nx2: lower TLD4 with a texel offset

Bit 76 puts a texel offset after the coordinates, as for TEX: x and y in Ra, Ra+1
and the offsets in Rb. Maxwell TLD4 takes it as AOFFI (bits 54-55) from Rb.

1 files changed, 16 insertions, 7 deletions

diff --git a/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h b/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
index d9c52522f3..5436b50d39 100644
--- a/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
+++ b/src/shader_recompiler/frontend/ampere/lower_to_maxwell.h
@@ -1649,26 +1649,35 @@ inline LoweredProgram LowerToMaxwellWith(std::span<const Instruction> code, u32
             const u32 mask = insn.Bits(72, 4), second_dest = insn.Bits(64, 8);
             const u32 source_a = insn.Bits(24, 8), source_b = insn.Bits(32, 8);
             const u32 binding = insn.Bits(54, 5);
-            if (insn.Bits(61, 3) != 1 || insn.Bits(76, 4) || insn.Bits(81, 3) != 7 || insn.Bits(59, 1) ||
+            // Bit 76: a texel offset after the coordinates (x, y in Ra, Ra+1, offsets in Rb, as
+            // for TEX); Maxwell TLD4 takes it as AOFFI (bits 54-55) from Rb.
+            const bool offsets = insn.Bits(76, 1) != 0;
+            if (insn.Bits(61, 3) != 1 || insn.Bits(77, 3) || insn.Bits(81, 3) != 7 || insn.Bits(59, 1) ||
+                (offsets && (bindless || source_b == 255 || (source_a != 255 && source_a >= 245))) ||
                 insn.Bits(89, 1) || mask == 0 ||
                 (bindless ? (handle >= 246 || insn.Bits(46, 13) || !insn.Bits(91, 1))
                           : (insn.Bits(40, 14) > 0x1fff || binding >= 18 ||
                              (result.required_texture_buffer != ~0U && result.required_texture_buffer != binding))))
                 return fail(index, "native texture gather mode");
             if (!bindless) result.required_texture_buffer = binding;
-            // Two coordinates: x in Ra, y in Rb (the coordinate list splits evenly across them).
+            // Two coordinates: x in Ra, y in Rb (the coordinate list splits evenly across them);
+            // with an offset, x and y in Ra, Ra+1 and the offset in Rb.
+            const u32 y = !offsets ? source_b : source_a == 255 ? 255U : source_a + 1;
             emit(0x5c98000000000000ULL | predicate | 250 | (u64{15} << 39) | (u64{source_a} << 20));
-            emit(0x5c98000000000000ULL | predicate | 251 | (u64{15} << 39) |
-                 (u64{source_b == 255 ? 255U : source_b} << 20));
+            emit(0x5c98000000000000ULL | predicate | 251 | (u64{15} << 39) | (u64{y} << 20));
+            if (offsets) {
+                emit(0x5c98000000000000ULL | predicate | 246 | (u64{15} << 39) | (u64{source_b} << 20));
+            }
             if (bindless) {
                 emit(0x5c98000000000000ULL | predicate | 246 | (u64{15} << 39) | (u64{handle} << 20));
                 emit(0xdec0000000000000ULL | predicate | 250 | (u64{250} << 8) | (u64{246} << 20) |
                      (u64{2} << 28) | (u64{mask} << 31) | (u64{insn.Bits(87, 2)} << 38) |
                      (u64{insn.Bits(90, 1)} << 49) | (u64{7} << 51));
             } else {
-            emit(0xc800000000000000ULL | predicate | 250 | (u64{250} << 8) | (u64{255} << 20) |
-                 (u64{2} << 28) | (u64{mask} << 31) | (u64{insn.Bits(40, 14)} << 36) |
-                 (u64{insn.Bits(90, 1)} << 49) | (u64{7} << 51) | (u64{insn.Bits(87, 2)} << 56));
+            emit(0xc800000000000000ULL | predicate | 250 | (u64{250} << 8) |
+                 (u64{offsets ? 246U : 255U} << 20) | (u64{2} << 28) | (u64{mask} << 31) |
+                 (u64{insn.Bits(40, 14)} << 36) | (u64{insn.Bits(90, 1)} << 49) | (u64{7} << 51) |
+                 (u64{offsets ? 1U : 0U} << 54) | (u64{insn.Bits(87, 2)} << 56));
             }
             const u32 components = static_cast<u32>(std::popcount(mask));
             const bool split = second_dest != 255;