Quelle  upsampling_neon.c

// Copyright 2011 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING 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.
// -----------------------------------------------------------------------------
//
// NEON version of YUV to RGB upsampling functions.
//
// Author: mans@mansr.com (Mans Rullgard)
// Based on SSE code by: somnath@google.com (Somnath Banerjee)

#include "src/dsp/dsp.h"

#if defined(WEBP_USE_NEON)

#include <assert.h>
#include <arm_neon.h>
#include <string.h>
#include "src/dsp/neon.h"
#include "src/dsp/yuv.h"

#ifdef FANCY_UPSAMPLING

//-----------------------------------------------------------------------------
// U/V upsampling

// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
#define UPSAMPLE_16PIXELS(r1, r2, out) do {                             \
  const uint8x8_t a = vld1_u8(r1 + 0);                                  \
  const uint8x8_t b = vld1_u8(r1 + 1);                                  \
  const uint8x8_t c = vld1_u8(r2 + 0);                                  \
  const uint8x8_t d = vld1_u8(r2 + 1);                                  \
  /* a + b + c + d */                                                   \
  const uint16x8_t ad = vaddl_u8(a,  d);                                \
  const uint16x8_t bc = vaddl_u8(b,  c);                                \
  const uint16x8_t abcd = vaddq_u16(ad, bc);                            \
  /* 3a +  b +  c + 3d */                                               \
  const uint16x8_t al = vaddq_u16(abcd, vshlq_n_u16(ad, 1));            \
  /*  a + 3b + 3c +  d */                                               \
  const uint16x8_t bl = vaddq_u16(abcd, vshlq_n_u16(bc, 1));            \
                                                                        \
  const uint8x8_t diag2 = vshrn_n_u16(al, 3);                           \
  const uint8x8_t diag1 = vshrn_n_u16(bl, 3);                           \
                                                                        \
  const uint8x8_t A = vrhadd_u8(a, diag1);                              \
  const uint8x8_t B = vrhadd_u8(b, diag2);                              \
  const uint8x8_t C = vrhadd_u8(c, diag2);                              \
  const uint8x8_t D = vrhadd_u8(d, diag1);                              \
                                                                        \
  uint8x8x2_t A_B, C_D;                                                 \
  INIT_VECTOR2(A_B, A, B);                                              \
  INIT_VECTOR2(C_D, C, D);                                              \
  vst2_u8(out +  0, A_B);                                               \
  vst2_u8(out + 32, C_D);                                               \
} while (0)

// Turn the macro into a function for reducing code-size when non-critical
static void Upsample16Pixels_NEON(const uint8_t* WEBP_RESTRICT const r1,
                                  const uint8_t* WEBP_RESTRICT const r2,
                                  uint8_t* WEBP_RESTRICT const out) {
  UPSAMPLE_16PIXELS(r1, r2, out);
}

#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                  \
  uint8_t r1[9], r2[9];                                                 \
  memcpy(r1, (tb), (num_pixels));                                       \
  memcpy(r2, (bb), (num_pixels));                                       \
  /* replicate last byte */                                             \
  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels));    \
  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels));    \
  Upsample16Pixels_NEON(r1, r2, out);                                   \
}

//-----------------------------------------------------------------------------
// YUV->RGB conversion

// note: we represent the 33050 large constant as 32768 + 282
static const int16_t kCoeffs1[4] = { 19077, 26149, 6419, 13320 };

#define v255 vdup_n_u8(255)

#define STORE_Rgb(out, r, g, b) do {                                    \
  uint8x8x3_t r_g_b;                                                    \
  INIT_VECTOR3(r_g_b, r, g, b);                                         \
  vst3_u8(out, r_g_b);                                                  \
} while (0)

#define STORE_Bgr(out, r, g, b) do {                                    \
  uint8x8x3_t b_g_r;                                                    \
  INIT_VECTOR3(b_g_r, b, g, r);                                         \
  vst3_u8(out, b_g_r);                                                  \
} while (0)

#define STORE_Rgba(out, r, g, b) do {                                   \
  uint8x8x4_t r_g_b_v255;                                               \
  INIT_VECTOR4(r_g_b_v255, r, g, b, v255);                              \
  vst4_u8(out, r_g_b_v255);                                             \
} while (0)

#define STORE_Bgra(out, r, g, b) do {                                   \
  uint8x8x4_t b_g_r_v255;                                               \
  INIT_VECTOR4(b_g_r_v255, b, g, r, v255);                              \
  vst4_u8(out, b_g_r_v255);                                             \
} while (0)

