///////////////////////////////////////////////////////////////////////
// File:        dotproductsse.cpp
// Description: Architecture-specific dot-product function.
// Author:      Ray Smith
// Created:     Wed Jul 22 10:57:45 PDT 2015
//
// (C) Copyright 2015, Google Inc.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// http://www.apache.org/licenses/LICENSE-2.0
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
///////////////////////////////////////////////////////////////////////

#if !defined(__SSE4_1__)
// This code can't compile with "-msse4.1", so use dummy stubs.

#include "dotproductsse.h"
#include <stdio.h>
#include <stdlib.h>

namespace tesseract {
double DotProductSSE(const double* u, const double* v, int n) {
  fprintf(stderr, "DotProductSSE can't be used on Android\n");
  abort();
}
int32_t IntDotProductSSE(const int8_t* u, const int8_t* v, int n) {
  fprintf(stderr, "IntDotProductSSE can't be used on Android\n");
  abort();
}
}  // namespace tesseract

#else  // !defined(__SSE4_1__)
// Non-Android code here

#include <emmintrin.h>
#include <smmintrin.h>
#include <stdint.h>
#include "dotproductsse.h"
#include "host.h"

namespace tesseract {

// Computes and returns the dot product of the n-vectors u and v.
// Uses Intel SSE intrinsics to access the SIMD instruction set.
double DotProductSSE(const double* u, const double* v, int n) {
  int max_offset = n - 2;
  int offset = 0;
  // Accumulate a set of 2 sums in sum, by loading pairs of 2 values from u and
  // v, and multiplying them together in parallel.
  __m128d sum = _mm_setzero_pd();
  if (offset <= max_offset) {
    offset = 2;
    // Aligned load is reputedly faster but requires 16 byte aligned input.
    if ((reinterpret_cast<const uintptr_t>(u) & 15) == 0 &&
        (reinterpret_cast<const uintptr_t>(v) & 15) == 0) {
      // Use aligned load.
      sum = _mm_load_pd(u);
      __m128d floats2 = _mm_load_pd(v);
      // Multiply.
      sum = _mm_mul_pd(sum, floats2);
      while (offset <= max_offset) {
        __m128d floats1 = _mm_load_pd(u + offset);
        floats2 = _mm_load_pd(v + offset);
        offset += 2;
        floats1 = _mm_mul_pd(floats1, floats2);
        sum = _mm_add_pd(sum, floats1);
      }
    } else {
      // Use unaligned load.
      sum = _mm_loadu_pd(u);
      __m128d floats2 = _mm_loadu_pd(v);
      // Multiply.
      sum = _mm_mul_pd(sum, floats2);
      while (offset <= max_offset) {
        __m128d floats1 = _mm_loadu_pd(u + offset);
        floats2 = _mm_loadu_pd(v + offset);
        offset += 2;
        floats1 = _mm_mul_pd(floats1, floats2);
        sum = _mm_add_pd(sum, floats1);
      }
    }
  }
  // Add the 2 sums in sum horizontally.
  sum = _mm_hadd_pd(sum, sum);
  // Extract the low result.
  double result = _mm_cvtsd_f64(sum);
  // Add on any left-over products.
  while (offset < n) {
    result += u[offset] * v[offset];
    ++offset;
  }
  return result;
}

// Computes and returns the dot product of the n-vectors u and v.
// Uses Intel SSE intrinsics to access the SIMD instruction set.
int32_t IntDotProductSSE(const int8_t* u, const int8_t* v, int n) {
  int max_offset = n - 8;
  int offset = 0;
  // Accumulate a set of 4 32-bit sums in sum, by loading 8 pairs of 8-bit
  // values, extending to 16 bit, multiplying to make 32 bit results.
  __m128i sum = _mm_setzero_si128();
  if (offset <= max_offset) {
    offset = 8;
    __m128i packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(u));
    __m128i packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(v));
    sum = _mm_cvtepi8_epi16(packed1);
    packed2 = _mm_cvtepi8_epi16(packed2);
    // The magic _mm_add_epi16 is perfect here. It multiplies 8 pairs of 16 bit
    // ints to make 32 bit results, which are then horizontally added in pairs
    // to make 4 32 bit results that still fit in a 128 bit register.
    sum = _mm_madd_epi16(sum, packed2);
    while (offset <= max_offset) {
      packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(u + offset));
      packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(v + offset));
      offset += 8;
      packed1 = _mm_cvtepi8_epi16(packed1);
      packed2 = _mm_cvtepi8_epi16(packed2);
      packed1 = _mm_madd_epi16(packed1, packed2);
      sum = _mm_add_epi32(sum, packed1);
    }
  }
  // Sum the 4 packed 32 bit sums and extract the low result.
  sum = _mm_hadd_epi32(sum, sum);
  sum = _mm_hadd_epi32(sum, sum);
  int32_t result = _mm_cvtsi128_si32(sum);
  while (offset < n) {
    result += u[offset] * v[offset];
    ++offset;
  }
  return result;
}

}  // namespace tesseract.

#endif  // ANDROID_BUILD