optimize upsample

Author: frasercrmck

libclc/clc/lib/generic/integer/clc_upsample.cl

@@ -1,35 +1,24 @@
 #include <clc/internal/clc.h>
 
+// TODO: Replace with __clc_convert_<type> when available
+#define __CLC_CONVERT_TY(X, TY) __builtin_convertvector(X, TY)
+
+#define __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE, GENTYPE, UGENTYPE, GENSIZE)          \
+  _CLC_OVERLOAD _CLC_DEF BGENTYPE __clc_upsample(GENTYPE hi, UGENTYPE lo) {    \
+    BGENTYPE large_hi = __CLC_CONVERT_TY(hi, BGENTYPE);                        \
+    BGENTYPE large_lo = __CLC_CONVERT_TY(lo, BGENTYPE);                        \
+    return (large_hi << (BGENTYPE)GENSIZE) | large_lo;                         \
+  }
+
 #define __CLC_UPSAMPLE_IMPL(BGENTYPE, GENTYPE, UGENTYPE, GENSIZE)              \
   _CLC_OVERLOAD _CLC_DEF BGENTYPE __clc_upsample(GENTYPE hi, UGENTYPE lo) {    \
     return ((BGENTYPE)hi << GENSIZE) | lo;                                     \
   }                                                                            \
-  _CLC_OVERLOAD _CLC_DEF BGENTYPE##2 __clc_upsample(GENTYPE##2 hi,             \
-                                                    UGENTYPE##2 lo) {          \
-    return (BGENTYPE##2){__clc_upsample(hi.s0, lo.s0),                         \
-                         __clc_upsample(hi.s1, lo.s1)};                        \
-  }                                                                            \
-  _CLC_OVERLOAD _CLC_DEF BGENTYPE##3 __clc_upsample(GENTYPE##3 hi,             \
-                                                    UGENTYPE##3 lo) {          \
-    return (BGENTYPE##3){__clc_upsample(hi.s0, lo.s0),                         \
-                         __clc_upsample(hi.s1, lo.s1),                         \
-                         __clc_upsample(hi.s2, lo.s2)};                        \
-  }                                                                            \
-  _CLC_OVERLOAD _CLC_DEF BGENTYPE##4 __clc_upsample(GENTYPE##4 hi,             \
-                                                    UGENTYPE##4 lo) {          \
-    return (BGENTYPE##4){__clc_upsample(hi.lo, lo.lo),                         \
-                         __clc_upsample(hi.hi, lo.hi)};                        \
-  }                                                                            \
-  _CLC_OVERLOAD _CLC_DEF BGENTYPE##8 __clc_upsample(GENTYPE##8 hi,             \
-                                                    UGENTYPE##8 lo) {          \
-    return (BGENTYPE##8){__clc_upsample(hi.lo, lo.lo),                         \
-                         __clc_upsample(hi.hi, lo.hi)};                        \
-  }                                                                            \
-  _CLC_OVERLOAD _CLC_DEF BGENTYPE##16 __clc_upsample(GENTYPE##16 hi,           \
-                                                     UGENTYPE##16 lo) {        \
-    return (BGENTYPE##16){__clc_upsample(hi.lo, lo.lo),                        \
-                          __clc_upsample(hi.hi, lo.hi)};                       \
-  }
+  __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE##2, GENTYPE##2, UGENTYPE##2, GENSIZE)       \
+  __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE##3, GENTYPE##3, UGENTYPE##3, GENSIZE)       \
+  __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE##4, GENTYPE##4, UGENTYPE##4, GENSIZE)       \
+  __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE##8, GENTYPE##8, UGENTYPE##8, GENSIZE)       \
+  __CLC_UPSAMPLE_VEC_IMPL(BGENTYPE##16, GENTYPE##16, UGENTYPE##16, GENSIZE)
 
 #define __CLC_UPSAMPLE_TYPES()                                                 \
   __CLC_UPSAMPLE_IMPL(short, char, uchar, 8)                                   \
@@ -43,3 +32,4 @@ __CLC_UPSAMPLE_TYPES()
 
 #undef __CLC_UPSAMPLE_TYPES
 #undef __CLC_UPSAMPLE_IMPL
+#undef __CLC_CONVERT_TY