Support builtins for min and max operations in CMCL

.

Author: vmustya

IGC/VectorCompiler/CMCL/lib/Headers/cm-cl/detail/builtins.h

@@ -118,6 +118,11 @@ template <typename T> uint32_t __cm_cl_cbit(T src);
 template <typename T, int width>
 vector_impl<uint32_t, width> __cm_cl_cbit(vector_impl<T, width> src);
 
+template <typename T> T __cm_cl_smin(T src0, T src1);
+template <typename T> T __cm_cl_smax(T src0, T src1);
+template <typename T> T __cm_cl_umin(T src0, T src1);
+template <typename T> T __cm_cl_umax(T src0, T src1);
+
 template <typename T> T __cm_cl_fma(T src0, T src1, T src2);
 
 uint32_t __cm_cl_bfrev(uint32_t src);
@@ -394,6 +399,30 @@ template <typename T> T absolute(T src) {
   return src;
 }
 
+template <typename T> T min(T src0, T src1) {
+  static_assert(cl::is_arithmetic<T>::value,
+                "Min only supports arithmetic types");
+  if constexpr (cl::is_floating_point<T>::value)
+    return __cm_cl_minnum(src0, src1);
+
+  if constexpr (cl::is_signed<T>::value)
+    return __cm_cl_smin(src0, src1);
+
+  return __cm_cl_umin(src0, src1);
+}
+
+template <typename T> T max(T src0, T src1) {
+  static_assert(cl::is_arithmetic<T>::value,
+                "Max only supports arithmetic types");
+  if constexpr (cl::is_floating_point<T>::value)
+    return __cm_cl_maxnum(src0, src1);
+
+  if constexpr (cl::is_signed<T>::value)
+    return __cm_cl_smax(src0, src1);
+
+  return __cm_cl_umax(src0, src1);
+}
+
 template <typename T> T ceil(T src) {
   static_assert(cl::is_floating_point<T>::value,
                 "Ceil function expects floating poing type.");
@@ -418,18 +447,6 @@ template <typename T> T roundne(T src) {
   return __cm_cl_roundne(src);
 }
 
-template <typename T> T min_float(T src0, T src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "illegal type provided in min_float");
-  return __cm_cl_minnum(src0, src1);
-}
-
-template <typename T> T max_float(T src0, T src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "illegal type provided in max_float");
-  return __cm_cl_maxnum(src0, src1);
-}
-
 template <bool use_fast, typename T> T sqrt(T src) {
   static_assert(cl::is_floating_point<T>::value,
                 "illegal type provided in sqrt");

IGC/VectorCompiler/CMCL/lib/Headers/cm-cl/math.h

@@ -271,28 +271,20 @@ vector<T, width> square_root(vector<T, width> src, cm::tag::fast_t) {
 
 template <typename T, int width>
 vector<T, width> minimum(vector<T, width> src0, vector<T, width> src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "only floating point types supported yet for minimum");
-  return detail::min_float(src0.cl_vector(), src1.cl_vector());
+  return detail::min(src0.cl_vector(), src1.cl_vector());
 }
 
 template <typename T, int width>
 vector<T, width> maximum(vector<T, width> src0, vector<T, width> src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "only floating point types supported yet for maximum");
-  return detail::max_float(src0.cl_vector(), src1.cl_vector());
+  return detail::max(src0.cl_vector(), src1.cl_vector());
 }
 
 template <typename T> T minimum(T src0, T src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "only floating point types supported yet for minimum");
-  return detail::min_float(src0, src1);
+  return detail::min(src0, src1);
 }
 
 template <typename T> T maximum(T src0, T src1) {
-  static_assert(cl::is_floating_point<T>::value,
-                "only floating point types supported yet for maximum");
-  return detail::max_float(src0, src1);
+  return detail::max(src0, src1);
 }
 
 /*====================== log_base_2 =====================*/

IGC/VectorCompiler/CMCL/lib/Headers/opencl_type_traits.h

@@ -101,7 +101,8 @@ template <typename T> struct pointer_traits<__constant T *> {
 template <typename T>
 struct is_floating_point
     : integral_constant<
-          bool, is_same<float, typename remove_cv<T>::type>::value ||
+          bool, is_same<half, typename remove_cv<T>::type>::value ||
+                    is_same<float, typename remove_cv<T>::type>::value ||
                     is_same<double, typename remove_cv<T>::type>::value ||
                     is_same<long double, typename remove_cv<T>::type>::value> {
 };

IGC/VectorCompiler/CMCL/lib/Support/TranslationDescription.json

@@ -699,6 +699,66 @@
         ]
       }
     },
+    "SMin": {
+      "Name": "smin",
+      "Operands": [
+        {"Name": "Source0", "Kind": "Input"},
+        {"Name": "Source1", "Kind": "Input"}
+      ],
+      "TranslateInto": {
+        "LLVM-Intrinsic": "smin",
+        "ReturnType": {"GetBuiltinReturnType": []},
+        "Operands": [
+          {"GetBuiltinOperand": ["Source0"]},
+          {"GetBuiltinOperand": ["Source1"]}
+        ]
+      }
+    },
+    "SMax": {
+      "Name": "smax",
+      "Operands": [
+        {"Name": "Source0", "Kind": "Input"},
+        {"Name": "Source1", "Kind": "Input"}
+      ],
+      "TranslateInto": {
+        "LLVM-Intrinsic": "smax",
+        "ReturnType": {"GetBuiltinReturnType": []},
+        "Operands": [
+          {"GetBuiltinOperand": ["Source0"]},
+          {"GetBuiltinOperand": ["Source1"]}
+        ]
+      }
+    },
+    "UMin": {
+      "Name": "umin",
+      "Operands": [
+        {"Name": "Source0", "Kind": "Input"},
+        {"Name": "Source1", "Kind": "Input"}
+      ],
+      "TranslateInto": {
+        "LLVM-Intrinsic": "umin",
+        "ReturnType": {"GetBuiltinReturnType": []},
+        "Operands": [
+          {"GetBuiltinOperand": ["Source0"]},
+          {"GetBuiltinOperand": ["Source1"]}
+        ]
+      }
+    },
+    "UMax": {
+      "Name": "umax",
+      "Operands": [
+        {"Name": "Source0", "Kind": "Input"},
+        {"Name": "Source1", "Kind": "Input"}
+      ],
+      "TranslateInto": {
+        "LLVM-Intrinsic": "umax",
+        "ReturnType": {"GetBuiltinReturnType": []},
+        "Operands": [
+          {"GetBuiltinOperand": ["Source0"]},
+          {"GetBuiltinOperand": ["Source1"]}
+        ]
+      }
+    },
     "Log2": {
       "Name": "log2",
       "Operands": [