generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge changes I53f8a160,I48f282bf

Browse Source 
* changes:
  quantize ssse3: copy style from sse2
  quantize sse2: copy opts from ssse3
This commit is contained in:
Johann Koenig
2017-08-22 22:27:56 +00:00
committed by Gerrit Code Review
parent 7f2993f5e4 b9c1dcc5fa
commit 103e4e50a8
2 changed files with 192 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files

292
vpx_dsp/x86/quantize_sse2.c
View File

@@ -23,185 +23,151 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan_ptr, uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) { const int16_t *iscan_ptr) {
__m128i zero; const __m128i zero = _mm_setzero_si128();
__m128i eob; int index = 16;
__m128i zbin;
__m128i round, quant, dequant, shift; __m128i zbin, round, quant, dequant, shift;
__m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i cmp_mask0, cmp_mask1;
__m128i qtmp0, qtmp1;
__m128i zero_coeff0, zero_coeff1, iscan0, iscan1;
__m128i eob, eob0, eob1;
(void)scan_ptr; (void)scan_ptr;
(void)skip_block; (void)skip_block;
assert(!skip_block); assert(!skip_block);
coeff_ptr += n_coeffs; // Setup global values.
iscan_ptr += n_coeffs; zbin = _mm_load_si128((const __m128i *)zbin_ptr);
qcoeff_ptr += n_coeffs; round = _mm_load_si128((const __m128i *)round_ptr);
dqcoeff_ptr += n_coeffs; quant = _mm_load_si128((const __m128i *)quant_ptr);
n_coeffs = -n_coeffs; zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1));
zero = _mm_setzero_si128(); dequant = _mm_load_si128((const __m128i *)dequant_ptr);
{ shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
__m128i coeff0, coeff1;
// Setup global values // Do DC and first 15 AC.
{ coeff0 = load_tran_low(coeff_ptr);
__m128i pw_1; coeff1 = load_tran_low(coeff_ptr + 8);
zbin = _mm_load_si128((const __m128i *)zbin_ptr);
round = _mm_load_si128((const __m128i *)round_ptr);
quant = _mm_load_si128((const __m128i *)quant_ptr);
pw_1 = _mm_set1_epi16(1);
zbin = _mm_sub_epi16(zbin, pw_1);
dequant = _mm_load_si128((const __m128i *)dequant_ptr);
shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
}
{ // Poor man's abs().
__m128i coeff0_sign, coeff1_sign; coeff0_sign = _mm_srai_epi16(coeff0, 15);
__m128i qcoeff0, qcoeff1; coeff1_sign = _mm_srai_epi16(coeff1, 15);
__m128i qtmp0, qtmp1; qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
__m128i cmp_mask0, cmp_mask1; qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
// Do DC and first 15 AC qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
coeff0 = load_tran_low(coeff_ptr + n_coeffs); qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
coeff0_sign = _mm_srai_epi16(coeff0, 15); zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
coeff1_sign = _mm_srai_epi16(coeff1, 15); cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); qcoeff0 = _mm_adds_epi16(qcoeff0, round);
zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC round = _mm_unpackhi_epi64(round, round);
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
shift = _mm_unpackhi_epi64(shift, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
// Reinsert signs qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign); quant = _mm_unpackhi_epi64(quant, quant);
qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign); qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
// Mask out zbin threshold coeffs qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs); qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8); shift = _mm_unpackhi_epi64(shift, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant); // Reinsert signs
dequant = _mm_unpackhi_epi64(dequant, dequant); qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant); qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs); // Mask out zbin threshold coeffs
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8); qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
} qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
{ store_tran_low(qcoeff0, qcoeff_ptr);
// Scan for eob store_tran_low(qcoeff1, qcoeff_ptr + 8);
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1; coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
__m128i iscan0, iscan1; dequant = _mm_unpackhi_epi64(dequant, dequant);
__m128i eob1; coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); store_tran_low(coeff0, dqcoeff_ptr);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); store_tran_low(coeff1, dqcoeff_ptr + 8);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); // Scan for eob.
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
// Add one to convert from indices to counts zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8));
eob = _mm_and_si128(iscan0, nzero_coeff0); // Add one to convert from indices to counts
eob1 = _mm_and_si128(iscan1, nzero_coeff1); iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
eob = _mm_max_epi16(eob, eob1); iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
} eob = _mm_andnot_si128(zero_coeff0, iscan0);
n_coeffs += 8 * 2; eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
eob = _mm_max_epi16(eob, eob1);
// AC only loop.
while (index < n_coeffs) {
coeff0 = load_tran_low(coeff_ptr + index);
coeff1 = load_tran_low(coeff_ptr + index + 8);
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_tran_low(qcoeff0, qcoeff_ptr + index);
store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + index);
store_tran_low(coeff1, dqcoeff_ptr + index + 8);
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
eob0 = _mm_andnot_si128(zero_coeff0, iscan0);
eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
eob0 = _mm_max_epi16(eob0, eob1);
eob = _mm_max_epi16(eob, eob0);
index += 16;
} }
// AC only loop // Accumulate eob.
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
coeff0 = load_tran_low(coeff_ptr + n_coeffs);
coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
// Mask out zbin threshold coeffs
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob0, eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob0 = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob0 = _mm_max_epi16(eob0, eob1);
eob = _mm_max_epi16(eob, eob0);
}
n_coeffs += 8 * 2;
}
// Accumulate EOB
{ {
__m128i eob_shuffled; __m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe); eob_shuffled = _mm_shuffle_epi32(eob, 0xe);

