// Copyright 2020 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include #include #include void xnn_math_f32_roundd__scalar_cvt( size_t n, const float* input, float* output) { assert(n % sizeof(float) == 0); // Threshold of non-integral values in single-precision floating-point representation. // All inputs above this threshold (by absolute value) are integer numbers. const float vintegral_threshold = 0x1.000000p+23f; // Unit constant to decrement results rounded "wrong way" (i.e. up) in the round-towards-zero operation. const float vone = 1.0f; for (; n != 0; n -= sizeof(float)) { const float vx = *input++; // Convert floating-point value x to integer, with rounding towards zero, and then back to floating-point. // Note: the result is valid only for abs(x) < 2**31, but we further restrict its use to 2**23. const float vprerndx = (float) (int32_t) vx; // Compute abs(x) to check if the FP->INT->FP conversion result is valid. const float vabsx = fabsf(vx); // Select between the x rounded via FP->INT->FP conversion and the original x value. const float vrndx = XNN_UNPREDICTABLE(vabsx < vintegral_threshold) ? vprerndx : vx; // Restore the sign of -0.0f lost in the FP->INT->FP conversion. const float vadjrndx = copysignf(vrndx, vx); // Adjust x rounded towards zero to get x rounded down. // Note: addition implicitly converts SNaN inputs to QNaNs. const float vy = XNN_UNPREDICTABLE(vrndx <= vx) ? vadjrndx : vrndx - vone; *output++ = vy; } }