[SYCL][libdevice] Add SIMD emulation APIs for imf libdevice (#6327)

* [SYCL][libdevice] Add SIMD emulation APIs for imf libdevice
* Add __imf_vcmp* functions part1
* add __imf_vset* APIs
* Add __imf_vsadu2/4 __imf_vsads2/4
* Add all SIMD APIs
* fix incorrect __devicelib_imf_vhaddu2 name
* Fix bug in __imf_vcmplt4

Signed-off-by: jinge90 <[email protected]>

Author: jinge90

libdevice/cmake/modules/ImfSrcConcate.cmake

@@ -1,6 +1,7 @@
 set(imf_fp32_fallback_src_list imf_utils/integer_misc.cpp
                                imf_utils/half_convert.cpp
                                imf_utils/float_convert.cpp
+                               imf_utils/simd_emulate.cpp
                                imf/imf_inline_fp32.cpp)
 
 set(imf_fp64_fallback_src_list imf_utils/double_convert.cpp

libdevice/cmake/modules/SYCLLibdevice.cmake

@@ -123,6 +123,7 @@ set(imf_fallback_fp32_deps device.h device_imf.hpp imf_half.hpp
                            imf_utils/integer_misc.cpp
                            imf_utils/float_convert.cpp
                            imf_utils/half_convert.cpp
+                           imf_utils/simd_emulate.cpp
                            imf/imf_inline_fp32.cpp)
 set(imf_fallback_fp64_deps device.h device_imf.hpp imf_half.hpp
                            imf_utils/double_convert.cpp

libdevice/device_imf.hpp

@@ -12,8 +12,8 @@
 #include "device.h"
 #include "imf_half.hpp"
 #include <cstddef>
+#include <limits>
 #include <type_traits>
-
 #ifdef __LIBDEVICE_IMF_ENABLED__
 
 #if !defined(__SPIR__) && !defined(__LIBDEVICE_HOST_IMPL__)
@@ -458,12 +458,21 @@ static inline int __popcll(unsigned long long int x) {
 #endif
 }
 
-static inline unsigned int __abs(int x) { return x < 0 ? -x : x; }
-
-static inline unsigned long long int __abs(long long int x) {
+template <typename T>
+static inline typename std::make_unsigned<T>::type __abs(T x) {
+  static_assert((std::is_signed<T>::value && std::is_integral<T>::value),
+                "__abs can only accept signed integral type.");
   return x < 0 ? -x : x;
 }
 
+template <typename T> static inline void __swap(T &x, T &y) {
+  static_assert(std::is_integral<T>::value,
+                "__swap can only accept integral type.");
+  T tmp = x;
+  x = y;
+  y = tmp;
+}
+
 template <typename Ty1, typename Ty2>
 static inline Ty2 __get_bytes_by_index(Ty1 x, size_t idx) {
   static_assert(!std::is_signed<Ty1>::value && !std::is_signed<Ty2>::value,