Mesa (master): vc4: Only render tiles where the scissor ever intersected them.
Eric Anholt
anholt at kemper.freedesktop.org
Tue Dec 30 22:35:05 UTC 2014
Module: Mesa
Branch: master
Commit: 3ba57bae47666ada1145259755fc326b1b9f9463
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=3ba57bae47666ada1145259755fc326b1b9f9463
Author: Eric Anholt <eric at anholt.net>
Date: Sun Dec 28 08:14:19 2014 -1000
vc4: Only render tiles where the scissor ever intersected them.
This gives a 2.7x improvement in x11perf -rect100, since we only end up
load/storing the x11perf window, not the whole screen.
---
src/gallium/drivers/vc4/vc4_context.c | 37 ++++++++++++++++++++++++++-------
src/gallium/drivers/vc4/vc4_context.h | 10 +++++++++
src/gallium/drivers/vc4/vc4_draw.c | 4 ++++
src/gallium/drivers/vc4/vc4_emit.c | 11 ++++++++--
4 files changed, 52 insertions(+), 10 deletions(-)
diff --git a/src/gallium/drivers/vc4/vc4_context.c b/src/gallium/drivers/vc4/vc4_context.c
index 401eb21..e4e4330 100644
--- a/src/gallium/drivers/vc4/vc4_context.c
+++ b/src/gallium/drivers/vc4/vc4_context.c
@@ -94,8 +94,15 @@ vc4_setup_rcl(struct vc4_context *vc4)
uint32_t resolve_uncleared = vc4->resolve & ~vc4->cleared;
uint32_t width = vc4->framebuffer.width;
uint32_t height = vc4->framebuffer.height;
- uint32_t xtiles = align(width, 64) / 64;
- uint32_t ytiles = align(height, 64) / 64;
+ uint32_t stride_in_tiles = align(width, 64) / 64;
+
+ assert(vc4->draw_min_x != ~0 && vc4->draw_min_y != ~0);
+ uint32_t min_x_tile = vc4->draw_min_x / 64;
+ uint32_t min_y_tile = vc4->draw_min_y / 64;
+ uint32_t max_x_tile = (vc4->draw_max_x - 1) / 64;
+ uint32_t max_y_tile = (vc4->draw_max_y - 1) / 64;
+ uint32_t xtiles = max_x_tile - min_x_tile + 1;
+ uint32_t ytiles = max_y_tile - min_y_tile + 1;
#if 0
fprintf(stderr, "RCL: resolve 0x%x clear 0x%x resolve uncleared 0x%x\n",
@@ -171,10 +178,10 @@ vc4_setup_rcl(struct vc4_context *vc4)
uint32_t depth_hindex = ztex ? vc4_gem_hindex(vc4, ztex->bo) : 0;
uint32_t tile_alloc_hindex = vc4_gem_hindex(vc4, vc4->tile_alloc);
- for (int y = 0; y < ytiles; y++) {
- for (int x = 0; x < xtiles; x++) {
- bool end_of_frame = (x == xtiles - 1 &&
- y == ytiles - 1);
+ for (int y = min_y_tile; y <= max_y_tile; y++) {
+ for (int x = min_x_tile; x <= max_x_tile; x++) {
+ bool end_of_frame = (x == max_x_tile &&
+ y == max_y_tile);
bool coords_emitted = false;
/* Note that the load doesn't actually occur until the
@@ -225,13 +232,13 @@ vc4_setup_rcl(struct vc4_context *vc4)
/* Wait for the binner before jumping to the first
* tile's lists.
*/
- if (x == 0 && y == 0)
+ if (x == min_x_tile && y == min_y_tile)
cl_u8(&vc4->rcl, VC4_PACKET_WAIT_ON_SEMAPHORE);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_BRANCH_TO_SUB_LIST);
cl_reloc_hindex(&vc4->rcl, tile_alloc_hindex,
- (y * xtiles + x) * 32);
+ (y * stride_in_tiles + x) * 32);
if (vc4->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
@@ -313,6 +320,11 @@ vc4_draw_reset(struct vc4_context *vc4)
vc4->dirty = ~0;
vc4->resolve = 0;
vc4->cleared = 0;
+
+ vc4->draw_min_x = ~0;
+ vc4->draw_min_y = ~0;
+ vc4->draw_max_x = 0;
+ vc4->draw_max_y = 0;
}
void
@@ -323,6 +335,15 @@ vc4_flush(struct pipe_context *pctx)
if (!vc4->needs_flush)
return;
+ /* The RCL setup would choke if the draw bounds cause no drawing, so
+ * just drop the drawing if that's the case.