#define STORE_Argb(out, r, g, b) do {                                   \
  uint8x8x4_t v255_r_g_b;                                               \
  INIT_VECTOR4(v255_r_g_b, v255, r, g, b);                              \
  vst4_u8(out, v255_r_g_b);                                             \
} while (0)

#if (WEBP_SWAP_16BIT_CSP == 0)
#define ZIP_U8(lo, hi) vzip_u8((lo), (hi))
#else
#define ZIP_U8(lo, hi) vzip_u8((hi), (lo))
#endif

#define STORE_Rgba4444(out, r, g, b) do {                               \
  const uint8x8_t rg = vsri_n_u8(r, g, 4);      /* shift g, insert r */ \
  const uint8x8_t ba = vsri_n_u8(b, v255, 4);   /* shift a, insert b */ \
  const uint8x8x2_t rgba4444 = ZIP_U8(rg, ba);                          \
  vst1q_u8(out, vcombine_u8(rgba4444.val[0], rgba4444.val[1]));         \
} while (0)

#define STORE_Rgb565(out, r, g, b) do {                                 \
  const uint8x8_t rg = vsri_n_u8(r, g, 5);   /* shift g and insert r */ \
  const uint8x8_t g1 = vshl_n_u8(g, 3);      /* pre-shift g: 3bits */   \
  const uint8x8_t gb = vsri_n_u8(g1, b, 3);  /* shift b and insert g */ \
  const uint8x8x2_t rgb565 = ZIP_U8(rg, gb);                            \
  vst1q_u8(out, vcombine_u8(rgb565.val[0], rgb565.val[1]));             \
} while (0)

#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) do {         \
  int i;                                                                \
  for (i = 0; i < N; i += 8) {                                          \
    const int off = ((cur_x) + i) * XSTEP;                              \
    const uint8x8_t y  = vld1_u8((src_y) + (cur_x)  + i);               \
    const uint8x8_t u  = vld1_u8((src_uv) + i +  0);                    \
    const uint8x8_t v  = vld1_u8((src_uv) + i + 16);                    \
    const int16x8_t Y0 = vreinterpretq_s16_u16(vshll_n_u8(y, 7));       \
    const int16x8_t U0 = vreinterpretq_s16_u16(vshll_n_u8(u, 7));       \
    const int16x8_t V0 = vreinterpretq_s16_u16(vshll_n_u8(v, 7));       \
    const int16x8_t Y1 = vqdmulhq_lane_s16(Y0, coeff1, 0);              \
    const int16x8_t R0 = vqdmulhq_lane_s16(V0, coeff1, 1);              \
    const int16x8_t G0 = vqdmulhq_lane_s16(U0, coeff1, 2);              \
    const int16x8_t G1 = vqdmulhq_lane_s16(V0, coeff1, 3);              \
    const int16x8_t B0 = vqdmulhq_n_s16(U0, 282);                       \
    const int16x8_t R1 = vqaddq_s16(Y1, R_Rounder);                     \
    const int16x8_t G2 = vqaddq_s16(Y1, G_Rounder);                     \
    const int16x8_t B1 = vqaddq_s16(Y1, B_Rounder);                     \
    const int16x8_t R2 = vqaddq_s16(R0, R1);                            \
    const int16x8_t G3 = vqaddq_s16(G0, G1);                            \
    const int16x8_t B2 = vqaddq_s16(B0, B1);                            \
    const int16x8_t G4 = vqsubq_s16(G2, G3);                            \
    const int16x8_t B3 = vqaddq_s16(B2, U0);                            \
    const uint8x8_t R = vqshrun_n_s16(R2, YUV_FIX2);                    \
    const uint8x8_t G = vqshrun_n_s16(G4, YUV_FIX2);                    \
    const uint8x8_t B = vqshrun_n_s16(B3, YUV_FIX2);                    \
    STORE_ ## FMT(out + off, R, G, B);                                  \
  }                                                                     \
} while (0)

#define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) {           \
  int i;                                                                \
  for (i = 0; i < N; i++) {                                             \
    const int off = ((cur_x) + i) * XSTEP;                              \
    const int y = src_y[(cur_x) + i];                                   \
    const int u = (src_uv)[i];                                          \
    const int v = (src_uv)[i + 16];                                     \
    FUNC(y, u, v, rgb + off);                                           \
  }                                                                     \
}

