delete clc_sw_fma

Author: wenju-he

libclc/clc/include/clc/internal/math/clc_sw_fma.h

@@ -23,15 +23,6 @@
 #define PNOR 0x100
 #define PINF 0x200
 
-#if (defined __AMDGCN__ || defined __R600__) && !defined __HAS_FMAF__
-#define __CLC_HAVE_HW_FMA32() (0)
-#elif defined(CLC_SPIRV)
-bool __attribute__((noinline)) __clc_runtime_has_hw_fma32(void);
-#define __CLC_HAVE_HW_FMA32() __clc_runtime_has_hw_fma32()
-#else
-#define __CLC_HAVE_HW_FMA32() (1)
-#endif
-
 #define HAVE_BITALIGN() (0)
 #define HAVE_FAST_FMA32() (0)
 

libclc/clc/include/clc/math/math.h

@@ -1,2 +1 @@
-math/clc_sw_fma.cl
 integer/clc_mul_hi.cl

libclc/clc/lib/clspv/SOURCES

@@ -138,7 +138,6 @@ math/clc_sinh.cl
 math/clc_sinpi.cl
 math/clc_sqrt.cl
 math/clc_subnormal_config.cl
-math/clc_sw_fma.cl
 math/clc_tables.cl
 math/clc_tan.cl
 math/clc_tanh.cl

libclc/clc/lib/clspv/math/clc_sw_fma.cl

@@ -8,9 +8,5 @@
 
 _CLC_DEF _CLC_OVERLOAD __CLC_GENTYPE __clc_fma(__CLC_GENTYPE a, __CLC_GENTYPE b,
                                                __CLC_GENTYPE c) {
-#if __CLC_FPSIZE == 32
-  if (!__CLC_HAVE_HW_FMA32())
-    return __clc_sw_fma(a, b, c);
-#endif
   return __builtin_elementwise_fma(a, b, c);
 }

libclc/clc/lib/generic/SOURCES

@@ -97,19 +97,9 @@ _CLC_DEF _CLC_OVERLOAD void __clc_fullMulS(private __CLC_FLOATN *hi,
                                            private __CLC_FLOATN *lo,
                                            __CLC_FLOATN a, __CLC_FLOATN b,
                                            __CLC_FLOATN bh, __CLC_FLOATN bt) {
-  if (__CLC_HAVE_HW_FMA32()) {
-    __CLC_FLOATN ph = a * b;
-    *hi = ph;
-    *lo = __clc_fma(a, b, -ph);
-  } else {
-    __CLC_FLOATN ah = __CLC_AS_FLOATN(__CLC_AS_UINTN(a) & 0xfffff000U);
-    __CLC_FLOATN at = a - ah;
-    __CLC_FLOATN ph = a * b;
-    __CLC_FLOATN pt = __clc_mad(
-        at, bt, __clc_mad(at, bh, __clc_mad(ah, bt, __clc_mad(ah, bh, -ph))));
-    *hi = ph;
-    *lo = pt;
-  }
+  __CLC_FLOATN ph = a * b;
+  *hi = ph;
+  *lo = __clc_fma(a, b, -ph);
 }
 
 _CLC_DEF _CLC_OVERLOAD __CLC_FLOATN __clc_removePi2S(private __CLC_FLOATN *hi,
@@ -280,20 +270,8 @@ _CLC_DEF _CLC_OVERLOAD __CLC_INTN __clc_argReductionLargeS(
   const __CLC_FLOATN pio2t = (__CLC_FLOATN)0xa22168 / 0x1.0p+47f;
 
   __CLC_FLOATN rh, rt;
-
-  if (__CLC_HAVE_HW_FMA32()) {
-    rh = q1 * pio2h;
-    rt = __clc_fma(q0, pio2h, __clc_fma(q1, pio2t, __clc_fma(q1, pio2h, -rh)));
-  } else {
-    __CLC_FLOATN q1h = __CLC_AS_FLOATN(__CLC_AS_UINTN(q1) & 0xfffff000);
-    __CLC_FLOATN q1t = q1 - q1h;
-    rh = q1 * pio2h;
-    rt = __clc_mad(
-        q1t, pio2ht,
-        __clc_mad(q1t, pio2hh,
-                  __clc_mad(q1h, pio2ht, __clc_mad(q1h, pio2hh, -rh))));
-    rt = __clc_mad(q0, pio2h, __clc_mad(q1, pio2t, rt));
-  }
+  rh = q1 * pio2h;
+  rt = __clc_fma(q0, pio2h, __clc_fma(q1, pio2t, __clc_fma(q1, pio2h, -rh)));
 
   __CLC_FLOATN t = rh + rt;
   rt = rt - (t - rh);