freedreno/a6xx: Let the GPU track streamout offsets

[mesa.git] / src / gallium / drivers / freedreno / freedreno_batch.h

diff --git a/src/gallium/drivers/freedreno/freedreno_batch.h b/src/gallium/drivers/freedreno/freedreno_batch.h
index 6bb88a62916c809b5c38020ff453d542507ab1d1..edf0840825be470855be6c77bc112d1512b9aad7 100644 (file)
--- a/src/gallium/drivers/freedreno/freedreno_batch.h
+++ b/src/gallium/drivers/freedreno/freedreno_batch.h
@@ -84,6 +84,10 @@ struct fd_batch {
         * The 'cleared' bits will be set for buffers which are *entirely*
         * cleared, and 'partial_cleared' bits will be set if you must
         * check cleared_scissor.
+        *
+        * The 'invalidated' bits are set for cleared buffers, and buffers
+        * where the contents are undefined, ie. what we don't need to restore
+        * to gmem.
         */
        enum {
                /* align bitmask values w/ PIPE_CLEAR_*.. since that is convenient.. */
@@ -91,7 +95,7 @@ struct fd_batch {
                FD_BUFFER_DEPTH   = PIPE_CLEAR_DEPTH,
                FD_BUFFER_STENCIL = PIPE_CLEAR_STENCIL,
                FD_BUFFER_ALL     = FD_BUFFER_COLOR | FD_BUFFER_DEPTH | FD_BUFFER_STENCIL,
-       } cleared, partial_cleared, restore, resolve;
+       } invalidated, cleared, fast_cleared, restore, resolve;
 
        /* is this a non-draw batch (ie compute/blit which has no pfb state)? */
        bool nondraw : 1;
@@ -118,8 +122,10 @@ struct fd_batch {
 
                FD_GMEM_BLEND_ENABLED        = 0x10,
                FD_GMEM_LOGICOP_ENABLED      = 0x20,
+               FD_GMEM_FB_READ              = 0x40,
        } gmem_reason;
        unsigned num_draws;   /* number of draws in current batch */
+       unsigned num_vertices;   /* number of vertices in current batch */
 
        /* Track the maximal bounds of the scissor of all the draws within a
         * batch.  Used at the tile rendering step (fd_gmem_render_tiles(),
@@ -127,23 +133,13 @@ struct fd_batch {
         */
        struct pipe_scissor_state max_scissor;
 
-       /* Track the cleared scissor for color/depth/stencil, so we know
-        * which, if any, tiles need to be restored (mem2gmem).  Only valid
-        * if the corresponding bit in ctx->cleared is set.
-        */
-       struct {
-               struct pipe_scissor_state color, depth, stencil;
-       } cleared_scissor;
-
        /* Keep track of DRAW initiators that need to be patched up depending
         * on whether we using binning or not:
         */
        struct util_dynarray draw_patches;
 
-       /* Keep track of blitter GMEM offsets that need to be patched up once we
-        * know the gmem layout:
-        */
-       struct util_dynarray gmem_patches;
+       /* texture state that needs patching for fb_read: */
+       struct util_dynarray fb_read_patches;
 
        /* Keep track of writes to RB_RENDER_CONTROL which need to be patched
         * once we know whether or not to use GMEM, and GMEM tile pitch.
@@ -153,8 +149,22 @@ struct fd_batch {
         */
        struct util_dynarray rbrc_patches;
 
+       /* Keep track of GMEM related values that need to be patched up once we
+        * know the gmem layout:
+        */
+       struct util_dynarray gmem_patches;
+
+       /* Keep track of pointer to start of MEM exports for a20x binning shaders
+        *
+        * this is so the end of the shader can be cut off at the right point
+        * depending on the GMEM configuration
+        */
+       struct util_dynarray shader_patches;
+
        struct pipe_framebuffer_state framebuffer;
 
+       struct fd_submit *submit;
+
        /** draw pass cmdstream: */
        struct fd_ringbuffer *draw;
        /** binning pass cmdstream: */
@@ -164,6 +174,12 @@ struct fd_batch {
 
        // TODO maybe more generically split out clear and clear_binning rings?
        struct fd_ringbuffer *lrz_clear;
+       struct fd_ringbuffer *tile_setup;
+       struct fd_ringbuffer *tile_fini;
+
+       union pipe_color_union clear_color[MAX_RENDER_TARGETS];
+       double clear_depth;
+       unsigned clear_stencil;
 
        /**
         * hw query related state:
@@ -213,7 +229,7 @@ struct fd_batch * fd_batch_create(struct fd_context *ctx, bool nondraw);
 
 void fd_batch_reset(struct fd_batch *batch);
 void fd_batch_sync(struct fd_batch *batch);
-void fd_batch_flush(struct fd_batch *batch, bool sync, bool force);
+void fd_batch_flush(struct fd_batch *batch, bool sync);
 void fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep);
 void fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write);
 void fd_batch_check_size(struct fd_batch *batch);
@@ -231,6 +247,10 @@ void __fd_batch_destroy(struct fd_batch *batch);
  * WARNING the _locked() version can briefly drop the lock.  Without
  * recursive mutexes, I'm not sure there is much else we can do (since
  * __fd_batch_destroy() needs to unref resources)
+ *
+ * WARNING you must acquire the screen->lock and use the _locked()
+ * version in case that the batch being ref'd can disappear under
+ * you.
  */
 
 /* fwd-decl prototypes to untangle header dependency :-/ */