Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <assert.h> |
| 10 | #include <stdint.h> |
| 11 | #include <stddef.h> |
| 12 | |
| 13 | #include <nmmintrin.h> |
| 14 | |
| 15 | #include <xnnpack/requantization-stubs.h> |
| 16 | |
| 17 | |
| 18 | void xnn_qs8_requantize_fp32__sse4( |
| 19 | size_t n, |
| 20 | const int32_t* input, |
| 21 | float scale, |
| 22 | int8_t zero_point, |
| 23 | int8_t qmin, |
| 24 | int8_t qmax, |
| 25 | int8_t* output) |
| 26 | { |
| 27 | assert(n % 16 == 0); |
| 28 | assert(scale < 1.0f); |
| 29 | assert(scale >= 0x1.0p-32f); |
| 30 | |
| 31 | const __m128 vscale = _mm_set1_ps(scale); |
| 32 | const __m128i vzero_point = _mm_set1_epi16((short) zero_point); |
| 33 | const __m128i vqmin = _mm_set1_epi8((char) qmin); |
| 34 | const __m128i vqmax = _mm_set1_epi8((char) qmax); |
| 35 | for (; n != 0; n -= 16) { |
| 36 | const __m128i x = _mm_loadu_si128((const __m128i*) input); |
| 37 | const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4)); |
| 38 | const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8)); |
| 39 | const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12)); |
| 40 | input += 16; |
| 41 | |
| 42 | // Convert int32_t input to FP32 and multiply by FP32 scale. |
| 43 | // Both operations involve statistically unbiased roundings (with default MXCSR rounding mode): |
| 44 | // - Large int32_t values can't be exactly represented as FP32. CVTDQ2PS instruction on x86 would round it |
| 45 | // according to nearest FP32 value with ties to even (assuming default MXCSR rounding mode). |
| 46 | // - Product of two FP32 values is generally not exactly representation as an FP32 value, and will be rounded |
| 47 | // to nearest FP32 value with ties to even with default MXCSR rounding mode. |
| 48 | const __m128 x_scaled = _mm_mul_ps(_mm_cvtepi32_ps(x), vscale); |
| 49 | const __m128 y_scaled = _mm_mul_ps(_mm_cvtepi32_ps(y), vscale); |
| 50 | const __m128 z_scaled = _mm_mul_ps(_mm_cvtepi32_ps(z), vscale); |
| 51 | const __m128 w_scaled = _mm_mul_ps(_mm_cvtepi32_ps(w), vscale); |
| 52 | |
| 53 | // Convert scaled FP32 result to int32_t using CVTPS2DQ instruction from x86 SSE2. CVTPS2DQ instruction rounds |
| 54 | // result according to nearest FP32 value with ties to even (assuming default MXCSR rounding mode). |
| 55 | // However, when conversion overflows, it produces INT32_MIN as a result. For large positive inputs the result |
| 56 | // of conversion can become negative, which affects the final requantization result. Note that on x86 SSE2 we |
| 57 | // have e.g. int32_t(float(INT32_MAX)) == INT32_MIN! This happens because float(INT32_MAX) rounds to 2**31, |
| 58 | // which overflows int32_t when it is converted back to integer. |
| 59 | // |
| 60 | // Thankfully, we can prove that overflow never happens in this requantization scheme. The largest positive |
| 61 | // input is INT32_MAX (2**31 - 1), which turns into 2**31 when converted to float. The largest scale value |
| 62 | // is 0x1.FFFFFEp-1. When multiplied together, the result is 2147483520 (compare to INT32_MAX = 2147483647), |
| 63 | // which fits into int32_t without overflow. |
| 64 | const __m128i x_rounded = _mm_cvtps_epi32(x_scaled); |
| 65 | const __m128i y_rounded = _mm_cvtps_epi32(y_scaled); |
| 66 | const __m128i z_rounded = _mm_cvtps_epi32(z_scaled); |
| 67 | const __m128i w_rounded = _mm_cvtps_epi32(w_scaled); |
| 68 | |
| 69 | // Standard final sequence on x86 SSE2: |
| 70 | // - Pack to int16_t and saturate |
| 71 | // - Add zero point |
| 72 | // - Pack to int8_t and saturate |
| 73 | // - Clamp between qmin and qmax |
| 74 | const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_rounded, y_rounded), vzero_point); |
| 75 | const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_rounded, w_rounded), vzero_point); |
| 76 | const __m128i xyzw_packed = _mm_packs_epi16(xy_packed, zw_packed); |
| 77 | const __m128i xyzw_clamped = _mm_max_epi8(_mm_min_epi8(xyzw_packed, vqmax), vqmin); |
| 78 | |
| 79 | // 4x CVTDQ2PS |
| 80 | // 4x MULPS |
| 81 | // 4x CVTPS2DQ |
| 82 | // 2x PACKSSDW |
| 83 | // 2x PADDSW |
| 84 | // 1x PACKSSWB |
| 85 | // 1x PMAXSB |
| 86 | // 1x PMINSB |
| 87 | // --------------------- |
| 88 | // 19 instructions total |
| 89 | |
| 90 | _mm_storeu_si128((__m128i*) output, xyzw_clamped); |
| 91 | output += 16; |
| 92 | } |
| 93 | } |