Add support for wasm dot-product instruction (#5861)

* Add support for wasm dot-product instruction

Author: steven-johnson

src/CodeGen_WebAssembly.cpp

@@ -102,6 +102,8 @@ const WasmIntrinsic intrinsic_defs[] = {
     {"saturating_narrow_i16x16_to_u8x16", UInt(8, 16), "saturating_narrow", {Int(16, 16)}, Target::WasmSimd128},
     {"saturating_narrow_i32x8_to_i16x8", Int(16, 8), "saturating_narrow", {Int(32, 8)}, Target::WasmSimd128},
     {"saturating_narrow_i32x8_to_u16x8", UInt(16, 8), "saturating_narrow", {Int(32, 8)}, Target::WasmSimd128},
+
+    {"llvm.wasm.dot", Int(32, 4), "dot_product", {Int(16, 8), Int(16, 8)}, Target::WasmSimd128},
 #endif
 };
 // clang-format on
@@ -185,6 +187,8 @@ void CodeGen_WebAssembly::codegen_vector_reduce(const VectorReduce *op, const Ex
         {VectorReduce::Add, 2, i32(wild_i16x_), "pairwise_widening_add", Target::WasmSimd128},
         {VectorReduce::Add, 2, u32(wild_u16x_), "pairwise_widening_add", Target::WasmSimd128},
         {VectorReduce::Add, 2, i32(wild_u16x_), "pairwise_widening_add", Target::WasmSimd128},
+
+        {VectorReduce::Add, 2, i32(widening_mul(wild_i16x_, wild_i16x_)), "dot_product", Target::WasmSimd128},
     };
     // clang-format on
 

test/correctness/simd_op_check.cpp

@@ -1753,12 +1753,15 @@ class SimdOpCheck : public SimdOpCheckTest {
                 check("i32x4.mul", 4 * w, i32_1 * i32_2);
                 check("i64x2.mul", 2 * w, i64_1 * i64_2);
 
-                // Integer dot product (16 -> 32)
-                // TODO(https://github.com/halide/Halide/issues/5130): NOT BEING GENERATED AT TRUNK
-                // {
-                //     RDom r(0, 4);
-                //     check("i32x4.dot_i16x8_s", 2 * w, sum(i32(in_i16(x * 4 + r)) * in_i16(x * 4 + r + 32)));
-                // }
+                if (Halide::Internal::get_llvm_version() >= 130) {
+                    // Integer dot product (16 -> 32)
+                    for (int f : {2, 4, 8}) {
+                        RDom r(0, f);
+                        for (int v : {1, 2, 4}) {
+                            check("i32x4.dot_i16x8_s", w * v, sum(i32(in_i16(f * x + r)) * in_i16(f * x + r + 32)));
+                        }
+                    }
+                }
 
                 // Integer negation
                 check("i8x16.neg", 16 * w, -i8_1);