Blame src/dsp/lossless_enc_msa.c

Packit 9c6abc
// Copyright 2016 Google Inc. All Rights Reserved.
Packit 9c6abc
//
Packit 9c6abc
// Use of this source code is governed by a BSD-style license
Packit 9c6abc
// that can be found in the COPYING file in the root of the source
Packit 9c6abc
// tree. An additional intellectual property rights grant can be found
Packit 9c6abc
// in the file PATENTS. All contributing project authors may
Packit 9c6abc
// be found in the AUTHORS file in the root of the source tree.
Packit 9c6abc
// -----------------------------------------------------------------------------
Packit 9c6abc
//
Packit 9c6abc
// MSA variant of Image transform methods for lossless encoder.
Packit 9c6abc
//
Packit 9c6abc
// Authors: Prashant Patil (Prashant.Patil@imgtec.com)
Packit 9c6abc
Packit 9c6abc
#include "src/dsp/dsp.h"
Packit 9c6abc
Packit 9c6abc
#if defined(WEBP_USE_MSA)
Packit 9c6abc
Packit 9c6abc
#include "src/dsp/lossless.h"
Packit 9c6abc
#include "src/dsp/msa_macro.h"
Packit 9c6abc
Packit 9c6abc
#define TRANSFORM_COLOR_8(src0, src1, dst0, dst1, c0, c1, mask0, mask1) do {  \
Packit 9c6abc
  v8i16 g0, g1, t0, t1, t2, t3;                                               \
Packit 9c6abc
  v4i32 t4, t5;                                                               \
Packit 9c6abc
  VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1);                   \
Packit 9c6abc
  DOTP_SB2_SH(g0, g1, c0, c0, t0, t1);                                        \
Packit 9c6abc
  SRAI_H2_SH(t0, t1, 5);                                                      \
Packit 9c6abc
  t0 = __msa_subv_h((v8i16)src0, t0);                                         \
Packit 9c6abc
  t1 = __msa_subv_h((v8i16)src1, t1);                                         \
Packit 9c6abc
  t4 = __msa_srli_w((v4i32)src0, 16);                                         \
Packit 9c6abc
  t5 = __msa_srli_w((v4i32)src1, 16);                                         \
Packit 9c6abc
  DOTP_SB2_SH(t4, t5, c1, c1, t2, t3);                                        \
Packit 9c6abc
  SRAI_H2_SH(t2, t3, 5);                                                      \
Packit 9c6abc
  SUB2(t0, t2, t1, t3, t0, t1);                                               \
Packit 9c6abc
  VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1);                   \
Packit 9c6abc
} while (0)
Packit 9c6abc
Packit 9c6abc
#define TRANSFORM_COLOR_4(src, dst, c0, c1, mask0, mask1) do {  \
Packit 9c6abc
  const v16i8 g0 = VSHF_SB(src, src, mask0);                    \
Packit 9c6abc
  v8i16 t0 = __msa_dotp_s_h(c0, g0);                            \
Packit 9c6abc
  v8i16 t1;                                                     \
Packit 9c6abc
  v4i32 t2;                                                     \
Packit 9c6abc
  t0 = SRAI_H(t0, 5);                                           \
Packit 9c6abc
  t0 = __msa_subv_h((v8i16)src, t0);                            \
Packit 9c6abc
  t2 = __msa_srli_w((v4i32)src, 16);                            \
Packit 9c6abc
  t1 = __msa_dotp_s_h(c1, (v16i8)t2);                           \
Packit 9c6abc
  t1 = SRAI_H(t1, 5);                                           \
Packit 9c6abc
  t0 = t0 - t1;                                                 \
Packit 9c6abc
  dst = VSHF_UB(src, t0, mask1);                                \
Packit 9c6abc
} while (0)
Packit 9c6abc
Packit 9c6abc
static void TransformColor_MSA(const VP8LMultipliers* const m, uint32_t* data,
Packit 9c6abc
                               int num_pixels) {
Packit 9c6abc
  v16u8 src0, dst0;
Packit 9c6abc
  const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ |
Packit 9c6abc
                                         (m->green_to_red_ << 16));
Packit 9c6abc
  const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_);
Packit 9c6abc
  const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
Packit 9c6abc
                        13, 255, 13, 255 };
Packit 9c6abc