+ */
+ if (vc4->draw_max_x <= vc4->draw_min_x ||
+ vc4->draw_max_y <= vc4->draw_min_y) {
+ vc4_draw_reset(vc4);
+ return;
+ }
+
/* Increment the semaphore indicating that binning is done and
* unblocking the render thread. Note that this doesn't act until the
* FLUSH completes.
diff --git a/src/gallium/drivers/vc4/vc4_context.h b/src/gallium/drivers/vc4/vc4_context.h
index 962abbf..7e18a75 100644
--- a/src/gallium/drivers/vc4/vc4_context.h
+++ b/src/gallium/drivers/vc4/vc4_context.h
@@ -173,6 +173,16 @@ struct vc4_context {
struct vc4_cl bo_handles;
struct vc4_cl bo_pointers;
uint32_t shader_rec_count;
+ /** @{
+ * Bounding box of the scissor across all queued drawing.
+ *
+ * Note that the max values are exclusive.
+ */
+ uint32_t draw_min_x;
+ uint32_t draw_min_y;
+ uint32_t draw_max_x;
+ uint32_t draw_max_y;
+ /** @} */
struct vc4_bo *tile_alloc;
struct vc4_bo *tile_state;
diff --git a/src/gallium/drivers/vc4/vc4_draw.c b/src/gallium/drivers/vc4/vc4_draw.c
index d99faa4..0d91504 100644
--- a/src/gallium/drivers/vc4/vc4_draw.c
+++ b/src/gallium/drivers/vc4/vc4_draw.c
@@ -316,6 +316,10 @@ vc4_clear(struct pipe_context *pctx, unsigned buffers,
if (buffers & PIPE_CLEAR_STENCIL)
vc4->clear_stencil = stencil;
+ vc4->draw_min_x = 0;
+ vc4->draw_min_y = 0;
+ vc4->draw_max_x = vc4->framebuffer.width;
+ vc4->draw_max_y = vc4->framebuffer.height;
vc4->cleared |= buffers;
vc4->resolve |= buffers;
diff --git a/src/gallium/drivers/vc4/vc4_emit.c b/src/gallium/drivers/vc4/vc4_emit.c
index 6856441..d2b54fc 100644
--- a/src/gallium/drivers/vc4/vc4_emit.c
+++ b/src/gallium/drivers/vc4/vc4_emit.c
@@ -37,12 +37,19 @@ vc4_emit_state(struct pipe_context *pctx)
float vp_maxy = fabs(vpscale[1]) + vptranslate[1];
uint32_t minx = MAX2(vc4->scissor.minx, vp_minx);
uint32_t miny = MAX2(vc4->scissor.miny, vp_miny);
+ uint32_t maxx = MIN2(vc4->scissor.maxx, vp_maxx);
+ uint32_t maxy = MIN2(vc4->scissor.maxy, vp_maxy);
cl_u8(&vc4->bcl, VC4_PACKET_CLIP_WINDOW);
cl_u16(&vc4->bcl, minx);
cl_u16(&vc4->bcl, miny);
- cl_u16(&vc4->bcl, MIN2(vc4->scissor.maxx, vp_maxx) - minx);
- cl_u16(&vc4->bcl, MIN2(vc4->scissor.maxy, vp_maxy) - miny);
+ cl_u16(&vc4->bcl, maxx - minx);
+ cl_u16(&vc4->bcl, maxy - miny);
+
+ vc4->draw_min_x = MIN2(vc4->draw_min_x, minx);
+ vc4->draw_min_y = MIN2(vc4->draw_min_y, miny);
+ vc4->draw_max_x = MAX2(vc4->draw_max_x, maxx);
+ vc4->draw_max_y = MAX2(vc4->draw_max_y, maxy);
}
if (vc4->dirty & (VC4_DIRTY_RASTERIZER | VC4_DIRTY_ZSA)) {
More information about the mesa-commit
mailing list