src/qu8-requantization/gemmlowp-sse4.c

// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 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 <stdint.h>
#include <stddef.h>

#include <smmintrin.h>

#include <fp16/bitcasts.h>

#include <xnnpack/requantization-stubs.h>


void xnn_qu8_requantize_gemmlowp__sse4(
    size_t n,
    const int32_t* input,
    float scale,
    uint8_t zero_point,
    uint8_t qmin,
    uint8_t qmax,
    uint8_t* output)
{
  assert(n % 16 == 0);
  assert(scale < 1.0f);
  assert(scale >= 0x1.0p-32f);

  // Compute requantization parameters.
  const uint32_t scale_bits = fp32_to_bits(scale);

  // Multiplier is in [0x40000000, 0x7FFFFF80] range.
  const int32_t multiplier = (int32_t)(((scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000)) << 7);
  assert(multiplier >= INT32_C(0x40000000));
  assert(multiplier <= INT32_C(0x7FFFFF80));

  // Shift is in [0, 31] range.
  const int32_t shift = 127 + 31 - 32 - (fp32_to_bits(scale) >> 23);
  assert(shift >= 0);
  assert(shift < 32);

  const __m128i vmultiplier = _mm_set1_epi32(multiplier);
  const __m128i vzero_point = _mm_set1_epi16((short) (uint16_t) zero_point);
  const __m128i vqmin = _mm_set1_epi8((char) qmin);
  const __m128i vqmax = _mm_set1_epi8((char) qmax);
  const __m128i vshift = _mm_cvtsi32_si128((int) shift);
  const uint32_t remainder_mask = (UINT32_C(1) << shift) - UINT32_C(1);
  const __m128i vremainder_mask = _mm_set1_epi32((int) remainder_mask);
  const __m128i vthreshold = _mm_set1_epi32((int) (remainder_mask >> 1));
  const __m128i vq31rounding = _mm_set1_epi64x(UINT64_C(0x40000000));
  for (; n != 0; n -= 16) {
    const __m128i x = _mm_loadu_si128((const __m128i*) input);
    const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4));
    const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8));
    const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12));
    input += 16;

    const __m128i x_rev = _mm_shuffle_epi32(x, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i y_rev = _mm_shuffle_epi32(y, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i z_rev = _mm_shuffle_epi32(z, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i w_rev = _mm_shuffle_epi32(w, _MM_SHUFFLE(2, 3, 0, 1));

    const __m128i x_product_even = _mm_add_epi64(_mm_mul_epi32(x, vmultiplier), vq31rounding);
    const __m128i y_product_even = _mm_add_epi64(_mm_mul_epi32(y, vmultiplier), vq31rounding);
    const __m128i z_product_even = _mm_add_epi64(_mm_mul_epi32(z, vmultiplier), vq31rounding);
    const __m128i w_product_even = _mm_add_epi64(_mm_mul_epi32(w, vmultiplier), vq31rounding);

    const __m128i x_product_odd = _mm_add_epi64(_mm_mul_epi32(x_rev, vmultiplier), vq31rounding);
    const __m128i y_product_odd = _mm_add_epi64(_mm_mul_epi32(y_rev, vmultiplier), vq31rounding);
    const __m128i z_product_odd = _mm_add_epi64(_mm_mul_epi32(z_rev, vmultiplier), vq31rounding);
    const __m128i w_product_odd = _mm_add_epi64(_mm_mul_epi32(w_rev, vmultiplier), vq31rounding);

    const __m128i x_q31product_even = _mm_srli_epi64(x_product_even, 31);
    const __m128i x_q31product_odd = _mm_add_epi64(x_product_odd, x_product_odd);
    const __m128i y_q31product_even = _mm_srli_epi64(y_product_even, 31);
    const __m128i y_q31product_odd = _mm_add_epi64(y_product_odd, y_product_odd);
    const __m128i z_q31product_even = _mm_srli_epi64(z_product_even, 31);
    const __m128i z_q31product_odd = _mm_add_epi64(z_product_odd, z_product_odd);
    const __m128i w_q31product_even = _mm_srli_epi64(w_product_even, 31);
    const __m128i w_q31product_odd = _mm_add_epi64(w_product_odd, w_product_odd);

    const __m128i x_q31product = _mm_blend_epi16(x_q31product_even, x_q31product_odd, 0xCC);
    const __m128i y_q31product = _mm_blend_epi16(y_q31product_even, y_q31product_odd, 0xCC);
    const __m128i z_q31product = _mm_blend_epi16(z_q31product_even, z_q31product_odd, 0xCC);
    const __m128i w_q31product = _mm_blend_epi16(w_q31product_even, w_q31product_odd, 0xCC);

    const __m128i x_remainder =
        _mm_add_epi32(_mm_and_si128(x_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), x_q31product));
    const __m128i y_remainder =
        _mm_add_epi32(_mm_and_si128(y_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), y_q31product));
    const __m128i z_remainder =
        _mm_add_epi32(_mm_and_si128(z_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), z_q31product));
    const __m128i w_remainder =
        _mm_add_epi32(_mm_and_si128(w_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), w_q31product));

    const __m128i x_scaled =
        _mm_sub_epi32(_mm_sra_epi32(x_q31product, vshift), _mm_cmpgt_epi32(x_remainder, vthreshold));
    const __m128i y_scaled =
        _mm_sub_epi32(_mm_sra_epi32(y_q31product, vshift), _mm_cmpgt_epi32(y_remainder, vthreshold));
    const __m128i z_scaled =
        _mm_sub_epi32(_mm_sra_epi32(z_q31product, vshift), _mm_cmpgt_epi32(z_remainder, vthreshold));
    const __m128i w_scaled =
        _mm_sub_epi32(_mm_sra_epi32(w_q31product, vshift), _mm_cmpgt_epi32(w_remainder, vthreshold));

    const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point);
    const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point);
    const __m128i xyzw_packed = _mm_packus_epi16(xy_packed, zw_packed);
    const __m128i xyzw_clamped = _mm_max_epu8(_mm_min_epu8(xyzw_packed, vqmax), vqmin);

    // 4x PSHUFD
    // 8x PMULDQ
    // 12x PADDQ
    // 4x PADDD
    // 2x PADDW
    // 4x PSUBD
    // 4x PSLRQ (immediate)
    // 4x PSRAD (register)
    // 4x PBLENDW
    // 4x PAND
    // 4x PXOR (setzero)
    // 8x PCMPGTD
    // 2x PACKSSDW
    // 1x PACKUSWB
    // 1x PMAXUB
    // 1x PMINUB
    // ---------------------
    // 67 instructions total

    _mm_storeu_si128((__m128i*) output, xyzw_clamped);
    output += 16;
  }
}