[PATCH v4 1/2] vkd3d-shader: Add constant folding for the 'dot' operation.

[Nikolay Sivov]

From: Nikolay Sivov

Signed-off-by: Nikolay Sivov 
---
 libs/vkd3d-shader/hlsl_constant_ops.c | 32 +++++++++++++++++++++++++++
 tests/hlsl/dot.shader_test            | 13 +++++++++++
 2 files changed, 45 insertions(+)

diff --git a/libs/vkd3d-shader/hlsl_constant_ops.c b/libs/vkd3d-shader/hlsl_constant_ops.c
index 01c438ae2..5501b6e9c 100644
--- a/libs/vkd3d-shader/hlsl_constant_ops.c
+++ b/libs/vkd3d-shader/hlsl_constant_ops.c
@@ -348,6 +348,34 @@ static bool fold_bit_xor(struct hlsl_ctx *ctx, struct hlsl_constant_value *dst,
     return true;
 }

+static bool fold_dot(struct hlsl_ctx *ctx, struct hlsl_constant_value *dst, const struct hlsl_type *dst_type,
+        const struct hlsl_ir_constant *src1, const struct hlsl_ir_constant *src2)
+{
+    enum hlsl_base_type type = dst_type->base_type;
+    unsigned int k;
+
+    assert(type == src1->node.data_type->base_type);
+    assert(type == src2->node.data_type->base_type);
+    assert(src1->node.data_type->dimx == src2->node.data_type->dimx);
+
+    dst->u[0].f = 0.0f;
+    for (k = 0; k < src1->node.data_type->dimx; ++k)
+    {
+        switch (type)
+        {
+            case HLSL_TYPE_FLOAT:
+            case HLSL_TYPE_HALF:
+                dst->u[0].f += src1->value.u[k].f * src2->value.u[k].f;
+                break;
+            default:
+                FIXME("Fold 'dot' for type %s.\n", debug_hlsl_type(ctx, dst_type));
+                return false;
+        }
+    }
+
+    return true;
+}
+
 static bool fold_div(struct hlsl_ctx *ctx, struct hlsl_constant_value *dst, const struct hlsl_type *dst_type,
         const struct hlsl_ir_constant *src1, const struct hlsl_ir_constant *src2,
         const struct vkd3d_shader_location *loc)
@@ -788,6 +816,10 @@ bool hlsl_fold_constant_exprs(struct hlsl_ctx *ctx, struct hlsl_ir_node *instr,
             success = fold_bit_xor(ctx, &res, instr->data_type, arg1, arg2);
             break;

+        case HLSL_OP2_DOT:
+            success = fold_dot(ctx, &res, instr->data_type, arg1, arg2);
+            break;
+
         case HLSL_OP2_DIV:
             success = fold_div(ctx, &res, instr->data_type, arg1, arg2, &instr->loc);
             break;
diff --git a/tests/hlsl/dot.shader_test b/tests/hlsl/dot.shader_test
index e51d7cb8c..15f120f70 100644
--- a/tests/hlsl/dot.shader_test
+++ b/tests/hlsl/dot.shader_test
@@ -74,6 +74,19 @@ uniform 4 float4 3.0 0.0 0.0 0.0
 draw quad
 probe all rgba (6.0, 6.0, 6.0, 6.0)

+[pixel shader]
+static const float4 x = float4(2.0, 3.0, 4.0, 5.0);
+static const float4 y = float4(6.0, 7.0, 8.0, 9.0);
+
+float4 main() : sv_target
+{
+    return dot(x, y);
+}
+
+[test]
+draw quad
+probe all rgba (110.0, 110.0, 110.0, 110.0)
+
 [pixel shader fail]
 uniform float1x1 x;
 uniform float4 y;