KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

radeonsi: refine fast clears for small buffers, always use them for large HTILE 

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10003>
This commit is contained in:
Marek Olšák 2021-03-29 02:24:39 -04:00 committed by Marge Bot
parent 06b6af596c
commit f7c58559f5
1 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
src/gallium/drivers/radeonsi/si_clear.c
View File

@ -538,6 +538,7 @@ static void si_fast_clear(struct si_context *sctx, unsigned *buffers,
struct si_clear_info info[8 * 2 + 1]; /* MRTs * (CMASK + DCC) + ZS */
unsigned num_clears = 0;
unsigned clear_types = 0;
bool fb_too_small = fb->width * fb->height * fb->layers <= 512 * 512;
/* This function is broken in BE, so just disable this path for now */
#if UTIL_ARCH_BIG_ENDIAN
@ -591,8 +592,7 @@ static void si_fast_clear(struct si_context *sctx, unsigned *buffers,
*
* This helps on both dGPUs and APUs, even small APUs like Mullins.
*/
bool too_small = tex->buffer.b.b.nr_samples <= 1 &&
tex->buffer.b.b.width0 * tex->buffer.b.b.height0 <= 512 * 512;
bool too_small = tex->buffer.b.b.nr_samples <= 1 && fb_too_small;
bool eliminate_needed = false;
bool fmask_decompress_needed = false;
@ -843,12 +843,13 @@ static void si_fast_clear(struct si_context *sctx, unsigned *buffers,
zstex->surface.meta_offset, zstex->surface.meta_size, clear_value);
clear_types |= SI_CLEAR_TYPE_HTILE;
}
} else if (num_clears) {
} else if (num_clears || !fb_too_small) {
/* This is where the HTILE buffer clear is done.
*
* If there is no clear scheduled, we should use the draw-based clear that is without
* waits. If there is some other clear scheduled, we will have to wait anyway, so add
* the HTILE buffer clear to the batch here.
* If there is no clear scheduled and the framebuffer size is too small, we should use
* the draw-based clear that is without waits. If there is some other clear scheduled,
* we will have to wait anyway, so add the HTILE buffer clear to the batch here.
* If the framebuffer size is large enough, use this codepath too.
*/
uint64_t htile_offset = zstex->surface.meta_offset;
unsigned htile_size = 0;