Quelle  psnr.cc   Sprache: C

/*
 *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS. All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "./psnr.h"  // NOLINT

#ifdef _OPENMP
#include <omp.h>
#endif
#ifdef _MSC_VER
#include <intrin.h>  // For __cpuid()
#endif

#ifdef __cplusplus
extern "C" {
#endif

typedef unsigned int uint32_t;  // NOLINT
#ifdef _MSC_VER
typedef unsigned __int64 uint64_t;
#else  // COMPILER_MSVC
#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
typedef unsigned long uint64_t;  // NOLINT
#else   // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
typedef unsigned long long uint64_t;  // NOLINT
#endif  // __LP64__
#endif  // _MSC_VER

// libyuv provides this function when linking library for jpeg support.
#if !defined(HAVE_JPEG)

#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
    !defined(__aarch64__)
#define HAS_SUMSQUAREERROR_NEON
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
                                    const uint8_t* src_b,
                                    int count) {
  volatile uint32_t sse;
  asm volatile(
      "vmov.u8 q7, #0 \n"
      "vmov.u8 q9, #0 \n"
      "vmov.u8 q8, #0 \n"
      "vmov.u8 q10, #0 \n"

      "1: \n"
      "vld1.u8 {q0}, [%0]! \n"
      "vld1.u8 {q1}, [%1]! \n"
      "vsubl.u8 q2, d0, d2 \n"
      "vsubl.u8 q3, d1, d3 \n"
      "vmlal.s16 q7, d4, d4 \n"
      "vmlal.s16 q8, d6, d6 \n"
      "vmlal.s16 q8, d5, d5 \n"
      "vmlal.s16 q10, d7, d7 \n"
      "subs %2, %2, #16 \n"
      "bhi 1b \n"

      "vadd.u32 q7, q7, q8 \n"
      "vadd.u32 q9, q9, q10 \n"
      "vadd.u32 q10, q7, q9 \n"
      "vpaddl.u32 q1, q10 \n"
      "vadd.u64 d0, d2, d3 \n"
      "vmov.32 %3, d0[0] \n"
      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
      :
      : "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
  return sse;
}
#elif !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
#define HAS_SUMSQUAREERROR_NEON
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
                                    const uint8_t* src_b,
                                    int count) {
  volatile uint32_t sse;
  asm volatile(
      "eor v16.16b, v16.16b, v16.16b \n"
      "eor v18.16b, v18.16b, v18.16b \n"
      "eor v17.16b, v17.16b, v17.16b \n"
      "eor v19.16b, v19.16b, v19.16b \n"

      "1: \n"
      "ld1 {v0.16b}, [%0], #16 \n"
      "ld1 {v1.16b}, [%1], #16 \n"
      "subs %w2, %w2, #16 \n"
      "usubl v2.8h, v0.8b, v1.8b \n"
      "usubl2 v3.8h, v0.16b, v1.16b \n"
      "smlal v16.4s, v2.4h, v2.4h \n"
      "smlal v17.4s, v3.4h, v3.4h \n"
      "smlal2 v18.4s, v2.8h, v2.8h \n"
      "smlal2 v19.4s, v3.8h, v3.8h \n"
      "b.gt 1b \n"

      "add v16.4s, v16.4s, v17.4s \n"
      "add v18.4s, v18.4s, v19.4s \n"
      "add v19.4s, v16.4s, v18.4s \n"
      "addv s0, v19.4s \n"
      "fmov %w3, s0 \n"
      : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
      :
      : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
  return sse;
}
#elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
#define HAS_SUMSQUAREERROR_SSE2
__declspec(naked) static uint32_t SumSquareError_SSE2(const uint8_t* /*src_a*/,
                                                      const uint8_t* /*src_b*/,
                                                      int /*count*/) {
  __asm {
    mov        eax, [esp + 4]  // src_a
    mov        edx, [esp + 8]  // src_b
    mov        ecx, [esp + 12]  // count
    pxor       xmm0, xmm0
    pxor       xmm5, xmm5
    sub        edx, eax

  wloop:
    movdqu     xmm1, [eax]
    movdqu     xmm2, [eax + edx]
    lea        eax,  [eax + 16]
    movdqu     xmm3, xmm1
    psubusb    xmm1, xmm2
    psubusb    xmm2, xmm3
    por        xmm1, xmm2
    movdqu     xmm2, xmm1
    punpcklbw  xmm1, xmm5
    punpckhbw  xmm2, xmm5
    pmaddwd    xmm1, xmm1
    pmaddwd    xmm2, xmm2
    paddd      xmm0, xmm1
    paddd      xmm0, xmm2
    sub        ecx, 16
    ja         wloop

    pshufd     xmm1, xmm0, 0EEh
    paddd      xmm0, xmm1
    pshufd     xmm1, xmm0, 01h
    paddd      xmm0, xmm1
    movd       eax, xmm0
    ret
  }
}
#elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
#define HAS_SUMSQUAREERROR_SSE2
static uint32_t SumSquareError_SSE2(const uint8_t* src_a,
                                    const uint8_t* src_b,
                                    int count) {
  uint32_t sse;
  asm volatile(  // NOLINT
      "pxor %%xmm0,%%xmm0 \n"
      "pxor %%xmm5,%%xmm5 \n"
      "sub %0,%1 \n"