  const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11,
Packit 9c6abc
                        28, 13, 30, 15 };
Packit 9c6abc
Packit 9c6abc
  while (num_pixels >= 8) {
Packit 9c6abc
    v16u8 src1, dst1;
Packit 9c6abc
    LD_UB2(data, 4, src0, src1);
Packit 9c6abc
    TRANSFORM_COLOR_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1);
Packit 9c6abc
    ST_UB2(dst0, dst1, data, 4);
Packit 9c6abc
    data += 8;
Packit 9c6abc
    num_pixels -= 8;
Packit 9c6abc
  }
Packit 9c6abc
  if (num_pixels > 0) {
Packit 9c6abc
    if (num_pixels >= 4) {
Packit 9c6abc
      src0 = LD_UB(data);
Packit 9c6abc
      TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1);
Packit 9c6abc
      ST_UB(dst0, data);
Packit 9c6abc
      data += 4;
Packit 9c6abc
      num_pixels -= 4;
Packit 9c6abc
    }
Packit 9c6abc
    if (num_pixels > 0) {
Packit 9c6abc
      src0 = LD_UB(data);
Packit 9c6abc
      TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1);
Packit 9c6abc
      if (num_pixels == 3) {
Packit 9c6abc
        const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
Packit 9c6abc
        const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2);
Packit 9c6abc
        SD(pix_d, data + 0);
Packit 9c6abc
        SW(pix_w, data + 2);
Packit 9c6abc
      } else if (num_pixels == 2) {
Packit 9c6abc
        const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0);
Packit 9c6abc
        SD(pix_d, data);
Packit 9c6abc
      } else {
Packit 9c6abc
        const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0);
Packit 9c6abc
        SW(pix_w, data);
Packit 9c6abc
      }
Packit 9c6abc
    }
Packit 9c6abc
  }
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
static void SubtractGreenFromBlueAndRed_MSA(uint32_t* argb_data,
Packit 9c6abc
                                            int num_pixels) {
Packit 9c6abc
  int i;
Packit 9c6abc
  uint8_t* ptemp_data = (uint8_t*)argb_data;
Packit 9c6abc
  v16u8 src0, dst0, tmp0;
Packit 9c6abc
  const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255,
Packit 9c6abc
                       13, 255, 13, 255 };
Packit 9c6abc
Packit 9c6abc
  while (num_pixels >= 8) {
Packit 9c6abc
    v16u8 src1, dst1, tmp1;
Packit 9c6abc
    LD_UB2(ptemp_data, 16, src0, src1);
Packit 9c6abc
    VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1);
Packit 9c6abc
    SUB2(src0, tmp0, src1, tmp1, dst0, dst1);
Packit 9c6abc
    ST_UB2(dst0, dst1, ptemp_data, 16);
Packit 9c6abc
    ptemp_data += 8 * 4;
Packit 9c6abc
    num_pixels -= 8;
Packit 9c6abc
  }
Packit 9c6abc
  if (num_pixels > 0) {
Packit 9c6abc
    if (num_pixels >= 4) {
Packit 9c6abc
      src0 = LD_UB(ptemp_data);
Packit 9c6abc
      tmp0 = VSHF_UB(src0, src0, mask);
Packit 9c6abc
      dst0 = src0 - tmp0;
Packit 9c6abc
      ST_UB(dst0, ptemp_data);
Packit 9c6abc
      ptemp_data += 4 * 4;
Packit 9c6abc
      num_pixels -= 4;
Packit 9c6abc
    }
Packit 9c6abc
    for (i = 0; i < num_pixels; i++) {
Packit 9c6abc
      const uint8_t b = ptemp_data[0];
Packit 9c6abc
      const uint8_t g = ptemp_data[1];
Packit 9c6abc
      const uint8_t r = ptemp_data[2];
Packit 9c6abc
      ptemp_data[0] = (b - g) & 0xff;
Packit 9c6abc
      ptemp_data[2] = (r - g) & 0xff;
Packit 9c6abc
      ptemp_data += 4;
Packit 9c6abc
    }
Packit 9c6abc
  }
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
//------------------------------------------------------------------------------
Packit 9c6abc
// Entry point
Packit 9c6abc
Packit 9c6abc
extern void VP8LEncDspInitMSA(void);
Packit 9c6abc
Packit 9c6abc
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMSA(void) {
Packit 9c6abc
  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_MSA;
Packit 9c6abc
  VP8LTransformColor = TransformColor_MSA;
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
#else  // !WEBP_USE_MSA
Packit 9c6abc
Packit 9c6abc
WEBP_DSP_INIT_STUB(VP8LEncDspInitMSA)
Packit 9c6abc
Packit 9c6abc
#endif  // WEBP_USE_MSA