From: Marek Olšák <marek.ol...@amd.com>

---
 src/gallium/drivers/radeonsi/si_clear.c | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_clear.c 
b/src/gallium/drivers/radeonsi/si_clear.c
index 520e5b94f4a..8aa3355afc8 100644
--- a/src/gallium/drivers/radeonsi/si_clear.c
+++ b/src/gallium/drivers/radeonsi/si_clear.c
@@ -425,51 +425,50 @@ static void si_do_fast_color_clear(struct si_context 
*sctx,
                 */
                if (tex->buffer.b.is_shared &&
                    !(tex->buffer.external_usage & 
PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
                        continue;
 
                if (sctx->chip_class <= VI &&
                    tex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D 
&&
                    !sctx->screen->info.htile_cmask_support_1d_tiling)
                        continue;
 
-               bool need_decompress_pass = false;
-
                /* Use a slow clear for small surfaces where the cost of
                 * the eliminate pass can be higher than the benefit of fast
                 * clear. The closed driver does this, but the numbers may 
differ.
                 *
                 * 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 eliminate_needed = false;
+               bool fmask_decompress_needed = false;
 
                /* Fast clear is the most appropriate place to enable DCC for
                 * displayable surfaces.
                 */
                if (sctx->family == CHIP_STONEY && !too_small) {
                        vi_separate_dcc_try_enable(sctx, tex);
 
                        /* RB+ isn't supported with a CMASK clear only on 
Stoney,
                         * so all clears are considered to be hypothetically 
slow
                         * clears, which is weighed when determining whether to
                         * enable separate DCC.
                         */
                        if (tex->dcc_gather_statistics) /* only for Stoney */
                                tex->num_slow_clears++;
                }
 
                /* Try to clear DCC first, otherwise try CMASK. */
                if (vi_dcc_enabled(tex, 0)) {
                        uint32_t reset_value;
-                       bool eliminate_needed;
 
                        if (sctx->screen->debug_flags & DBG(NO_DCC_CLEAR))
                                continue;
 
                        /* This can happen with mipmapping or MSAA. */
                        if (sctx->chip_class == VI &&
                            
!tex->surface.u.legacy.level[level].dcc_fast_clear_size)
                                continue;
 
                        if 
(!vi_get_fast_clear_parameters(tex->buffer.b.b.format,
@@ -484,28 +483,24 @@ static void si_do_fast_color_clear(struct si_context 
*sctx,
                        /* DCC fast clear with MSAA should clear CMASK to 0xC. 
*/
                        if (tex->buffer.b.b.nr_samples >= 2 && 
tex->cmask_buffer) {
                                /* TODO: This doesn't work with MSAA. */
                                if (eliminate_needed)
                                        continue;
 
                                uint32_t clear_value = 0xCCCCCCCC;
                                si_clear_buffer(sctx, &tex->cmask_buffer->b.b,
                                                tex->cmask_offset, 
tex->surface.cmask_size,
                                                &clear_value, 4, 
SI_COHERENCY_CB_META);
-                               need_decompress_pass = true;
+                               fmask_decompress_needed = true;
                        }
 
                        vi_dcc_clear_level(sctx, tex, 0, reset_value);
-
-                       if (eliminate_needed)
-                               need_decompress_pass = true;
-
                        tex->separate_dcc_dirty = true;
                } else {
                        if (too_small)
                                continue;
 
                        /* 128-bit formats are unusupported */
                        if (tex->surface.bpe > 8) {
                                continue;
                        }
 
@@ -516,37 +511,38 @@ static void si_do_fast_color_clear(struct si_context 
*sctx,
                        /* ensure CMASK is enabled */
                        si_alloc_separate_cmask(sctx->screen, tex);
                        if (!tex->cmask_buffer)
                                continue;
 
                        /* Do the fast clear. */
                        uint32_t clear_value = 0;
                        si_clear_buffer(sctx, &tex->cmask_buffer->b.b,
                                        tex->cmask_offset, 
tex->surface.cmask_size,
                                        &clear_value, 4, SI_COHERENCY_CB_META);
-                       need_decompress_pass = true;
+                       eliminate_needed = true;
                }
 
-               if (need_decompress_pass &&
+               if ((eliminate_needed || fmask_decompress_needed) &&
                    !(tex->dirty_level_mask & (1 << level))) {
                        tex->dirty_level_mask |= 1 << level;
                        
p_atomic_inc(&sctx->screen->compressed_colortex_counter);
                }
 
                /* We can change the micro tile mode before a full clear. */
                si_set_optimal_micro_tile_mode(sctx->screen, tex);
 
+               *buffers &= ~clear_bit;
+
                if (si_set_clear_color(tex, fb->cbufs[i]->format, color)) {
                        sctx->framebuffer.dirty_cbufs |= 1 << i;
                        si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
                }
-               *buffers &= ~clear_bit;
        }
 }
 
 static void si_clear(struct pipe_context *ctx, unsigned buffers,
                     const union pipe_color_union *color,
                     double depth, unsigned stencil)
 {
        struct si_context *sctx = (struct si_context *)ctx;
        struct pipe_framebuffer_state *fb = &sctx->framebuffer.state;
        struct pipe_surface *zsbuf = fb->zsbuf;
-- 
2.17.1

_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to