#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv,                  \
                      top_dst, bottom_dst, cur_x, len) {                \
  CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x);                 \
  if (bottom_y != NULL) {                                               \
    CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x);  \
  }                                                                     \
}

#define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv,                 \
                      top_dst, bottom_dst, cur_x, len) {                \
  CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x);                \
  if (bottom_y != NULL) {                                               \
    CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
  }                                                                     \
}

#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP)                              \
static void FUNC_NAME(const uint8_t* WEBP_RESTRICT top_y,                      \
                      const uint8_t* WEBP_RESTRICT bottom_y,                   \
                      const uint8_t* WEBP_RESTRICT top_u,                      \
                      const uint8_t* WEBP_RESTRICT top_v,                      \
                      const uint8_t* WEBP_RESTRICT cur_u,                      \
                      const uint8_t* WEBP_RESTRICT cur_v,                      \
                      uint8_t* WEBP_RESTRICT top_dst,                          \
                      uint8_t* WEBP_RESTRICT bottom_dst, int len) {            \
  int block;                                                                   \
  /* 16 byte aligned array to cache reconstructed u and v */                   \
  uint8_t uv_buf[2 * 32 + 15];                                                 \
  uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~(uintptr_t)15); \
  const int uv_len = (len + 1) >> 1;                                           \
  /* 9 pixels must be read-able for each block */                              \
  const int num_blocks = (uv_len - 1) >> 3;                                    \
  const int leftover = uv_len - num_blocks * 8;                                \
  const int last_pos = 1 + 16 * num_blocks;                                    \
                                                                               \
  const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                         \
  const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                         \
                                                                               \
  const int16x4_t coeff1 = vld1_s16(kCoeffs1);                                 \
  const int16x8_t R_Rounder = vdupq_n_s16(-14234);                             \
  const int16x8_t G_Rounder = vdupq_n_s16(8708);                               \
  const int16x8_t B_Rounder = vdupq_n_s16(-17685);                             \
                                                                               \
  /* Treat the first pixel in regular way */                                   \
  assert(top_y != NULL);                                                       \
  {                                                                            \
    const int u0 = (top_u[0] + u_diag) >> 1;                                   \
    const int v0 = (top_v[0] + v_diag) >> 1;                                   \
    VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst);                                \
  }                                                                            \
  if (bottom_y != NULL) {                                                      \
    const int u0 = (cur_u[0] + u_diag) >> 1;                                   \
    const int v0 = (cur_v[0] + v_diag) >> 1;                                   \
    VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst);                          \
  }                                                                            \
                                                                               \
  for (block = 0; block < num_blocks; ++block) {                               \
    UPSAMPLE_16PIXELS(top_u, cur_u, r_uv);                                     \
    UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16);                                \
    CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv,                           \
                  top_dst, bottom_dst, 16 * block + 1, 16);                    \
    top_u += 8;                                                                \
    cur_u += 8;                                                                \
    top_v += 8;                                                                \
    cur_v += 8;                                                                \
  }                                                                            \
                                                                               \
  UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv);                           \
  UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16);                      \
  CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv,                 \
                top_dst, bottom_dst, last_pos, len - last_pos);                \
}

// NEON variants of the fancy upsampler.
NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair_NEON, Rgba, 4)
NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair_NEON, Bgra, 4)
#if !defined(WEBP_REDUCE_CSP)
NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair_NEON,  Rgb,  3)
NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair_NEON,  Bgr,  3)
NEON_UPSAMPLE_FUNC(UpsampleArgbLinePair_NEON, Argb, 4)
NEON_UPSAMPLE_FUNC(UpsampleRgba4444LinePair_NEON, Rgba4444, 2)
NEON_UPSAMPLE_FUNC(UpsampleRgb565LinePair_NEON, Rgb565, 2)
#endif   // WEBP_REDUCE_CSP

//------------------------------------------------------------------------------
// Entry point

extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];

extern void WebPInitUpsamplersNEON(void);

WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersNEON(void) {
  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_NEON;
  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_NEON;
  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_NEON;
  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_NEON;
#if !defined(WEBP_REDUCE_CSP)
  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair_NEON;
  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair_NEON;
  WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_NEON;
  WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_NEON;
  WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_NEON;
  WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_NEON;
  WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_NEON;
#endif   // WEBP_REDUCE_CSP
}

#endif  // FANCY_UPSAMPLING

#endif  // WEBP_USE_NEON

#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_NEON))
WEBP_DSP_INIT_STUB(WebPInitUpsamplersNEON)
#endif