      "1: \n"
      "movdqu (%0),%%xmm1 \n"
      "movdqu (%0,%1,1),%%xmm2 \n"
      "lea 0x10(%0),%0 \n"
      "movdqu %%xmm1,%%xmm3 \n"
      "psubusb %%xmm2,%%xmm1 \n"
      "psubusb %%xmm3,%%xmm2 \n"
      "por %%xmm2,%%xmm1 \n"
      "movdqu %%xmm1,%%xmm2 \n"
      "punpcklbw %%xmm5,%%xmm1 \n"
      "punpckhbw %%xmm5,%%xmm2 \n"
      "pmaddwd %%xmm1,%%xmm1 \n"
      "pmaddwd %%xmm2,%%xmm2 \n"
      "paddd %%xmm1,%%xmm0 \n"
      "paddd %%xmm2,%%xmm0 \n"
      "sub $0x10,%2 \n"
      "ja 1b \n"

      "pshufd $0xee,%%xmm0,%%xmm1 \n"
      "paddd %%xmm1,%%xmm0 \n"
      "pshufd $0x1,%%xmm0,%%xmm1 \n"
      "paddd %%xmm1,%%xmm0 \n"
      "movd %%xmm0,%3 \n"

      : "+r"(src_a),  // %0
        "+r"(src_b),  // %1
        "+r"(count),  // %2
        "=g"(sse)     // %3
      :
      : "memory", "cc"
#if defined(__SSE2__)
        ,
        "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
#endif
  );  // NOLINT
  return sse;
}
#endif  // LIBYUV_DISABLE_X86 etc

#if defined(HAS_SUMSQUAREERROR_SSE2)
#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
static __inline void __cpuid(int cpu_info[4], int info_type) {
  asm volatile(  // NOLINT
      "mov %%ebx, %%edi \n"
      "cpuid \n"
      "xchg %%edi, %%ebx \n"
      : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]),
        "=d"(cpu_info[3])
      : "a"(info_type));
}
// For gcc/clang but not clangcl.
#elif !defined(_MSC_VER) && (defined(__i386__) || defined(__x86_64__))
static __inline void __cpuid(int cpu_info[4], int info_type) {
  asm volatile(  // NOLINT
      "cpuid \n"
      : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]),
        "=d"(cpu_info[3])
      : "a"(info_type));
}
#endif

static int CpuHasSSE2() {
#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)
  int cpu_info[4];
  __cpuid(cpu_info, 1);
  if (cpu_info[3] & 0x04000000) {
    return 1;
  }
#endif
  return 0;
}
#endif  // HAS_SUMSQUAREERROR_SSE2

static uint32_t SumSquareError_C(const uint8_t* src_a,
                                 const uint8_t* src_b,
                                 int count) {
  uint32_t sse = 0u;
  for (int x = 0; x < count; ++x) {
    int diff = src_a[x] - src_b[x];
    sse += static_cast<uint32_t>(diff * diff);
  }
  return sse;
}

double ComputeSumSquareError(const uint8_t* src_a,
                             const uint8_t* src_b,
                             int count) {
  uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
                             int count) = SumSquareError_C;
#if defined(HAS_SUMSQUAREERROR_NEON)
  SumSquareError = SumSquareError_NEON;
#endif
#if defined(HAS_SUMSQUAREERROR_SSE2)
  if (CpuHasSSE2()) {
    SumSquareError = SumSquareError_SSE2;
  }
#endif
  const int kBlockSize = 1 << 15;
  uint64_t sse = 0;
#ifdef _OPENMP
#pragma omp parallel for reduction(+ : sse)
#endif
  for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
  }
  src_a += count & ~(kBlockSize - 1);
  src_b += count & ~(kBlockSize - 1);
  int remainder = count & (kBlockSize - 1) & ~15;
  if (remainder) {
    sse += SumSquareError(src_a, src_b, remainder);
    src_a += remainder;
    src_b += remainder;
  }
  remainder = count & 15;
  if (remainder) {
    sse += SumSquareError_C(src_a, src_b, remainder);
  }
  return static_cast<double>(sse);
}
#endif

// PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
// Returns 128.0 (kMaxPSNR) if sse is 0 (perfect match).
double ComputePSNR(double sse, double size) {
  const double kMINSSE = 255.0 * 255.0 * size / pow(10.0, kMaxPSNR / 10.0);
  if (sse <= kMINSSE) {
    sse = kMINSSE;  // Produces max PSNR of 128
  }
  return 10.0 * log10(255.0 * 255.0 * size / sse);
}

#ifdef __cplusplus
}  // extern "C"
#endif