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 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;
FTransform(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]];
}
}
VP8SetHistogramData(distribution, histo);
}
static void CollectHistogram_SSE41(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]);
// v = abs(out) >> 3
const __m128i abs0 = _mm_abs_epi16(out0);
const __m128i abs1 = _mm_abs_epi16(out1);
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]];
}
}
VP8SetHistogramData(distribution, histo);
}
