iris: Don't flush the render cache for a compute batch

Make sure we comply with BSpec and ensure that certain flush flags
are not set for compute batches

Signed-off-by: Rohan Garg's avatarRohan Garg <rohan.garg@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15664>

src/gallium/drivers/iris/iris_context.h

@@ -379,6 +379,18 @@ enum pipe_control_flags
    (PIPE_CONTROL_L3_READ_ONLY_CACHE_INVALIDATE | \
     PIPE_CONTROL_CONST_CACHE_INVALIDATE)
 
+#define PIPE_CONTROL_GRAPHICS_BITS \
+   (PIPE_CONTROL_RENDER_TARGET_FLUSH |          \
+    PIPE_CONTROL_DEPTH_CACHE_FLUSH |            \
+    PIPE_CONTROL_TILE_CACHE_FLUSH |             \
+    PIPE_CONTROL_DEPTH_STALL |                  \
+    PIPE_CONTROL_STALL_AT_SCOREBOARD |          \
+    PIPE_CONTROL_PSS_STALL_SYNC |               \
+    PIPE_CONTROL_VF_CACHE_INVALIDATE |          \
+    PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET |  \
+    PIPE_CONTROL_L3_READ_ONLY_CACHE_INVALIDATE |\
+    PIPE_CONTROL_WRITE_DEPTH_COUNT)
+
 enum iris_predicate_state {
    /* The first two states are used if we can determine whether to draw
     * without having to look at the values in the query object buffer. This

src/gallium/drivers/iris/iris_fine_fence.c

@@ -69,6 +69,9 @@ iris_fine_fence_new(struct iris_batch *batch, unsigned flags)
            PIPE_CONTROL_TILE_CACHE_FLUSH |
            PIPE_CONTROL_DEPTH_CACHE_FLUSH |
            PIPE_CONTROL_DATA_CACHE_FLUSH;
+
+      if (batch->name == IRIS_BATCH_COMPUTE)
+         pc &= ~PIPE_CONTROL_GRAPHICS_BITS;
    }
    iris_emit_pipe_control_write(batch, "fence: fine", pc,
                                 iris_resource_bo(fine->ref.res),

src/gallium/drivers/iris/iris_pipe_control.c

@@ -313,20 +313,33 @@ iris_emit_buffer_barrier_for(struct iris_batch *batch,
    }
 
    if (bits) {
+      /* Stall-at-scoreboard is not supported by the compute pipeline, use the
+       * documented sequence of two PIPE_CONTROLs with PIPE_CONTROL_FLUSH_ENABLE
+       * set in the second PIPE_CONTROL in order to obtain a similar effect.
+       */
+      const bool compute_stall_sequence = batch->name == IRIS_BATCH_COMPUTE &&
+         (bits & PIPE_CONTROL_STALL_AT_SCOREBOARD) &&
+         !(bits & PIPE_CONTROL_CACHE_FLUSH_BITS);
+
       /* Stall-at-scoreboard is not expected to work in combination with other
        * flush bits.
        */
       if (bits & PIPE_CONTROL_CACHE_FLUSH_BITS)
          bits &= ~PIPE_CONTROL_STALL_AT_SCOREBOARD;
 
+      if (batch->name == IRIS_BATCH_COMPUTE)
+         bits &= ~PIPE_CONTROL_GRAPHICS_BITS;
+
       /* Emit any required flushes and invalidations. */
-      if (bits & all_flush_bits)
+      if ((bits & all_flush_bits) || compute_stall_sequence)
          iris_emit_end_of_pipe_sync(batch, "cache tracker: flush",
                                     bits & all_flush_bits);
 
-      if (bits & ~all_flush_bits)
+      if ((bits & ~all_flush_bits) || compute_stall_sequence)
          iris_emit_pipe_control_flush(batch, "cache tracker: invalidate",
-                                      bits & ~all_flush_bits);
+                                      (bits & ~all_flush_bits) |
+                                      (compute_stall_sequence ?
+                                       PIPE_CONTROL_FLUSH_ENABLE : 0));
    }
 }
 
@@ -403,9 +416,14 @@ iris_memory_barrier(struct pipe_context *ctx, unsigned flags)
    }
 
    iris_foreach_batch(ice, batch) {
+      const unsigned allowed_bits =
+         batch->name == IRIS_BATCH_COMPUTE ? ~PIPE_CONTROL_GRAPHICS_BITS : ~0u;
+
       if (batch->contains_draw) {
          iris_batch_maybe_flush(batch, 24);
-         iris_emit_pipe_control_flush(batch, "API: memory barrier", bits);
+         iris_emit_pipe_control_flush(batch,
+                                      "API: memory barrier",
+                                      bits & allowed_bits);
       }
    }
 }

src/gallium/drivers/iris/iris_query.c

@@ -173,10 +173,21 @@ write_value(struct iris_context *ice, struct iris_query *q, unsigned offset)
    struct iris_bo *bo = iris_resource_bo(q->query_state_ref.res);
 
    if (!iris_is_query_pipelined(q)) {
+      enum pipe_control_flags flags = PIPE_CONTROL_CS_STALL |
+                                      PIPE_CONTROL_STALL_AT_SCOREBOARD;
+      if (batch->name == IRIS_BATCH_COMPUTE) {
+         iris_emit_pipe_control_write(batch,
+                                      "query: write immediate for compute batches",
+                                      PIPE_CONTROL_WRITE_IMMEDIATE,
+                                      bo,
+                                      offset,
+                                      0ull);
+         flags = PIPE_CONTROL_FLUSH_ENABLE;
+      }
+
       iris_emit_pipe_control_flush(batch,
                                    "query: non-pipelined snapshot write",
-                                   PIPE_CONTROL_CS_STALL |
-                                   PIPE_CONTROL_STALL_AT_SCOREBOARD);
+                                   flags);
       q->stalled = true;
    }
 

src/gallium/drivers/iris/iris_state.c

@@ -654,7 +654,7 @@ emit_pipeline_select(struct iris_batch *batch, uint32_t pipeline)
    enum pipe_control_flags flags = PIPE_CONTROL_CS_STALL |
                                    PIPE_CONTROL_FLUSH_HDC;
 
-   if (pipeline == GPGPU) {
+   if (pipeline == GPGPU && batch->name == IRIS_BATCH_RENDER) {
       flags |= PIPE_CONTROL_RENDER_TARGET_FLUSH |
                PIPE_CONTROL_DEPTH_CACHE_FLUSH;
    }