Marat Dukhan | f9cf55d | 2021-12-09 18:54:00 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/qs8-f32-vcvt/sse4.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2021 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <immintrin.h> |
| 13 | |
| 14 | #include <xnnpack/common.h> |
| 15 | #include <xnnpack/intrinsics-polyfill.h> |
| 16 | #include <xnnpack/vcvt.h> |
| 17 | |
| 18 | |
| 19 | void xnn_qs8_f32_vcvt_ukernel__sse41_x16( |
| 20 | size_t n, |
| 21 | const int8_t* x, |
| 22 | float* y, |
| 23 | const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| 24 | { |
| 25 | assert(n != 0); |
| 26 | assert(n % sizeof(int8_t) == 0); |
| 27 | assert(x != NULL); |
| 28 | assert(y != NULL); |
| 29 | |
| 30 | const __m128i vminus_zero_point = _mm_load_si128((const __m128i*) params->sse4.minus_zero_point); |
| 31 | const __m128 vscale = _mm_load_ps(params->sse4.scale); |
| 32 | for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) { |
| 33 | __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(x)); |
| 34 | __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(x + 4)); |
| 35 | __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(x + 8)); |
| 36 | __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(x + 12)); |
| 37 | x += 16; |
| 38 | |
| 39 | vx0123 = _mm_add_epi32(vx0123, vminus_zero_point); |
| 40 | vx4567 = _mm_add_epi32(vx4567, vminus_zero_point); |
| 41 | vx89AB = _mm_add_epi32(vx89AB, vminus_zero_point); |
| 42 | vxCDEF = _mm_add_epi32(vxCDEF, vminus_zero_point); |
| 43 | |
| 44 | __m128 vy0123 = _mm_cvtepi32_ps(vx0123); |
| 45 | __m128 vy4567 = _mm_cvtepi32_ps(vx4567); |
| 46 | __m128 vy89AB = _mm_cvtepi32_ps(vx89AB); |
| 47 | __m128 vyCDEF = _mm_cvtepi32_ps(vxCDEF); |
| 48 | |
| 49 | vy0123 = _mm_mul_ps(vy0123, vscale); |
| 50 | vy4567 = _mm_mul_ps(vy4567, vscale); |
| 51 | vy89AB = _mm_mul_ps(vy89AB, vscale); |
| 52 | vyCDEF = _mm_mul_ps(vyCDEF, vscale); |
| 53 | |
| 54 | _mm_storeu_ps(y, vy0123); |
| 55 | _mm_storeu_ps(y + 4, vy4567); |
| 56 | _mm_storeu_ps(y + 8, vy89AB); |
| 57 | _mm_storeu_ps(y + 12, vyCDEF); |
| 58 | y += 16; |
| 59 | } |
| 60 | for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_t)) { |
| 61 | __m128i vx = _mm_cvtepi8_epi32(_mm_loadu_si32(x)); |
| 62 | vx = _mm_add_epi32(vx, vminus_zero_point); |
| 63 | x += 4; |
| 64 | |
| 65 | __m128 vy = _mm_cvtepi32_ps(vx); |
| 66 | vy = _mm_mul_ps(vy, vscale); |
| 67 | |
| 68 | _mm_storeu_ps(y, vy); |
| 69 | y += 4; |
| 70 | } |
| 71 | if XNN_UNLIKELY(n != 0) { |
| 72 | assert(n >= 1 * sizeof(int8_t)); |
| 73 | assert(n <= 3 * sizeof(int8_t)); |
| 74 | |
| 75 | __m128i vx = _mm_cvtepi8_epi32(_mm_loadu_si32(x)); |
| 76 | vx = _mm_add_epi32(vx, vminus_zero_point); |
| 77 | |
| 78 | __m128 vy = _mm_cvtepi32_ps(vx); |
| 79 | vy = _mm_mul_ps(vy, vscale); |
| 80 | |
| 81 | if (n & (2 * sizeof(int8_t))) { |
| 82 | _mm_storel_pi((__m64*) y, vy); |
| 83 | vy = _mm_movehl_ps(vy, vy); |
| 84 | y += 2; |
| 85 | } |
| 86 | if (n & (1 * sizeof(int8_t))) { |
| 87 | _mm_store_ss(y, vy); |
| 88 | } |
| 89 | } |
| 90 | } |