src/requantization/fp32-psimd.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 <psimd.h>

#include <xnnpack/requantization-stubs.h>


void xnn_requantize_fp32__psimd(
    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);

  const psimd_f32 vscale = psimd_splat_f32(scale);
  const psimd_f32 vfmin = psimd_splat_f32((float) ((int32_t)(uint32_t) qmin - (int32_t)(uint32_t) zero_point));
  const psimd_f32 vfmax = psimd_splat_f32((float) ((int32_t)(uint32_t) qmax - (int32_t)(uint32_t) zero_point));
  const psimd_f32 vfmagic = psimd_splat_f32(12582912.0f);
  const psimd_s32 vimagic = psimd_splat_s32(INT32_C(0x4B400000) - (int32_t)(uint32_t) zero_point);
  for (; n != 0; n -= 16) {
    const psimd_s32 x = psimd_load_s32(input);
    const psimd_s32 y = psimd_load_s32(input + 4);
    const psimd_s32 z = psimd_load_s32(input + 8);
    const psimd_s32 w = psimd_load_s32(input + 12);
    input += 16;

    // Convert int32_t input to FP32 and multiply by FP32 scale.
    // Both operations involve roundings:
    // - Large int32_t values can't be exactly represented as FP32. We expect that conversion instruction would
    //   round it to nearest FP32 value with ties to even, but Clang documentation for __builtin_convertvector does
    //   not guaratee that.
    // - Product of two FP32 values is generally not exactly representation as an FP32 value, and will be rounded
    //   to nearest FP32 value with ties to even.
    const psimd_f32 x_scaled = psimd_cvt_s32_f32(x) * vscale;
    const psimd_f32 y_scaled = psimd_cvt_s32_f32(y) * vscale;
    const psimd_f32 z_scaled = psimd_cvt_s32_f32(z) * vscale;
    const psimd_f32 w_scaled = psimd_cvt_s32_f32(w) * vscale;

    // Clang/gcc vector extension does not provide an intrinsics for a floating-point to integer conversion
    // operation with rounding-to-nearest-even. In lieu of such intrinsic, we use a magic trick of adding a large
    // number (1.5 * 2**23) to scaled value to cause rounding to integer, and then substracing this magic number as
    // integer. This trick works only in a limited range (absolute value of input must be less than 2**22), so
    // generally we have to clamp input to this range before using the magic. However, clamping to any smaller range
    // works just as well, and thus we clamp to [qmin - zero point, qmax - zero point] range so that after we add
    // zero point to the result, it gets into target [qmin, qmax] range.
    const psimd_f32 x_clamped = psimd_min_f32(psimd_max_f32(x_scaled, vfmin), vfmax);
    const psimd_f32 y_clamped = psimd_min_f32(psimd_max_f32(y_scaled, vfmin), vfmax);
    const psimd_f32 z_clamped = psimd_min_f32(psimd_max_f32(z_scaled, vfmin), vfmax);
    const psimd_f32 w_clamped = psimd_min_f32(psimd_max_f32(w_scaled, vfmin), vfmax);

    // Conversion to integer using the "magic trick". Rounding is performed in the output of addition operation,
    // and result is rounded to nearest even integer with ties to even.
    const psimd_s32 x_biased = (psimd_s32)(x_clamped + vfmagic) - vimagic;
    const psimd_s32 y_biased = (psimd_s32)(y_clamped + vfmagic) - vimagic;
    const psimd_s32 z_biased = (psimd_s32)(z_clamped + vfmagic) - vimagic;
    const psimd_s32 w_biased = (psimd_s32)(w_clamped + vfmagic) - vimagic;

    // Select low 8 bits of each 32-bit integer in the vectors for the output.
    // Since result is already clamped to [qmin, qmax] subrange of [0, 255], saturation is not needed.
    const psimd_u16 xy_packed = psimd_concat_even_u16((psimd_u16) x_biased, (psimd_u16) y_biased);
    const psimd_u16 zw_packed = psimd_concat_even_u16((psimd_u16) z_biased, (psimd_u16) w_biased);

    const psimd_u8 xyzw_packed = psimd_concat_even_u8((psimd_u8) xy_packed, (psimd_u8) zw_packed);

    psimd_store_u8(output, xyzw_packed);
    output += 16;
  }
}