Mesa (main): aco: return 0x76543210 for NULL FMASK fetch

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: bf0cc05227d0a4a90f4bec13e826123158c2a27b
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=bf0cc05227d0a4a90f4bec13e826123158c2a27b

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Wed Aug  4 16:13:47 2021 +0100

aco: return 0x76543210 for NULL FMASK fetch

This can replace several v_cndmask_b32 with a single v_cndmask_b32, and
will be useful when we lower sample index adjustment in NIR.

fossil-db (Sienna Cichlid):
Totals from 955 (0.64% of 150170) affected shaders:
VGPRs: 53232 -> 53208 (-0.05%)
CodeSize: 4712548 -> 4722016 (+0.20%); split: -0.02%, +0.23%
MaxWaves: 19052 -> 19056 (+0.02%)
Instrs: 875891 -> 875619 (-0.03%); split: -0.04%, +0.00%
Latency: 14070164 -> 14069089 (-0.01%); split: -0.02%, +0.01%
InvThroughput: 2322982 -> 2321419 (-0.07%); split: -0.08%, +0.01%
VClause: 23070 -> 23080 (+0.04%); split: -0.00%, +0.05%
SClause: 32463 -> 32426 (-0.11%); split: -0.12%, +0.01%
Copies: 42840 -> 42787 (-0.12%); split: -0.19%, +0.07%
Branches: 17907 -> 17900 (-0.04%); split: -0.06%, +0.02%
PreSGPRs: 43585 -> 43229 (-0.82%)
PreVGPRs: 47676 -> 47625 (-0.11%)

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12214>

---

 src/amd/compiler/aco_instruction_selection.cpp | 56 +++++++++++++++++---------
 1 file changed, 36 insertions(+), 20 deletions(-)

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 2f04619a7a3..e1915ce40c6 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -6025,6 +6025,17 @@ adjust_sample_index_using_fmask(isel_context* ctx, bool da, std::vector<Temp>& c
    load->da = da;
    load->dim = dim;
 
+   /* Don't adjust the sample index if WORD1.DATA_FORMAT of the FMASK
+    * resource descriptor is 0 (invalid),
+    */
+   Temp is_not_null = bld.tmp(bld.lm);
+   bld.vopc_e64(aco_opcode::v_cmp_lg_u32, Definition(is_not_null), Operand::zero(),
+                emit_extract_vector(ctx, fmask_desc_ptr, 1, s1))
+      .def(0)
+      .setHint(vcc);
+   fmask =
+      bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), Operand::c32(0x76543210), fmask, is_not_null);
+
    Operand sample_index4;
    if (sample_index.isConstant()) {
       if (sample_index.constantValue() < 16) {
@@ -6041,28 +6052,12 @@ adjust_sample_index_using_fmask(isel_context* ctx, bool da, std::vector<Temp>& c
          bld.vop2(aco_opcode::v_lshlrev_b32, bld.def(v1), Operand::c32(2u), sample_index);
    }
 
-   Temp final_sample;
    if (sample_index4.isConstant() && sample_index4.constantValue() == 0)
-      final_sample = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand::c32(15u), fmask);
+      return bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand::c32(15u), fmask);
    else if (sample_index4.isConstant() && sample_index4.constantValue() == 28)
-      final_sample = bld.vop2(aco_opcode::v_lshrrev_b32, bld.def(v1), Operand::c32(28u), fmask);
+      return bld.vop2(aco_opcode::v_lshrrev_b32, bld.def(v1), Operand::c32(28u), fmask);
    else
-      final_sample =
-         bld.vop3(aco_opcode::v_bfe_u32, bld.def(v1), fmask, sample_index4, Operand::c32(4u));
-
-   /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
-    * resource descriptor is 0 (invalid),
-    */
-   Temp compare = bld.tmp(bld.lm);
-   bld.vopc_e64(aco_opcode::v_cmp_lg_u32, Definition(compare), Operand::zero(),
-                emit_extract_vector(ctx, fmask_desc_ptr, 1, s1))
-      .def(0)
-      .setHint(vcc);
-
-   Temp sample_index_v = bld.copy(bld.def(v1), sample_index);
-
-   /* Replace the MSAA sample index. */
-   return bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), sample_index_v, final_sample, compare);
+      return bld.vop3(aco_opcode::v_bfe_u32, bld.def(v1), fmask, sample_index4, Operand::c32(4u));
 }
 
 static std::vector<Temp>
@@ -9556,7 +9551,8 @@ visit_tex(isel_context* ctx, nir_tex_instr* instr)
          dmask = 1 << instr->component;
       if (tg4_integer_cube_workaround || dst.type() == RegType::sgpr)
          tmp_dst = bld.tmp(instr->is_sparse ? v5 : v4);
-   } else if (instr->op == nir_texop_samples_identical) {
+   } else if (instr->op == nir_texop_samples_identical ||
+              instr->op == nir_texop_fragment_mask_fetch_amd) {
       tmp_dst = bld.tmp(v1);
    } else if (util_bitcount(dmask) != instr->dest.ssa.num_components ||
               dst.type() == RegType::sgpr) {
@@ -9758,6 +9754,26 @@ visit_tex(isel_context* ctx, nir_tex_instr* instr)
          bld.vopc(aco_opcode::v_cmp_eq_u32, Definition(dst), Operand::zero(), tmp_dst)
             .def(0)
             .setHint(vcc);
+      } else if (instr->op == nir_texop_fragment_mask_fetch_amd) {
+         /* Use 0x76543210 if the image doesn't have FMASK. */
+         assert(dmask == 1 && dst.bytes() == 4);
+         assert(dst.id() != tmp_dst.id());
+
+         if (dst.regClass() == s1) {
+            Temp is_not_null = bld.sopc(aco_opcode::s_cmp_lg_u32, bld.def(s1, scc), Operand::zero(),
+                                        emit_extract_vector(ctx, resource, 1, s1));
+            bld.sop2(aco_opcode::s_cselect_b32, Definition(dst),
+                     bld.as_uniform(tmp_dst), Operand::c32(0x76543210),
+                     bld.scc(is_not_null));
+         } else {
+            Temp is_not_null = bld.tmp(bld.lm);
+            bld.vopc_e64(aco_opcode::v_cmp_lg_u32, Definition(is_not_null), Operand::zero(),
+                         emit_extract_vector(ctx, resource, 1, s1))
+               .def(0)
+               .setHint(vcc);
+            bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst),
+                     bld.copy(bld.def(v1), Operand::c32(0x76543210)), tmp_dst, is_not_null);
+         }
       } else {
          expand_vector(ctx, tmp_dst, dst, instr->dest.ssa.num_components, dmask);
       }