146
vpx_dsp/x86/quantize_ssse3.c
View File

@@ -23,104 +23,88 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) { const int16_t *scan_ptr, const int16_t *iscan_ptr) {
const __m128i zero = _mm_setzero_si128(); const __m128i zero = _mm_setzero_si128();
intptr_t index = 16;
__m128i zbin, round, quant, dequant, shift;
__m128i coeff0, coeff1; __m128i coeff0, coeff1;
__m128i eob; __m128i qcoeff0, qcoeff1;
__m128i zbin; __m128i cmp_mask0, cmp_mask1;
__m128i round, quant, dequant, shift; __m128i qtmp0, qtmp1;
intptr_t index = 0; __m128i zero_coeff0, zero_coeff1, iscan0, iscan1;
__m128i eob, eob0, eob1;
(void)scan_ptr; (void)scan_ptr;
(void)skip_block; (void)skip_block;
assert(!skip_block); assert(!skip_block);
// Setup global values // Setup global values.
{ zbin = _mm_load_si128((const __m128i *)zbin_ptr);
const __m128i one = _mm_set1_epi16(1); // x86 has no "greater *or equal* comparison. Subtract 1 from zbin so
zbin = _mm_load_si128((const __m128i *)zbin_ptr); // it is a strict "greater" comparison.
// x86 has no "greater *or equal* comparison. Subtract 1 from zbin so zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1));
// it is a strict "greater" comparison. round = _mm_load_si128((const __m128i *)round_ptr);
zbin = _mm_sub_epi16(zbin, one); quant = _mm_load_si128((const __m128i *)quant_ptr);
round = _mm_load_si128((const __m128i *)round_ptr); dequant = _mm_load_si128((const __m128i *)dequant_ptr);
quant = _mm_load_si128((const __m128i *)quant_ptr); shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
dequant = _mm_load_si128((const __m128i *)dequant_ptr);
shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
}
{ // Do DC and first 15 AC.
__m128i qcoeff0, qcoeff1; coeff0 = load_tran_low(coeff_ptr);
__m128i qtmp0, qtmp1; coeff1 = load_tran_low(coeff_ptr + 8);
__m128i cmp_mask0, cmp_mask1;
__m128i zero_coeff0, zero_coeff1;
__m128i iscan0, iscan1;
__m128i eob1;
// Do DC and first 15 AC qcoeff0 = _mm_abs_epi16(coeff0);
coeff0 = load_tran_low(coeff_ptr + index); qcoeff1 = _mm_abs_epi16(coeff1);
coeff1 = load_tran_low(coeff_ptr + index + 8);
qcoeff0 = _mm_abs_epi16(coeff0); cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
qcoeff1 = _mm_abs_epi16(coeff1); zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); qcoeff0 = _mm_adds_epi16(qcoeff0, round);
// Overwrite DC component. round = _mm_unpackhi_epi64(round, round);
zbin = _mm_unpackhi_epi64(zbin, zbin); qcoeff1 = _mm_adds_epi16(qcoeff1, round);
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_adds_epi16(qcoeff0, round); qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
round = _mm_unpackhi_epi64(round, round); quant = _mm_unpackhi_epi64(quant, quant);
qcoeff1 = _mm_adds_epi16(qcoeff1, round); qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
quant = _mm_unpackhi_epi64(quant, quant); qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); shift = _mm_unpackhi_epi64(shift, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); // Reinsert signs
shift = _mm_unpackhi_epi64(shift, shift); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
// Reinsert signs // Mask out zbin threshold coeffs
qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
// Mask out zbin threshold coeffs store_tran_low(qcoeff0, qcoeff_ptr);
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); store_tran_low(qcoeff1, qcoeff_ptr + 8);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_tran_low(qcoeff0, qcoeff_ptr + index); coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
store_tran_low(qcoeff1, qcoeff_ptr + index + 8); dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant); store_tran_low(coeff0, dqcoeff_ptr);
dequant = _mm_unpackhi_epi64(dequant, dequant); store_tran_low(coeff1, dqcoeff_ptr + 8);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + index); // Scan for eob.
store_tran_low(coeff1, dqcoeff_ptr + index + 8); zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8));
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
eob = _mm_andnot_si128(zero_coeff0, iscan0);
eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
eob = _mm_max_epi16(eob, eob1);
// Scan for eob // AC only loop.
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
eob = _mm_andnot_si128(zero_coeff0, iscan0);
eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
eob = _mm_max_epi16(eob, eob1);
}
index += 16;
// AC only loop
while (index < n_coeffs) { while (index < n_coeffs) {
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
__m128i zero_coeff0, zero_coeff1;
__m128i iscan0, iscan1;
__m128i eob0, eob1;
coeff0 = load_tran_low(coeff_ptr + index); coeff0 = load_tran_low(coeff_ptr + index);
coeff1 = load_tran_low(coeff_ptr + index + 8); coeff1 = load_tran_low(coeff_ptr + index + 8);
@@ -142,11 +126,9 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
// Reinsert signs
qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0); qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1); qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
// Mask out zbin threshold coeffs
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
@@ -159,12 +141,10 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
store_tran_low(coeff0, dqcoeff_ptr + index); store_tran_low(coeff0, dqcoeff_ptr + index);
store_tran_low(coeff1, dqcoeff_ptr + index + 8); store_tran_low(coeff1, dqcoeff_ptr + index + 8);
// Scan for eob
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index)); iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8)); iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, cmp_mask0); iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
iscan1 = _mm_sub_epi16(iscan1, cmp_mask1); iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
eob0 = _mm_andnot_si128(zero_coeff0, iscan0); eob0 = _mm_andnot_si128(zero_coeff0, iscan0);
@@ -175,7 +155,7 @@ void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
index += 16; index += 16;
} }
// Accumulate EOB // Accumulate eob.
{ {
__m128i eob_shuffled; __m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe); eob_shuffled = _mm_shuffle_epi32(eob, 0xe);