// 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 void xnn_math_f32_sqrt__neon_nr3rsqrts( size_t n, const float* input, float* output) { assert(n % (4 * sizeof(float)) == 0); for (; n != 0; n -= 4 * sizeof(float)) { const float32x4_t vx = vld1q_f32(input); input += 4; // Initial approximation float32x4_t vrsqrtx = vrsqrteq_f32(vx); // Netwon-Raphson iteration: rsqrt_x <- rsqrt_x * ((3 - x * rsqrt_x * rsqrt_x) / 2) // Note: x * (rsqrt_x * rsqrt_x) for the first iteration and (x * rsqrt_x) * rsqrt_x for the next two improves accuracy // Note: vrsqrtsq_f32(x, y) := (3 - x * y) / 2 vrsqrtx = vmulq_f32(vrsqrtx, vrsqrtsq_f32(vx, vmulq_f32(vrsqrtx, vrsqrtx))); vrsqrtx = vmulq_f32(vrsqrtx, vrsqrtsq_f32(vmulq_f32(vrsqrtx, vx), vrsqrtx)); vrsqrtx = vmulq_f32(vrsqrtx, vrsqrtsq_f32(vmulq_f32(vrsqrtx, vx), vrsqrtx)); // Reconstruct sqrt(x) = rsqrt(x) * x const float32x4_t vy = vmulq_f32(vrsqrtx, vx); vst1q_f32(output, vy); output += 4; } }