freedreno/a6xx: do streamout only in binning pass

Use VPC_SO_OVERRIDE to control whether we do streamout in binning or
draw pass.  Normally we want to do streamout in binning pass, except
when there is a single tile and binning passed is skipped.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_emit.c b/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
index 3e8d41c4b24bee7da5f33c4650b2da0065aadf8b..4edc312e9932755911fdfe6f6f62e44ae6b627c1 100644 (file)
--- a/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
@@ -757,20 +757,13 @@ fd6_emit_streamout(struct fd_ringbuffer *ring, struct fd6_emit *emit, struct ir3
                        OUT_RING(ring, REG_A6XX_VPC_SO_PROG);
                        OUT_RING(ring, tf->prog[i]);
                }
-
-               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
-               OUT_RING(ring, 0x0);
        } else {
                OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 4);
                OUT_RING(ring, REG_A6XX_VPC_SO_CNTL);
                OUT_RING(ring, 0);
                OUT_RING(ring, REG_A6XX_VPC_SO_BUF_CNTL);
                OUT_RING(ring, 0);
-
-               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
-               OUT_RING(ring, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
        }
-
 }
 
 void

diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
index 7ac1bae13ee4046a4145ea84ee9d4d6e05faf634..678541c0b54681c33c734dda0b91bdfe2f6b6d21 100644 (file)
--- a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
@@ -762,11 +762,19 @@ fd6_emit_tile_init(struct fd_batch *batch)
        patch_fb_read(batch);
 
        if (use_hw_binning(batch)) {
+               /* enable stream-out during binning pass: */
+               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
+               OUT_RING(ring, 0);
+
                set_bin_size(ring, gmem->bin_w, gmem->bin_h,
                                A6XX_RB_BIN_CONTROL_BINNING_PASS | 0x6000000);
                update_render_cntl(batch, pfb, true);
                emit_binning_pass(batch);
 
+               /* and disable stream-out for draw pass: */
+               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
+               OUT_RING(ring, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
+
                /*
                 * NOTE: even if we detect VSC overflow and disable use of
                 * visibility stream in draw pass, it is still safe to execute
@@ -788,6 +796,10 @@ fd6_emit_tile_init(struct fd_batch *batch)
                OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
                OUT_RING(ring, 0x1);
        } else {
+               /* no binning pass, so enable stream-out for draw pass:: */
+               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
+               OUT_RING(ring, 0);
+
                set_bin_size(ring, gmem->bin_w, gmem->bin_h, 0x6000000);
        }
 
@@ -834,9 +846,6 @@ fd6_emit_tile_prep(struct fd_batch *batch, struct fd_tile *tile)
 
        set_scissor(ring, x1, y1, x2, y2);
 
-       OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
-       OUT_RING(ring, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
-
        if (use_hw_binning(batch)) {
                struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[tile->p];
 
@@ -886,9 +895,6 @@ fd6_emit_tile_prep(struct fd_batch *batch, struct fd_tile *tile)
                struct fd_gmem_stateobj *gmem = &batch->ctx->gmem;
                set_bin_size(ring, gmem->bin_w, gmem->bin_h, 0x6000000);
 
-               OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
-               OUT_RING(ring, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
-
                OUT_PKT7(ring, CP_SET_MODE, 1);
                OUT_RING(ring, 0x0);
 
@@ -1450,6 +1456,10 @@ fd6_emit_sysmem_prep(struct fd_batch *batch)
        OUT_PKT4(ring, REG_A6XX_RB_CCU_CNTL, 1);
        OUT_RING(ring, 0x10000000);   /* RB_CCU_CNTL */
 
+       /* enable stream-out, with sysmem there is only one pass: */
+       OUT_PKT4(ring, REG_A6XX_VPC_SO_OVERRIDE, 1);
+       OUT_RING(ring, 0);
+
        set_scissor(ring, 0, 0, pfb->width - 1, pfb->height - 1);
 
        set_window_offset(ring, 0, 0);