diff options
author | Marek Olšák <marek.olsak@amd.com> | 2021-02-09 18:56:04 -0500 |
---|---|---|
committer | Marek Olšák <marek.olsak@amd.com> | 2021-02-17 04:49:24 -0500 |
commit | 98ea523e007efa71adecfcce92a168efcf9b54dd (patch) | |
tree | b7b9397a25c5c100632385b0515f66f8cc582cd8 | |
parent | 4fe37b850a4c00221a210481eeb3b9f5ec68e3ea (diff) |
radeonsi: for tess, determine the minimum num_patches before optimizing tg size
Doing these MINs at the end could have undone optimizations for the LDS
size and threadgroup size, so move the MINs up.
Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9028>
-rw-r--r-- | src/gallium/drivers/radeonsi/si_state_draw.cpp | 32 |
1 files changed, 16 insertions, 16 deletions
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.cpp b/src/gallium/drivers/radeonsi/si_state_draw.cpp index a7aa9612e6c..62f4f3e1422 100644 --- a/src/gallium/drivers/radeonsi/si_state_draw.cpp +++ b/src/gallium/drivers/radeonsi/si_state_draw.cpp @@ -271,6 +271,22 @@ static void si_emit_derived_tess_state(struct si_context *sctx, unsigned max_verts_per_patch = MAX2(num_tcs_input_cp, num_tcs_output_cp); *num_patches = 256 / max_verts_per_patch; + /* Not necessary for correctness, but higher numbers are slower. + * The hardware can do more, but the radeonsi shader constant is + * limited to 6 bits. + */ + *num_patches = MIN2(*num_patches, 64); /* e.g. 64 triangles in exactly 3 waves */ + + /* When distributed tessellation is unsupported, switch between SEs + * at a higher frequency to manually balance the workload between SEs. + */ + if (!sctx->screen->info.has_distributed_tess && sctx->screen->info.max_se > 1) + *num_patches = MIN2(*num_patches, 16); /* recommended */ + + /* Make sure the output data fits in the offchip buffer */ + *num_patches = + MIN2(*num_patches, (sctx->screen->tess_offchip_block_dw_size * 4) / output_patch_size); + /* Make sure that the data fits in LDS. This assumes the shaders only * use LDS for the inputs and outputs. * @@ -286,22 +302,6 @@ static void si_emit_derived_tess_state(struct si_context *sctx, *num_patches = MAX2(*num_patches, 1); assert(*num_patches * lds_per_patch <= max_lds_size); - /* Make sure the output data fits in the offchip buffer */ - *num_patches = - MIN2(*num_patches, (sctx->screen->tess_offchip_block_dw_size * 4) / output_patch_size); - - /* Not necessary for correctness, but improves performance. - * The hardware can do more, but the radeonsi shader constant is - * limited to 6 bits. - */ - *num_patches = MIN2(*num_patches, 64); /* triangles: 3 full waves */ - - /* When distributed tessellation is unsupported, switch between SEs - * at a higher frequency to compensate for it. - */ - if (!sctx->screen->info.has_distributed_tess && sctx->screen->info.max_se > 1) - *num_patches = MIN2(*num_patches, 16); /* recommended */ - /* Make sure that vector lanes are reasonably occupied. It probably * doesn't matter much because this is LS-HS, and TES is likely to * occupy significantly more CUs. |