[clang] [llvm] [HLSL] Implement dot2add intrinsic (PR #131237)

[Ashley Coleman via llvm-commits]

================
@@ -0,0 +1,78 @@
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -triple \
+// RUN:   dxil-pc-shadermodel6.3-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-DXIL
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -triple \
+// RUN:   spirv-pc-vulkan-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-SPIRV
+
+// Test basic lowering to runtime function call.
+
+// CHECK-LABEL: define {{.*}}test
+float test(half2 p1, half2 p2, float p3) {
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %1, <2 x half> {{.*}} %2, float {{.*}} %3) #3 {{.*}}
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %0, <2 x half> {{.*}} %1, float {{.*}} %2) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg2_type
+float test_float_arg2_type(half2 p1, float2 p2, float p3) {
+  // CHECK-SPIRV:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %2 to <2 x half>
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %1, <2 x half> {{.*}} %conv, float {{.*}} %3) #3 {{.*}}
+  // CHECK-DXIL:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %1 to <2 x half>
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %0, <2 x half> {{.*}} %conv, float {{.*}} %2) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg1_type
+float test_float_arg1_type(float2 p1, half2 p2, float p3) {
+  // CHECK-SPIRV:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %1 to <2 x half>
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %2, float {{.*}} %3) #3 {{.*}}
+  // CHECK-DXIL:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %0 to <2 x half>
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %1, float {{.*}} %2) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_double_arg3_type
+float test_double_arg3_type(half2 p1, half2 p2, double p3) {
+  // CHECK-SPIRV:  %conv = fptrunc reassoc nnan ninf nsz arcp afn double %3 to float
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %1, <2 x half> {{.*}} %2, float {{.*}} %conv) #3 {{.*}}
+  // CHECK-DXIL:  %conv = fptrunc reassoc nnan ninf nsz arcp afn double %2 to float
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %0, <2 x half> {{.*}} %1, float {{.*}} %conv) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg1_arg2_type
+float test_float_arg1_arg2_type(float2 p1, float2 p2, float p3) {
+  // CHECK-SPIRV:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %1 to <2 x half>
+  // CHECK-SPIRV:  %conv1 = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %2 to <2 x half>
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %conv1, float {{.*}} %3) #3 {{.*}}
+  // CHECK-DXIL:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %0 to <2 x half>
+  // CHECK-DXIL:  %conv1 = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %1 to <2 x half>
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %conv1, float {{.*}} %2) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_int16_arg1_arg2_type
+float test_int16_arg1_arg2_type(int16_t2 p1, int16_t2 p2, float p3) {
+  // CHECK-SPIRV:  %conv = sitofp <2 x i16> %1 to <2 x half>
+  // CHECK-SPIRV:  %conv1 = sitofp <2 x i16> %2 to <2 x half>
+  // CHECK-SPIRV:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %conv1, float {{.*}} %3) #3 {{.*}}
+  // CHECK-DXIL:  %conv = sitofp <2 x i16> %0 to <2 x half>
+  // CHECK-DXIL:  %conv1 = sitofp <2 x i16> %1 to <2 x half>
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @_ZN4hlsl7dot2addEDv2_DhS0_f(<2 x half> {{.*}} %conv, <2 x half> {{.*}} %conv1, float {{.*}} %2) #2
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}dot2add_impl
----------------
V-FEXrt wrote:

> I'll let others weigh in on this

I suggested that Sumit do it this way but that was mainly because I forgot about inlining. I do think enabling the passes risks testing the wrong thing/hiding an unintentional output but I also agree that it would be better to directly test that the opcode of interest is emitted

> But I also understand that for dot2add we would be repeating these same checks verbatim for each test case.
imo its okay to repeat a lot in test so this isn't a strong argument to keep the current thing

https://github.com/llvm/llvm-project/pull/131237