radeonsi/gfx9: don't flush TC L2 between rendering and texturing if not needed

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This commit is contained in:
Marek Olšák 2017-08-18 15:51:59 +02:00
parent 287b0a28f4
commit 5b62eb237c
3 changed files with 47 additions and 34 deletions

View file

@ -391,29 +391,29 @@ si_decompress_depth(struct si_context *sctx,
/* Only in-place decompression needs to flush DB caches, or
* when we don't decompress but TC-compatible planes are dirty.
*/
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_VMEM_L1;
si_make_DB_shader_coherent(sctx, tex->resource.b.b.nr_samples,
inplace_planes & PIPE_MASK_S);
/* If we flush DB caches for TC-compatible depth, the dirty
* state becomes 0 for the whole mipmap tree and all planes.
* (there is nothing else to flush)
*/
if (tex->tc_compatible_htile) {
if (r600_can_sample_zs(tex, false))
/* Only clear the mask that we are flushing, because
* si_make_DB_shader_coherent() can treat depth and
* stencil differently.
*/
if (inplace_planes & PIPE_MASK_Z)
tex->dirty_level_mask = 0;
if (r600_can_sample_zs(tex, true))
if (inplace_planes & PIPE_MASK_S)
tex->stencil_dirty_level_mask = 0;
}
}
/* set_framebuffer_state takes care of coherency for single-sample.
* The DB->CB copy uses CB for the final writes.
*/
if (copy_planes && tex->resource.b.b.nr_samples > 1) {
sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_FLUSH_AND_INV_CB;
}
if (copy_planes && tex->resource.b.b.nr_samples > 1)
si_make_CB_shader_coherent(sctx, tex->resource.b.b.nr_samples);
}
static void
@ -524,10 +524,7 @@ static void si_blit_decompress_color(struct pipe_context *ctx,
}
sctx->decompression_enabled = false;
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_VMEM_L1;
si_make_CB_shader_coherent(sctx, rtex->resource.b.b.nr_samples);
}
static void
@ -1216,9 +1213,7 @@ static void si_do_CB_resolve(struct si_context *sctx,
si_blitter_end(&sctx->b.b);
/* Flush caches for possible texturing. */
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_VMEM_L1;
si_make_CB_shader_coherent(sctx, 1);
}
static bool do_hardware_msaa_resolve(struct pipe_context *ctx,

View file

@ -611,4 +611,27 @@ si_saved_cs_reference(struct si_saved_cs **dst, struct si_saved_cs *src)
*dst = src;
}
static inline void
si_make_CB_shader_coherent(struct si_context *sctx, unsigned num_samples)
{
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
SI_CONTEXT_INV_VMEM_L1;
/* Single-sample color is coherent with shaders on GFX9. */
if (sctx->b.chip_class <= VI || num_samples >= 2)
sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
}
static inline void
si_make_DB_shader_coherent(struct si_context *sctx, unsigned num_samples,
bool include_stencil)
{
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
SI_CONTEXT_INV_VMEM_L1;
/* Single-sample depth (not stencil) is coherent with shaders on GFX9. */
if (sctx->b.chip_class <= VI || num_samples >= 2 || include_stencil)
sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
}
#endif

View file

@ -2572,11 +2572,9 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
* Only flush and wait for CB if there is actually a bound color buffer.
*/
if (sctx->framebuffer.nr_samples <= 1 &&
sctx->framebuffer.state.nr_cbufs) {
sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_FLUSH_AND_INV_CB;
}
sctx->framebuffer.state.nr_cbufs)
si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples);
sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
/* u_blitter doesn't invoke depth decompression when it does multiple
@ -2585,11 +2583,8 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
* individual generate_mipmap blits.
* Note that lower mipmap levels aren't compressed.
*/
if (sctx->generate_mipmap_for_depth) {
sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_FLUSH_AND_INV_DB;
}
if (sctx->generate_mipmap_for_depth)
si_make_DB_shader_coherent(sctx, 1, false);
/* Take the maximum of the old and new count. If the new count is lower,
* dirtying is needed to disable the unbound colorbuffers.
@ -4026,11 +4021,8 @@ static void si_texture_barrier(struct pipe_context *ctx, unsigned flags)
/* Multisample surfaces are flushed in si_decompress_textures. */
if (sctx->framebuffer.nr_samples <= 1 &&
sctx->framebuffer.state.nr_cbufs) {
sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_FLUSH_AND_INV_CB;
}
sctx->framebuffer.state.nr_cbufs)
si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples);
}
/* This only ensures coherency for shader image/buffer stores. */
@ -4073,8 +4065,11 @@ static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
if (flags & PIPE_BARRIER_FRAMEBUFFER &&
sctx->framebuffer.nr_samples <= 1 &&
sctx->framebuffer.state.nr_cbufs) {
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
SI_CONTEXT_WRITEBACK_GLOBAL_L2;
sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
/* Single-sample color is coherent with TC on GFX9. */
if (sctx->screen->b.chip_class <= VI)
sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
}
/* Indirect buffers use TC L2 on GFX9, but not older hw. */