// // Copyright (c) 2017 The Khronos Group Inc. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. // #include "rounding_mode.h" #if (defined(__arm__) || defined(__aarch64__)) #define FPSCR_FZ (1 << 24) // Flush-To-Zero mode #define FPSCR_ROUND_MASK (3 << 22) // Rounding mode: #define _ARM_FE_FTZ 0x1000000 #define _ARM_FE_NFTZ 0x0 #if defined(__aarch64__) #define _FPU_GETCW(cw) __asm__("MRS %0,FPCR" : "=r"(cw)) #define _FPU_SETCW(cw) __asm__("MSR FPCR,%0" : : "ri"(cw)) #else #define _FPU_GETCW(cw) __asm__("VMRS %0,FPSCR" : "=r"(cw)) #define _FPU_SETCW(cw) __asm__("VMSR FPSCR,%0" : : "ri"(cw)) #endif #endif #if (defined(__arm__) || defined(__aarch64__)) && defined(__GNUC__) #define _ARM_FE_TONEAREST 0x0 #define _ARM_FE_UPWARD 0x400000 #define _ARM_FE_DOWNWARD 0x800000 #define _ARM_FE_TOWARDZERO 0xc00000 RoundingMode set_round(RoundingMode r, Type outType) { static const int flt_rounds[kRoundingModeCount] = { _ARM_FE_TONEAREST, _ARM_FE_TONEAREST, _ARM_FE_UPWARD, _ARM_FE_DOWNWARD, _ARM_FE_TOWARDZERO }; static const int int_rounds[kRoundingModeCount] = { _ARM_FE_TOWARDZERO, _ARM_FE_TONEAREST, _ARM_FE_UPWARD, _ARM_FE_DOWNWARD, _ARM_FE_TOWARDZERO }; const int *p = int_rounds; if (outType == kfloat || outType == kdouble) p = flt_rounds; int fpscr = 0; RoundingMode oldRound = get_round(); _FPU_GETCW(fpscr); _FPU_SETCW(p[r] | (fpscr & ~FPSCR_ROUND_MASK)); return oldRound; } RoundingMode get_round(void) { int fpscr; int oldRound; _FPU_GETCW(fpscr); oldRound = (fpscr & FPSCR_ROUND_MASK); switch (oldRound) { case _ARM_FE_TONEAREST: return kRoundToNearestEven; case _ARM_FE_UPWARD: return kRoundUp; case _ARM_FE_DOWNWARD: return kRoundDown; case _ARM_FE_TOWARDZERO: return kRoundTowardZero; } return kDefaultRoundingMode; } #elif !(defined(_WIN32) && defined(_MSC_VER)) RoundingMode set_round(RoundingMode r, Type outType) { static const int flt_rounds[kRoundingModeCount] = { FE_TONEAREST, FE_TONEAREST, FE_UPWARD, FE_DOWNWARD, FE_TOWARDZERO }; static const int int_rounds[kRoundingModeCount] = { FE_TOWARDZERO, FE_TONEAREST, FE_UPWARD, FE_DOWNWARD, FE_TOWARDZERO }; const int *p = int_rounds; if (outType == kfloat || outType == kdouble) p = flt_rounds; int oldRound = fegetround(); fesetround(p[r]); switch (oldRound) { case FE_TONEAREST: return kRoundToNearestEven; case FE_UPWARD: return kRoundUp; case FE_DOWNWARD: return kRoundDown; case FE_TOWARDZERO: return kRoundTowardZero; default: abort(); // ??! } return kDefaultRoundingMode; // never happens } RoundingMode get_round(void) { int oldRound = fegetround(); switch (oldRound) { case FE_TONEAREST: return kRoundToNearestEven; case FE_UPWARD: return kRoundUp; case FE_DOWNWARD: return kRoundDown; case FE_TOWARDZERO: return kRoundTowardZero; } return kDefaultRoundingMode; } #else RoundingMode set_round(RoundingMode r, Type outType) { static const int flt_rounds[kRoundingModeCount] = { _RC_NEAR, _RC_NEAR, _RC_UP, _RC_DOWN, _RC_CHOP }; static const int int_rounds[kRoundingModeCount] = { _RC_CHOP, _RC_NEAR, _RC_UP, _RC_DOWN, _RC_CHOP }; const int *p = (outType == kfloat || outType == kdouble) ? flt_rounds : int_rounds; unsigned int oldRound; int err = _controlfp_s(&oldRound, 0, 0); // get rounding mode into oldRound if (err) { vlog_error("\t\tERROR: -- cannot get rounding mode in %s:%d\n", __FILE__, __LINE__); return kDefaultRoundingMode; // what else never happens } oldRound &= _MCW_RC; RoundingMode old = (oldRound == _RC_NEAR) ? kRoundToNearestEven : (oldRound == _RC_UP) ? kRoundUp : (oldRound == _RC_DOWN) ? kRoundDown : (oldRound == _RC_CHOP) ? kRoundTowardZero : kDefaultRoundingMode; _controlfp_s(&oldRound, p[r], _MCW_RC); // setting new rounding mode return old; // returning old rounding mode } RoundingMode get_round(void) { unsigned int oldRound; int err = _controlfp_s(&oldRound, 0, 0); // get rounding mode into oldRound oldRound &= _MCW_RC; return (oldRound == _RC_NEAR) ? kRoundToNearestEven : (oldRound == _RC_UP) ? kRoundUp : (oldRound == _RC_DOWN) ? kRoundDown : (oldRound == _RC_CHOP) ? kRoundTowardZero : kDefaultRoundingMode; } #endif // // FlushToZero() sets the host processor into ftz mode. It is intended to have // a remote effect on the behavior of the code in basic_test_conversions.c. Some // host processors may not support this mode, which case you'll need to do some // clamping in software by testing against FLT_MIN or DBL_MIN in that file. // // Note: IEEE-754 says conversions are basic operations. As such they do *NOT* // have the behavior in section 7.5.3 of the OpenCL spec. They *ALWAYS* flush to // zero for subnormal inputs or outputs when FTZ mode is on like other basic // operators do (e.g. add, subtract, multiply, divide, etc.) // // Configuring hardware to FTZ mode varies by platform. // CAUTION: Some C implementations may also fail to behave properly in this // mode. // // On PowerPC, it is done by setting the FPSCR into non-IEEE mode. // On Intel, you can do this by turning on the FZ and DAZ bits in the MXCSR -- // provided that SSE/SSE2 // is used for floating point computation! If your OS uses x87, you'll // need to figure out how to turn that off for the conversions code in // basic_test_conversions.c so that they flush to zero properly. // Otherwise, you'll need to add appropriate software clamping to // basic_test_conversions.c in which case, these function are at // liberty to do nothing. // #if defined(__i386__) || defined(__x86_64__) || defined(_WIN32) #include #elif defined(__PPC__) #include #endif void *FlushToZero(void) { #if defined(__APPLE__) || defined(__linux__) || defined(_WIN32) #if defined(__i386__) || defined(__x86_64__) || defined(_MSC_VER) union { int i; void *p; } u = { _mm_getcsr() }; _mm_setcsr(u.i | 0x8040); return u.p; #elif defined(__arm__) || defined(__aarch64__) int fpscr; _FPU_GETCW(fpscr); _FPU_SETCW(fpscr | FPSCR_FZ); return NULL; #elif defined(__PPC__) fpu_control_t flags = 0; _FPU_GETCW(flags); flags |= _FPU_MASK_NI; _FPU_SETCW(flags); return NULL; #else #error Unknown arch #endif #else #error Please configure FlushToZero and UnFlushToZero to behave properly on this operating system. #endif } // Undo the effects of FlushToZero above, restoring the host to default // behavior, using the information passed in p. void UnFlushToZero(void *p) { #if defined(__APPLE__) || defined(__linux__) || defined(_WIN32) #if defined(__i386__) || defined(__x86_64__) || defined(_MSC_VER) union { void *p; int i; } u = { p }; _mm_setcsr(u.i); #elif defined(__arm__) || defined(__aarch64__) int fpscr; _FPU_GETCW(fpscr); _FPU_SETCW(fpscr & ~FPSCR_FZ); #elif defined(__PPC__) fpu_control_t flags = 0; _FPU_GETCW(flags); flags &= ~_FPU_MASK_NI; _FPU_SETCW(flags); #else #error Unknown arch #endif #else #error Please configure FlushToZero and UnFlushToZero to behave properly on this operating system. #endif }