nx2: lower 32-bit MATCH.ALL

Only MATCH.ALL.U64 was lowered; Nintendo Switch Sports Resort and Kirby
Air Riders use the 32-bit form (MATCH.ALL P0, RZ, Rx) in compute and
pixel shaders. Compare one word instead of two.

1 files changed, 9 insertions, 6 deletions

diff --git a/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h b/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
index 4cc2d0eb84..9a5b55033b 100644
--- a/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
+++ b/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
@@ -836,23 +836,26 @@ inline LoweredProgram LowerToMaxwell(std::span<const Instruction> code) {
             emit(0x3828000000000000ULL | predicate | (dest + 216) | (u64{insn.words[1] + 216} << 8) |
                  (u64{1} << 39) | (u64{1} << 40));
             break;
-        case 0x3a1: { // MATCH.ALL.U64 with only the all-equal predicate retained.
+        case 0x3a1: { // MATCH.ALL (.U32, or .U64 on bit 73) with only the all-equal predicate retained.
             const u32 source = insn.Bits(24, 8), target = insn.Bits(81, 3);
-            if (dest != 255 || source >= 254 || (source & 1) || target >= 6 ||
-                native_predicate != 7 || insn.words[1] || (insn.words[2] & ~0xe0000U) != 0x200)
+            const u32 words = insn.Bits(73, 1) ? 2 : 1;
+            if (dest != 255 || source >= 254 || (words == 2 && (source & 1)) || target >= 6 ||
+                native_predicate != 7 || insn.words[1] || (insn.words[2] & ~0xe0200U) != 0)
                 return fail(index, "native warp match mode");
             save_predicate();
+            // Compare every lane with the lowest active one, then require all to agree.
             emit(0x50d8000000070000ULL | 250 | (u64{7} << 39) | (u64{7} << 45) | (u64{1} << 48));
             emit(0x5c10000000070000ULL | 251 | (u64{250} << 8) | (u64{255} << 20) | (u64{1} << 49));
             emit(0x5c40000000070000ULL | 250 | (u64{250} << 8) | (u64{251} << 20) | (u64{7} << 48));
             emit(0x5c30000000070000ULL | 250 | (u64{250} << 20));
-            for (u32 word = 0; word < 2; ++word)
+            for (u32 word = 0; word < words; ++word)
                 emit(0xef10000000070000ULL | (251 + word) | (u64{source + word} << 8) |
                     (u64{250} << 20) | (u64{1} << 29) | (u64{31} << 34) | (u64{7} << 48));
             emit(0x5b60000000070000ULL | 7 | (u64{6} << 3) | (u64{source} << 8) |
                 (u64{251} << 20) | (u64{7} << 39) | (u64{2} << 49));
-            emit(0x5b60000000070000ULL | 7 | (u64{6} << 3) | (u64{source + 1} << 8) |
-                (u64{252} << 20) | (u64{6} << 39) | (u64{2} << 49));
+            if (words == 2)
+                emit(0x5b60000000070000ULL | 7 | (u64{6} << 3) | (u64{source + 1} << 8) |
+                    (u64{252} << 20) | (u64{6} << 39) | (u64{2} << 49));
             emit(0x50d8000000070000ULL | 255 | (u64{6} << 39) | (u64{target} << 45));
             restore_predicate();
             break;