[SYCL][ESIMD] Move fmax to SPIR-V intrinsic (#14020)

Couldn't do this earlier because of a driver issue. Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>
intel · Jul 8, 2024 · 1f1be9c · 1f1be9c
1 parent bd31be7
commit 1f1be9c
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 7 deletions.
diff --git a/sycl/include/sycl/ext/intel/esimd/detail/math_intrin.hpp b/sycl/include/sycl/ext/intel/esimd/detail/math_intrin.hpp
@@ -85,6 +85,10 @@ __ESIMD_INTRIN __ESIMD_raw_vec_t(T, N)
     __spirv_ocl_fmin(__ESIMD_raw_vec_t(T, N),
                      __ESIMD_raw_vec_t(T, N)) __ESIMD_INTRIN_END;
 
+template <typename T, int N>
+__ESIMD_INTRIN __ESIMD_raw_vec_t(T, N)
+    __spirv_ocl_fmax(__ESIMD_raw_vec_t(T, N),
+                     __ESIMD_raw_vec_t(T, N)) __ESIMD_INTRIN_END;
 // saturation intrinsics
 template <typename T0, typename T1, int SZ>
 __ESIMD_INTRIN __ESIMD_raw_vec_t(T0, SZ)
@@ -114,11 +118,7 @@ template <typename T0, typename T1, int SZ>
 __ESIMD_INTRIN __ESIMD_raw_vec_t(T0, SZ)
     __esimd_sstrunc_sat(__ESIMD_raw_vec_t(T1, SZ) src) __ESIMD_INTRIN_END;
 
-/// 3 kinds of max
-template <typename T, int SZ>
-__ESIMD_INTRIN __ESIMD_raw_vec_t(T, SZ)
-    __esimd_fmax(__ESIMD_raw_vec_t(T, SZ) src0,
-                 __ESIMD_raw_vec_t(T, SZ) src1) __ESIMD_INTRIN_END;
+/// 3 kinds of max, the missing fmax uses spir-v intrinsics above
 template <typename T, int SZ>
 __ESIMD_INTRIN __ESIMD_raw_vec_t(T, SZ)
     __esimd_umax(__ESIMD_raw_vec_t(T, SZ) src0,

diff --git a/sycl/include/sycl/ext/intel/esimd/math.hpp b/sycl/include/sycl/ext/intel/esimd/math.hpp
@@ -185,7 +185,7 @@ __ESIMD_API simd<T, SZ>(max)(simd<T, SZ> src0, simd<T, SZ> src1, Sat sat = {}) {
   constexpr bool is_sat = std::is_same_v<Sat, saturation_on_tag>;
 
   if constexpr (std::is_floating_point<T>::value) {
-    auto Result = __esimd_fmax<T, SZ>(src0.data(), src1.data());
+    auto Result = __spirv_ocl_fmax<T, SZ>(src0.data(), src1.data());
     if constexpr (is_sat)
       Result = __esimd_sat<T, T, SZ>(Result);
     return simd<T, SZ>(Result);
@@ -1466,7 +1466,7 @@ template <typename T0, typename T1, int SZ> struct esimd_apply_reduced_max {
   template <typename... T>
   simd<T0, SZ> operator()(simd<T1, SZ> v1, simd<T1, SZ> v2) {
     if constexpr (std::is_floating_point<T1>::value) {
-      return __esimd_fmax<T1, SZ>(v1.data(), v2.data());
+      return __spirv_ocl_fmax<T1, SZ>(v1.data(), v2.data());
     } else if constexpr (std::is_unsigned<T1>::value) {
       return __esimd_umax<T1, SZ>(v1.data(), v2.data());
     } else {