| // Copyright 2022 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 <assert.h> |
| #include <stddef.h> |
| |
| #include <xnnpack/math.h> |
| #include <xnnpack/math-stubs.h> |
| |
| |
| void xnn_math_u32_sqrt__scalar_clz_newton( |
| size_t n, |
| const uint32_t* input, |
| uint32_t* output) |
| { |
| assert(n % sizeof(uint32_t) == 0); |
| |
| for (; n != 0; n -= sizeof(uint32_t)) { |
| const uint32_t vx = *input++; |
| |
| uint32_t vy = vx; |
| |
| // Based on Hacker's Delight, Figure 11-1. |
| if (vx != 0) { |
| const uint32_t vs = 16 - (math_clz_nonzero_u32(vx - 1) >> 1); |
| |
| uint32_t vg0 = UINT32_C(1) << vs; |
| uint32_t vg1 = (vg0 + (vx >> vs)) >> 1; |
| while XNN_LIKELY(vg1 < vg0) { |
| vg0 = vg1; |
| vg1 = (vg0 + vx / vg0) >> 1; |
| } |
| |
| // vg0 is sqrt(vx) rounded down. Do the final rounding up if needed. |
| if XNN_UNPREDICTABLE(vg0 * vg0 < vx - vg0) { |
| vg0 += 1; |
| } |
| vy = vg0; |
| } |
| |
| *output++ = vy; |
| } |
| } |