Conor McCarthy : vkd3d-shader/spirv: Implement the WAVE_OP_* instructions.

[Alexandre Julliard]

Module: vkd3d
Branch: master
Commit: cfcc789b42ea707ed701cbcb0600be6156cfbbb1
URL:    https://gitlab.winehq.org/wine/vkd3d/-/commit/cfcc789b42ea707ed701cbcb0600be...

Author: Conor McCarthy 
Date:   Tue Apr 23 21:35:27 2024 +1000

vkd3d-shader/spirv: Implement the WAVE_OP_* instructions.

---

libs/vkd3d-shader/spirv.c                 | 26 +++++++++++++++++++++++++-
 tests/hlsl/wave-ops-float.shader_test     | 12 ++++++------
 tests/hlsl/wave-ops-int.shader_test       |  4 ++--
 tests/hlsl/wave-ops-uint.shader_test      |  6 +++---
 tests/hlsl/wave-reconvergence.shader_test |  4 ++--
 5 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c
index 60c08be2..7ac9f3cd 100644
--- a/libs/vkd3d-shader/spirv.c
+++ b/libs/vkd3d-shader/spirv.c
@@ -9826,6 +9826,22 @@ static SpvOp map_wave_alu_op(enum vkd3d_shader_opcode handler_idx, bool is_float
             return SpvOpGroupNonUniformBitwiseOr;
         case VKD3DSIH_WAVE_ACTIVE_BIT_XOR:
             return SpvOpGroupNonUniformBitwiseXor;
+        case VKD3DSIH_WAVE_OP_ADD:
+            return is_float ? SpvOpGroupNonUniformFAdd : SpvOpGroupNonUniformIAdd;
+        case VKD3DSIH_WAVE_OP_IMAX:
+            return SpvOpGroupNonUniformSMax;
+        case VKD3DSIH_WAVE_OP_IMIN:
+            return SpvOpGroupNonUniformSMin;
+        case VKD3DSIH_WAVE_OP_MAX:
+            return SpvOpGroupNonUniformFMax;
+        case VKD3DSIH_WAVE_OP_MIN:
+            return SpvOpGroupNonUniformFMin;
+        case VKD3DSIH_WAVE_OP_MUL:
+            return is_float ? SpvOpGroupNonUniformFMul : SpvOpGroupNonUniformIMul;
+        case VKD3DSIH_WAVE_OP_UMAX:
+            return SpvOpGroupNonUniformUMax;
+        case VKD3DSIH_WAVE_OP_UMIN:
+            return SpvOpGroupNonUniformUMin;
         default:
             vkd3d_unreachable();
     }
@@ -9849,7 +9865,7 @@ static void spirv_compiler_emit_wave_alu_op(struct spirv_compiler *compiler,
     vkd3d_spirv_enable_capability(builder, SpvCapabilityGroupNonUniformArithmetic);
     val_id = vkd3d_spirv_build_op_tr3(builder, &builder->function_stream, op, type_id,
             vkd3d_spirv_get_op_scope_subgroup(builder),
-            SpvGroupOperationReduce,
+            (instruction->flags & VKD3DSI_WAVE_PREFIX) ? SpvGroupOperationExclusiveScan : SpvGroupOperationReduce,
             val_id);

