aco: set TRUNC_COORD=0 for nir_texop_tg4

Fixes black squares in Assassin's Creed: Valhalla and rendering of
FidelityFX-CACAO demo.

fossil-db (sienna cichlid):
Totals from 3052 (2.09% of 146267) affected shaders:
SpillSGPRs: 8437 -> 8646 (+2.48%)
CodeSize: 30993832 -> 31116916 (+0.40%); split: -0.00%, +0.40%
Instrs: 5869934 -> 5886783 (+0.29%); split: -0.00%, +0.29%
Latency: 250330521 -> 250463770 (+0.05%); split: -0.00%, +0.05%
InvThroughput: 59797617 -> 59814584 (+0.03%); split: -0.00%, +0.03%
VClause: 92114 -> 92132 (+0.02%)
SClause: 197373 -> 197338 (-0.02%); split: -0.02%, +0.01%
Copies: 479482 -> 482394 (+0.61%); split: -0.01%, +0.61%
Branches: 219629 -> 219635 (+0.00%)
PreSGPRs: 248970 -> 249366 (+0.16%)

fossil-db (polaris10):
Totals from 3050 (2.06% of 147787) affected shaders:
SGPRs: 282864 -> 282912 (+0.02%); split: -0.01%, +0.02%
VGPRs: 242572 -> 242612 (+0.02%)
SpillSGPRs: 10387 -> 10675 (+2.77%)
CodeSize: 31872460 -> 31996128 (+0.39%)
MaxWaves: 10924 -> 10925 (+0.01%)
Instrs: 6222217 -> 6239072 (+0.27%)
Latency: 317482545 -> 317773685 (+0.09%); split: -0.00%, +0.09%
InvThroughput: 156149624 -> 156242072 (+0.06%); split: -0.00%, +0.06%
VClause: 92295 -> 92254 (-0.04%); split: -0.05%, +0.01%
SClause: 243342 -> 243321 (-0.01%); split: -0.01%, +0.00%
Copies: 678902 -> 681700 (+0.41%); split: -0.00%, +0.41%
Branches: 219698 -> 219703 (+0.00%)
PreSGPRs: 244251 -> 244644 (+0.16%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Fixes: 58f25098a0d ("radv: Use TRUNC_COORD on samplers")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3110
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10036>

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 510020a..97d80eb 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -5437,8 +5437,10 @@ Temp get_sampler_desc(isel_context *ctx, nir_deref_instr *deref_instr,
          constant_index = 0;
 
       const uint32_t *samplers = radv_immutable_samplers(layout, binding);
+      uint32_t dword0_mask = tex_instr->op == nir_texop_tg4 ?
+                             C_008F30_TRUNC_COORD : 0xffffffffu;
       return bld.pseudo(aco_opcode::p_create_vector, bld.def(s4),
-                        Operand(samplers[constant_index * 4 + 0]),
+                        Operand(samplers[constant_index * 4 + 0] & dword0_mask),
                         Operand(samplers[constant_index * 4 + 1]),
                         Operand(samplers[constant_index * 4 + 2]),
                         Operand(samplers[constant_index * 4 + 3]));
@@ -5500,6 +5502,23 @@ Temp get_sampler_desc(isel_context *ctx, nir_deref_instr *deref_instr,
       res = bld.pseudo(aco_opcode::p_create_vector, bld.def(s8),
                        components[0], components[1], components[2], components[3],
                        components[4], components[5], components[6], components[7]);
+   } else if (desc_type == ACO_DESC_SAMPLER && tex_instr->op == nir_texop_tg4) {
+      Temp components[4];
+      for (unsigned i = 0; i < 4; i++)
+         components[i] = bld.tmp(s1);
+
+      bld.pseudo(aco_opcode::p_split_vector,
+                 Definition(components[0]), Definition(components[1]),
+                 Definition(components[2]), Definition(components[3]), res);
+
+      /* We want to always use the linear filtering truncation behaviour for
+       * nir_texop_tg4, even if the sampler uses nearest/point filtering.
+       */
+      components[0] = bld.sop2(aco_opcode::s_and_b32, bld.def(s1), bld.def(s1, scc),
+                               components[0], Operand((uint32_t)C_008F30_TRUNC_COORD));
+
+      res = bld.pseudo(aco_opcode::p_create_vector, bld.def(s4),
+                       components[0], components[1], components[2], components[3]);
    }
 
    return res;