static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
int start_block, int end_block,
VP8Histogram* const histo) {
int j;
int distribution[MAX_COEFF_THRESH + 1] = { 0 };
for (j = start_block; j < end_block; ++j) {
int16_t out[16];
VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
{
int k;
v8i16 coeff0, coeff1;
const v8i16 zero = { 0 };
const v8i16 max_coeff_thr = __msa_ldi_h(MAX_COEFF_THRESH);
LD_SH2(&out[0], 8, coeff0, coeff1);
coeff0 = __msa_add_a_h(coeff0, zero);
coeff1 = __msa_add_a_h(coeff1, zero);
SRAI_H2_SH(coeff0, coeff1, 3);
coeff0 = __msa_min_s_h(coeff0, max_coeff_thr);
coeff1 = __msa_min_s_h(coeff1, max_coeff_thr);
ST_SH2(coeff0, coeff1, &out[0], 8);
for (k = 0; k < 16; ++k) {
++distribution[out[k]];
}
}
}
VP8SetHistogramData(distribution, histo);
}
static int ReconstructIntra16(VP8EncIterator* const it,
VP8ModeScore* const rd,
uint8_t* const yuv_out,
int mode) {
VP8Encoder* const enc = it->enc_;
const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
const uint8_t* const src = it->yuv_in_ + Y_OFF;
VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
int nz = 0;
int n;
int16_t tmp[16][16], dc_tmp[16];
for (n = 0; n < 16; ++n) {
VP8FTransform(src + VP8Scan[n], ref + VP8Scan[n], tmp[n]);
}
VP8FTransformWHT(tmp[0], dc_tmp);
nz |= VP8EncQuantizeBlock(dc_tmp, rd->y_dc_levels, 0, &dqm->y2_) << 24;
if (DO_TRELLIS_I16 && it->do_trellis_) {
int x, y;
VP8IteratorNzToBytes(it);
for (y = 0, n = 0; y < 4; ++y) {
for (x = 0; x < 4; ++x, ++n) {
const int ctx = it->top_nz_[x] + it->left_nz_[y];
const int non_zero =
TrellisQuantizeBlock(it, tmp[n], rd->y_ac_levels[n], ctx, 0,
&dqm->y1_, dqm->lambda_trellis_i16_);
it->top_nz_[x] = it->left_nz_[y] = non_zero;
nz |= non_zero << n;
}
}
} else {
for (n = 0; n < 16; ++n) {
nz |= VP8EncQuantizeBlock(tmp[n], rd->y_ac_levels[n], 1, &dqm->y1_) << n;
}
}
// Transform back
VP8ITransformWHT(dc_tmp, tmp[0]);
for (n = 0; n < 16; n += 2) {
VP8ITransform(ref + VP8Scan[n], tmp[n], yuv_out + VP8Scan[n], 1);
}
return nz;
}
static int ReconstructIntra4(VP8EncIterator* const it,
int16_t levels[16],
const uint8_t* const src,
uint8_t* const yuv_out,
int mode) {
const VP8Encoder* const enc = it->enc_;
const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
int nz = 0;
int16_t tmp[16];
VP8FTransform(src, ref, tmp);
if (DO_TRELLIS_I4 && it->do_trellis_) {
const int x = it->i4_ & 3, y = it->i4_ >> 2;
const int ctx = it->top_nz_[x] + it->left_nz_[y];
nz = TrellisQuantizeBlock(it, tmp, levels, ctx, 3, &dqm->y1_,
dqm->lambda_trellis_i4_);
} else {
static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
int start_block, int end_block,
VP8Histogram* const histo) {
const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
int j;
int distribution[MAX_COEFF_THRESH + 1] = { 0 };
for (j = start_block; j < end_block; ++j) {
int16_t out[16];
int k;
VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
// Convert coefficients to bin (within out[]).
{
// Load.
const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
// sign(out) = out >> 15 (0x0000 if positive, 0xffff if negative)
const __m128i sign0 = _mm_srai_epi16(out0, 15);
const __m128i sign1 = _mm_srai_epi16(out1, 15);
// abs(out) = (out ^ sign) - sign
const __m128i xor0 = _mm_xor_si128(out0, sign0);
const __m128i xor1 = _mm_xor_si128(out1, sign1);
const __m128i abs0 = _mm_sub_epi16(xor0, sign0);
const __m128i abs1 = _mm_sub_epi16(xor1, sign1);
// v = abs(out) >> 3
const __m128i v0 = _mm_srai_epi16(abs0, 3);
const __m128i v1 = _mm_srai_epi16(abs1, 3);
// bin = min(v, MAX_COEFF_THRESH)
const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
// Store.
_mm_storeu_si128((__m128i*)&out[0], bin0);
_mm_storeu_si128((__m128i*)&out[8], bin1);
}
// Convert coefficients to bin.
for (k = 0; k < 16; ++k) {
++distribution[out[k]];
}
}
VP8LSetHistogramData(distribution, histo);
}
static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
int start_block, int end_block) {
int histo[MAX_COEFF_THRESH + 1] = { 0 };
int16_t out[16];
int j, k;
const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
for (j = start_block; j < end_block; ++j) {
VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
// Convert coefficients to bin (within out[]).
{
// Load.
const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
// sign(out) = out >> 15 (0x0000 if positive, 0xffff if negative)
const __m128i sign0 = _mm_srai_epi16(out0, 15);
const __m128i sign1 = _mm_srai_epi16(out1, 15);
// abs(out) = (out ^ sign) - sign
const __m128i xor0 = _mm_xor_si128(out0, sign0);
const __m128i xor1 = _mm_xor_si128(out1, sign1);
const __m128i abs0 = _mm_sub_epi16(xor0, sign0);
const __m128i abs1 = _mm_sub_epi16(xor1, sign1);
// v = abs(out) >> 2
const __m128i v0 = _mm_srai_epi16(abs0, 2);
const __m128i v1 = _mm_srai_epi16(abs1, 2);
// bin = min(v, MAX_COEFF_THRESH)
const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
// Store.
_mm_storeu_si128((__m128i*)&out[0], bin0);
_mm_storeu_si128((__m128i*)&out[8], bin1);
}
// Use bin to update histogram.
for (k = 0; k < 16; ++k) {
histo[out[k]]++;
}
}
return VP8GetAlpha(histo);
}