spirv_compiler_emit_store_dst(compiler, dst, val_id);
@@ -10210,6 +10226,14 @@ static int spirv_compiler_handle_instruction(struct spirv_compiler *compiler,
         case VKD3DSIH_WAVE_ACTIVE_BIT_AND:
         case VKD3DSIH_WAVE_ACTIVE_BIT_OR:
         case VKD3DSIH_WAVE_ACTIVE_BIT_XOR:
+        case VKD3DSIH_WAVE_OP_ADD:
+        case VKD3DSIH_WAVE_OP_IMAX:
+        case VKD3DSIH_WAVE_OP_IMIN:
+        case VKD3DSIH_WAVE_OP_MAX:
+        case VKD3DSIH_WAVE_OP_MIN:
+        case VKD3DSIH_WAVE_OP_MUL:
+        case VKD3DSIH_WAVE_OP_UMAX:
+        case VKD3DSIH_WAVE_OP_UMIN:
             spirv_compiler_emit_wave_alu_op(compiler, instruction);
             break;
         case VKD3DSIH_DCL:
diff --git a/tests/hlsl/wave-ops-float.shader_test b/tests/hlsl/wave-ops-float.shader_test
index 2c402db1..60774fa8 100644
--- a/tests/hlsl/wave-ops-float.shader_test
+++ b/tests/hlsl/wave-ops-float.shader_test
@@ -133,7 +133,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (14.0, 4.5, 16.5, 10.0)
 probe uav 1 (1) rgba (14.0, 4.5, 16.5, 10.0)
 probe uav 1 (2) rgba (14.0, 4.5, 16.5, 10.0)
@@ -151,7 +151,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (150.0625, 1.5, 288.0, 39.0625)
 probe uav 1 (1) rgba (150.0625, 1.5, 288.0, 39.0625)
 probe uav 1 (2) rgba (150.0625, 1.5, 288.0, 39.0625)
@@ -169,7 +169,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (3.5, 1.0, 4.0, 2.5)
 probe uav 1 (1) rgba (3.5, 1.0, 4.0, 2.5)
 probe uav 1 (2) rgba (3.5, 1.0, 4.0, 2.5)
@@ -187,7 +187,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (3.5, 1.5, 4.5, 2.5)
 probe uav 1 (1) rgba (3.5, 1.5, 4.5, 2.5)
 probe uav 1 (2) rgba (3.5, 1.5, 4.5, 2.5)
@@ -205,7 +205,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (0.0, 0.0, 0.0, 0.0)
 probe uav 1 (1) rgba (3.5, 1.0, 4.0, 2.5)
 probe uav 1 (2) rgba (7.0, 2.0, 8.0, 5.0)
@@ -223,7 +223,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgba (1.0, 1.0, 1.0, 1.0)
 probe uav 1 (1) rgba (3.5, 1.0, 4.0, 2.5)
 probe uav 1 (2) rgba (12.25, 1.0, 16.0, 6.25)
diff --git a/tests/hlsl/wave-ops-int.shader_test b/tests/hlsl/wave-ops-int.shader_test
index 5c9a365f..bc6976ba 100644
--- a/tests/hlsl/wave-ops-int.shader_test
+++ b/tests/hlsl/wave-ops-int.shader_test
@@ -31,7 +31,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgbai (-3, 1, -4, 2)
 probe uav 1 (1) rgbai (-3, 1, -4, 2)
 probe uav 1 (2) rgbai (-3, 1, -4, 2)
@@ -49,7 +49,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rgbai (3, 5, -1, 2)
 probe uav 1 (1) rgbai (3, 5, -1, 2)
 probe uav 1 (2) rgbai (3, 5, -1, 2)
diff --git a/tests/hlsl/wave-ops-uint.shader_test b/tests/hlsl/wave-ops-uint.shader_test
index cb502e86..c16adab4 100644
--- a/tests/hlsl/wave-ops-uint.shader_test
+++ b/tests/hlsl/wave-ops-uint.shader_test
@@ -101,7 +101,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rui (8)
 probe uav 1 (1) rui (8)
 probe uav 1 (2) rui (8)
@@ -214,7 +214,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rui (41)
 probe uav 1 (1) rui (41)
 probe uav 1 (2) rui (41)
@@ -298,7 +298,7 @@ void main(uint id : SV_GroupIndex)

[test]
 uniform 0 uint4 1 5 0 0
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 1 (0) rui (29)
 probe uav 1 (1) rui (29)
 probe uav 1 (2) rui (29)
diff --git a/tests/hlsl/wave-reconvergence.shader_test b/tests/hlsl/wave-reconvergence.shader_test
index 66698da8..0a1da3ee 100644
--- a/tests/hlsl/wave-reconvergence.shader_test
+++ b/tests/hlsl/wave-reconvergence.shader_test
@@ -56,7 +56,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 2 (0) rui (0xf01)
 probe uav 2 (1) rui (0x503)
 probe uav 2 (2) rui (0x511)
@@ -124,7 +124,7 @@ void main(uint id : SV_GroupIndex)
 }

[test]
-todo dispatch 4 1 1
+dispatch 4 1 1
 probe uav 2 (0) rui (0x701)
 probe uav 2 (1) rui (0x503)
 probe uav 2 (2) rui (0x511)