Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <assert.h> |
| 7 | #include <stddef.h> |
| 8 | |
| 9 | #include <emmintrin.h> |
| 10 | |
| 11 | #include <xnnpack/math-stubs.h> |
| 12 | |
| 13 | |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 14 | void xnn_math_f32_roundne__sse2_cvt( |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 15 | size_t n, |
| 16 | const float* input, |
| 17 | float* output) |
| 18 | { |
| 19 | assert(n % (4 * sizeof(float)) == 0); |
| 20 | |
Marat Dukhan | 03723f5 | 2020-05-13 00:59:24 -0700 | [diff] [blame^] | 21 | // This magic number serves two purposes: |
| 22 | // 1. Set the bit corresponding to the sign of a floating-point number in a bitmask. |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 23 | // 2. Check if the input to CVTPS2DQ (_mm_cvtps_epi32) is out-of-range, which results in 0x80000000 output. |
Marat Dukhan | 03723f5 | 2020-05-13 00:59:24 -0700 | [diff] [blame^] | 24 | const __m128i vmagic = _mm_set1_epi32(0x80000000); |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 25 | |
| 26 | for (; n != 0; n -= 4 * sizeof(float)) { |
| 27 | const __m128 vx = _mm_load_ps(input); |
| 28 | input += 4; |
| 29 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 30 | // Convert floating-point value x to integer, with default rounding (to nearest-even). |
| 31 | // If x is beyond [-2**31, 2**31-1] range or x is NaN, the result is -2**31 (0x80000000). |
| 32 | const __m128i vintx = _mm_cvtps_epi32(vx); |
| 33 | |
Marat Dukhan | 03723f5 | 2020-05-13 00:59:24 -0700 | [diff] [blame^] | 34 | // Compute bitmask for the bits we want to copy from the rounded x. Other bits will be copied from x. |
| 35 | // If x is out-of-range for CVTPS2DQ, we want all bits from x. |
| 36 | // If x is in-range for CVTPS2DQ, we want all but the sign bit from the rounded x and the sign bit from x. |
| 37 | const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 38 | |
| 39 | // Convert integer back to floating-point. |
| 40 | // We binary OR the result with the sign of x to restore the sign of negative zero. |
Marat Dukhan | 03723f5 | 2020-05-13 00:59:24 -0700 | [diff] [blame^] | 41 | const __m128 vrndx = _mm_cvtepi32_ps(vintx); |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 42 | |
| 43 | // Combine x rounded via conversion to integer and the initial x value. |
| 44 | // For -2**31 < x < 2**31, the result is x rounded via conversion to integer. |
| 45 | // Otherwise (including NaN inputs), the result is x itself. |
| 46 | const __m128 vy = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vrndx)); |
| 47 | |
| 48 | _mm_store_ps(output, vy); |
| 49 | output += 4; |
| 50 | } |
| 51 | } |