intel: Add a batch flush between front-buffer downsample and X protocol.
[mesa.git] / src / mesa / drivers / dri / i965 / gen6_sol.c
index 491b39cce12424628cd3966e87678a8d217e0649..5c294b1d768e31dfe5c4cd2451f2e5f24783f154 100644 (file)
  * Code to initialize the binding table entries used by transform feedback.
  */
 
+#include "main/macros.h"
 #include "brw_context.h"
+#include "intel_batchbuffer.h"
 #include "brw_defines.h"
+#include "brw_state.h"
+#include "main/transformfeedback.h"
 
 static void
 gen6_update_sol_surfaces(struct brw_context *brw)
 {
-   struct gl_context *ctx = &brw->intel.ctx;
-   /* _NEW_TRANSFORM_FEEDBACK */
+   struct gl_context *ctx = &brw->ctx;
+   /* BRW_NEW_TRANSFORM_FEEDBACK */
    struct gl_transform_feedback_object *xfb_obj =
       ctx->TransformFeedback.CurrentObject;
    /* BRW_NEW_VERTEX_PROGRAM */
@@ -45,27 +49,142 @@ gen6_update_sol_surfaces(struct brw_context *brw)
 
    for (i = 0; i < BRW_MAX_SOL_BINDINGS; ++i) {
       const int surf_index = SURF_INDEX_SOL_BINDING(i);
-      if (xfb_obj->Active && i < linked_xfb_info->NumOutputs) {
+      if (_mesa_is_xfb_active_and_unpaused(ctx) &&
+          i < linked_xfb_info->NumOutputs) {
          unsigned buffer = linked_xfb_info->Outputs[i].OutputBuffer;
          unsigned buffer_offset =
             xfb_obj->Offset[buffer] / 4 +
             linked_xfb_info->Outputs[i].DstOffset;
          brw_update_sol_surface(
-            brw, xfb_obj->Buffers[buffer], &brw->bind.surf_offset[surf_index],
+            brw, xfb_obj->Buffers[buffer], &brw->gs.surf_offset[surf_index],
             linked_xfb_info->Outputs[i].NumComponents,
             linked_xfb_info->BufferStride[buffer], buffer_offset);
       } else {
-         brw->bind.surf_offset[surf_index] = 0;
+         brw->gs.surf_offset[surf_index] = 0;
       }
    }
+
+   brw->state.dirty.brw |= BRW_NEW_SURFACES;
 }
 
 const struct brw_tracked_state gen6_sol_surface = {
    .dirty = {
-      .mesa = _NEW_TRANSFORM_FEEDBACK,
+      .mesa = 0,
       .brw = (BRW_NEW_BATCH |
-              BRW_NEW_VERTEX_PROGRAM),
+              BRW_NEW_VERTEX_PROGRAM |
+              BRW_NEW_TRANSFORM_FEEDBACK),
       .cache = 0
    },
    .emit = gen6_update_sol_surfaces,
 };
+
+/**
+ * Constructs the binding table for the WM surface state, which maps unit
+ * numbers to surface state objects.
+ */
+static void
+brw_gs_upload_binding_table(struct brw_context *brw)
+{
+   struct gl_context *ctx = &brw->ctx;
+   /* BRW_NEW_VERTEX_PROGRAM */
+   const struct gl_shader_program *shaderprog =
+      ctx->Shader.CurrentVertexProgram;
+   bool has_surfaces = false;
+   uint32_t *bind;
+
+   if (shaderprog) {
+      const struct gl_transform_feedback_info *linked_xfb_info =
+        &shaderprog->LinkedTransformFeedback;
+      /* Currently we only ever upload surfaces for SOL. */
+      has_surfaces = linked_xfb_info->NumOutputs != 0;
+   }
+
+   /* Skip making a binding table if we don't have anything to put in it. */
+   if (!has_surfaces) {
+      if (brw->gs.bind_bo_offset != 0) {
+        brw->state.dirty.brw |= BRW_NEW_GS_BINDING_TABLE;
+        brw->gs.bind_bo_offset = 0;
+      }
+      return;
+   }
+
+   /* Might want to calculate nr_surfaces first, to avoid taking up so much
+    * space for the binding table.
+    */
+   bind = brw_state_batch(brw, AUB_TRACE_BINDING_TABLE,
+                         sizeof(uint32_t) * BRW_MAX_GS_SURFACES,
+                         32, &brw->gs.bind_bo_offset);
+
+   /* BRW_NEW_SURFACES */
+   memcpy(bind, brw->gs.surf_offset, BRW_MAX_GS_SURFACES * sizeof(uint32_t));
+
+   brw->state.dirty.brw |= BRW_NEW_GS_BINDING_TABLE;
+}
+
+const struct brw_tracked_state gen6_gs_binding_table = {
+   .dirty = {
+      .mesa = 0,
+      .brw = (BRW_NEW_BATCH |
+             BRW_NEW_VERTEX_PROGRAM |
+             BRW_NEW_SURFACES),
+      .cache = 0
+   },
+   .emit = brw_gs_upload_binding_table,
+};
+
+void
+brw_begin_transform_feedback(struct gl_context *ctx, GLenum mode,
+                            struct gl_transform_feedback_object *obj)
+{
+   struct brw_context *brw = brw_context(ctx);
+   const struct gl_shader_program *vs_prog =
+      ctx->Shader.CurrentVertexProgram;
+   const struct gl_transform_feedback_info *linked_xfb_info =
+      &vs_prog->LinkedTransformFeedback;
+   struct gl_transform_feedback_object *xfb_obj =
+      ctx->TransformFeedback.CurrentObject;
+
+   assert(brw->gen == 6);
+
+   /* Compute the maximum number of vertices that we can write without
+    * overflowing any of the buffers currently being used for feedback.
+    */
+   unsigned max_index
+      = _mesa_compute_max_transform_feedback_vertices(xfb_obj,
+                                                      linked_xfb_info);
+
+   /* Initialize the SVBI 0 register to zero and set the maximum index. */
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
+   OUT_BATCH(0); /* SVBI 0 */
+   OUT_BATCH(0); /* starting index */
+   OUT_BATCH(max_index);
+   ADVANCE_BATCH();
+
+   /* Initialize the rest of the unused streams to sane values.  Otherwise,
+    * they may indicate that there is no room to write data and prevent
+    * anything from happening at all.
+    */
+   for (int i = 1; i < 4; i++) {
+      BEGIN_BATCH(4);
+      OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
+      OUT_BATCH(i << SVB_INDEX_SHIFT);
+      OUT_BATCH(0); /* starting index */
+      OUT_BATCH(0xffffffff);
+      ADVANCE_BATCH();
+   }
+}
+
+void
+brw_end_transform_feedback(struct gl_context *ctx,
+                           struct gl_transform_feedback_object *obj)
+{
+   /* After EndTransformFeedback, it's likely that the client program will try
+    * to draw using the contents of the transform feedback buffer as vertex
+    * input.  In order for this to work, we need to flush the data through at
+    * least the GS stage of the pipeline, and flush out the render cache.  For
+    * simplicity, just do a full flush.
+    */
+   struct brw_context *brw = brw_context(ctx);
+   intel_batchbuffer_emit_mi_flush(brw);
+}