<div dir="auto"><div> <div class="gmail_extra"> <div class="gmail_quote">On Jun 4, 2017 4:04 AM, "Dieter Nützel" <<a href="mailto:Dieter@nuetzel-hh.de">Dieter@nuetzel-hh.de</a>> wrote: <blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">It improves Uningine Heaven performance here (RX 580, 8 GB),too. without Tessellation Before: FPS: 79.0 Score: 1991 Min FPS: 20.8 Max FPS: 189.8 After: FPS: 79.1 Score: 1993 Min FPS: 19.6 Max FPS: 185.9 with Tessellation Before: FPS: 67.7 Score: 1705 Min FPS: 8.8 Max FPS: 182.6 After: FPS: 68.3 Score: 1720 Min FPS: 15.9 Max FPS: 179.6 System Platform: Linux 4.20.0-amd-staging-4.11-1.g7262353-default+ x86_64 CPU model: Intel(R) Xeon(R) CPU X3470 @ 2.93GHz (2925MHz) x8 GPU model: Unknown GPU (256MB) x1 Marek is this the 'tessellation regression' you tried to solve? </blockquote></div></div></div><div dir="auto"> </div><div dir="auto">No, the SI tessellation regression needs a different fix.</div><div dir="auto"> </div><div dir="auto">Marek</div><div dir="auto"> </div><div dir="auto"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> For this series: Tested-by: Dieter Nützel <<a href="mailto:Dieter@nuetzel-hh.de" target="_blank">Dieter@nuetzel-hh.de</a>> Dieter<div class="elided-text"> Am 03.06.2017 18:04, schrieb Marek Olšák: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> From: Marek Olšák <<a href="mailto:marek.olsak@amd.com" target="_blank">marek.olsak@amd.com</a>> Heaven LDS usage for LS+HS is below. The masks are "outputs_written" for LS and HS. Note that 32K is the maximum size. Before: heaven_x64: ls=1f1 tcs=1f1, lds=32K heaven_x64: ls=31 tcs=31, lds=24K heaven_x64: ls=71 tcs=71, lds=28K After: heaven_x64: ls=3f tcs=3f, lds=24K heaven_x64: ls=7 tcs=7, lds=13K heaven_x64: ls=f tcs=f, lds=17K All other apps have a similar decrease in LDS usage, because the "outputs_written" masks are similar. Also, most apps don't write POSITION in these shader stages, so there is room for improvement. (tight per-component input/output packing might help even more) It's unknown whether this improves performance. --- src/gallium/drivers/radeonsi/si_shader.c | 18 +++++++++++------- src/gallium/drivers/radeonsi/si_state_shaders.c | 4 +++- 2 files changed, 14 insertions(+), 8 deletions(-) diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index ddfaa3b..3a86c0b 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -129,32 +129,36 @@ unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned in /** * Returns a unique index for a semantic name and index. The index must be * less than 64, so that a 64-bit bitmask of used inputs or outputs can be * calculated. */ unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index) { switch (semantic_name) { case TGSI_SEMANTIC_POSITION: return 0; - case TGSI_SEMANTIC_PSIZE: - return 1; - case TGSI_SEMANTIC_CLIPDIST: - assert(index <= 1); - return 2 + index; case TGSI_SEMANTIC_GENERIC: + /* Since some shader stages use the the highest used IO index + * to determine the size to allocate for inputs/outputs + * (in LDS, tess and GS rings). GENERIC should be placed right + * after POSITION to make that size as small as possible. + */ if (index < SI_MAX_IO_GENERIC) - return 4 + index; + return 1 + index; assert(!"invalid generic index"); return 0; - + case TGSI_SEMANTIC_PSIZE: + return SI_MAX_IO_GENERIC + 1; + case TGSI_SEMANTIC_CLIPDIST: + assert(index <= 1); + return SI_MAX_IO_GENERIC + 2 + index; case TGSI_SEMANTIC_FOG: return SI_MAX_IO_GENERIC + 4; case TGSI_SEMANTIC_LAYER: return SI_MAX_IO_GENERIC + 5; case TGSI_SEMANTIC_VIEWPORT_INDEX: return SI_MAX_IO_GENERIC + 6; case TGSI_SEMANTIC_PRIMID: return SI_MAX_IO_GENERIC + 7; case TGSI_SEMANTIC_COLOR: /* these alias */ case TGSI_SEMANTIC_BCOLOR: diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c index 8ac4309..f36997b 100644 --- a/src/gallium/drivers/radeonsi/si_state_shaders.c +++ b/src/gallium/drivers/radeonsi/si_state_shaders.c @@ -1226,21 +1226,23 @@ static void si_shader_selector_key_hw_vs(struct si_context *sctx, ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard || (!ps_colormask && !ps_modifies_zs && !ps->info.writes_memory); } /* Find out which VS outputs aren't used by the PS. */ uint64_t outputs_written = vs->outputs_written; uint64_t inputs_read = 0; - outputs_written &= ~0x3; /* ignore POSITION, PSIZE */ + /* ignore POSITION, PSIZE */ + outputs_written &= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION, 0) | + (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE, 0)))); if (!ps_disabled) { inputs_read = ps->inputs_read; } uint64_t linked = outputs_written & inputs_read; key->opt.hw_vs.kill_outputs = ~linked & outputs_written; } </blockquote> </div></blockquote></div> </div></div></div>