radeonsi: Process multiple patches per threadgroup.

Using more than 1 wave per threadgroup does increase performance
generally.  Not using too many patches per threadgroup also
increases performance. Both catalyst and amdgpu-pro seem to
use 40 patches as their maximum, but I haven't really seen
any performance increase from limiting the number of patches
to 40 instead of 64.

Note that the trick where we overlap the input and output LDS
does not work anymore as the insertion of the tess factors
changes the patch stride.

v2: - Add comment about LDS assumptions.
    - Add constant for buffer size.
    - Fix code style.

v3: - Correct limits for not splitting patches between waves.
    - Set max num_patches to 40 as in the proprietary driver.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 6fe2619..c8b87a9 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -108,20 +108,7 @@ static void si_emit_derived_tess_state(struct si_context *sctx,
        unsigned input_patch_size, output_patch_size, output_patch0_offset;
        unsigned perpatch_output_offset, lds_size, ls_rsrc2;
        unsigned tcs_in_layout, tcs_out_layout, tcs_out_offsets;
-       unsigned offchip_layout;
-
-       *num_patches = 1; /* TODO: calculate this */
-
-       if (sctx->last_ls == ls->current &&
-           sctx->last_tcs == tcs &&
-           sctx->last_tes_sh_base == tes_sh_base &&
-           sctx->last_num_tcs_input_cp == num_tcs_input_cp)
-               return;
-
-       sctx->last_ls = ls->current;
-       sctx->last_tcs = tcs;
-       sctx->last_tes_sh_base = tes_sh_base;
-       sctx->last_num_tcs_input_cp = num_tcs_input_cp;
+       unsigned offchip_layout, hardware_lds_size;
 
        /* This calculates how shader inputs and outputs among VS, TCS, and TES
         * are laid out in LDS. */
@@ -146,7 +133,29 @@ static void si_emit_derived_tess_state(struct si_context *sctx,
        pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
        output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
 
-       output_patch0_offset = sctx->tcs_shader.cso ? input_patch_size * *num_patches : 0;
+       /* Ensure that we only need one wave per SIMD so we don't need to check
+        * resource usage. Also ensures that the number of tcs in and out
+        * vertices per threadgroup are at most 256.
+        */
+       *num_patches = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp) * 4;
+
+       /* Make sure that the data fits in LDS. This assumes the shaders only
+        * use LDS for the inputs and outputs.
+        */
+       hardware_lds_size = sctx->b.chip_class >= CIK ? 65536 : 32768;
+       *num_patches = MIN2(*num_patches, hardware_lds_size / (input_patch_size +
+                                                              output_patch_size));
+
+       /* Make sure the output data fits in the offchip buffer */
+       *num_patches = MIN2(*num_patches, SI_TESS_OFFCHIP_BLOCK_SIZE /
+                                         output_patch_size);
+
+       /* Not necessary for correctness, but improves performance. The
+        * specific value is taken from the proprietary driver.
+        */
+       *num_patches = MIN2(*num_patches, 40);
+
+       output_patch0_offset = input_patch_size * *num_patches;
        perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size;
 
        lds_size = output_patch0_offset + output_patch_size * *num_patches;
@@ -160,6 +169,17 @@ static void si_emit_derived_tess_state(struct si_context *sctx,
                ls_rsrc2 |= S_00B52C_LDS_SIZE(align(lds_size, 256) / 256);
        }
 
+       if (sctx->last_ls == ls->current &&
+           sctx->last_tcs == tcs &&
+           sctx->last_tes_sh_base == tes_sh_base &&
+           sctx->last_num_tcs_input_cp == num_tcs_input_cp)
+               return;
+
+       sctx->last_ls = ls->current;
+       sctx->last_tcs = tcs;
+       sctx->last_tes_sh_base = tes_sh_base;
+       sctx->last_num_tcs_input_cp = num_tcs_input_cp;
+
        /* Due to a hw bug, RSRC2_LS must be written twice with another
         * LS register written in between. */
        if (sctx->b.chip_class == CIK && sctx->b.family != CHIP_HAWAII)