONTROL_STALL_AT_SCOREBOARD);
 
    brw_emit_pipe_control_write(brw, PIPE_CONTROL_WRITE_IMMEDIATE,
-                               brw->batch.workaround_bo, 0, 0, 0);
+                               brw->workaround_bo, 0, 0, 0);
 }
 
 /* Emit a pipelined flush to either flush render and texture cache for
@@ -287,7 +338,7 @@ intel_emit_post_sync_nonzero_flush(struct brw_context *brw)
  * This is also used for the always_flush_cache driconf debug option.
  */
 void
-intel_batchbuffer_emit_mi_flush(struct brw_context *brw)
+brw_emit_mi_flush(struct brw_context *brw)
 {
    if (brw->batch.ring == BLT_RING && brw->gen >= 6) {
       BEGIN_BATCH_BLT(4);
@@ -299,33 +350,40 @@ intel_batchbuffer_emit_mi_flush(struct brw_context *brw)
    } else {
       int flags = PIPE_CONTROL_NO_WRITE | PIPE_CONTROL_RENDER_TARGET_FLUSH;
       if (brw->gen >= 6) {
-         if (brw->gen == 9) {
-            /* Hardware workaround: SKL
-             *
-             * Emit Pipe Control with all bits set to zero before emitting
-             * a Pipe Control with VF Cache Invalidate set.
-             */
-            brw_emit_pipe_control_flush(brw, 0);
-         }
-
          flags |= PIPE_CONTROL_INSTRUCTION_INVALIDATE |
                   PIPE_CONTROL_DEPTH_CACHE_FLUSH |
                   PIPE_CONTROL_VF_CACHE_INVALIDATE |
                   PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
                   PIPE_CONTROL_CS_STALL;
-
-         if (brw->gen == 6) {
-            /* Hardware workaround: SNB B-Spec says:
-             *
-             * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache
-             * Flush Enable =1, a PIPE_CONTROL with any non-zero
-             * post-sync-op is required.
-             */
-            intel_emit_post_sync_nonzero_flush(brw);
-         }
       }
       brw_emit_pipe_control_flush(brw, flags);
    }
+}
+
+int
+brw_init_pipe_control(struct brw_context *brw,
+                      const struct brw_device_info *devinfo)
+{
+   if (devinfo->gen < 6)
+      return 0;
+
+   /* We can't just use brw_state_batch to get a chunk of space for
+    * the gen6 workaround because it involves actually writing to
+    * the buffer, and the kernel doesn't let us write to the batch.
+    */
+   brw->workaround_bo = drm_intel_bo_alloc(brw->bufmgr,
+                                           "pipe_control workaround",
+                                           4096, 4096);
+   if (brw->workaround_bo == NULL)
+      return -ENOMEM;
+
+   brw->pipe_controls_since_last_cs_stall = 0;
 
-   brw_render_cache_set_clear(brw);
+   return 0;
+}
+
+void
+brw_fini_pipe_control(struct brw_context *brw)
+{
+   drm_intel_bo_unreference(brw->workaround_bo);
 }