Skip to content

GitLab

Commit c02fdd02 authored by Johann's avatar Johann
Browse files

quantize: ignore skip_block in x86 

Change-Id: I9a963e99f08761f0c8d6a305619270b2f1c4edf8
parent b527b473
Hide whitespace changes
Inline Side-by-side
Showing with 251 additions and 356 deletions
vpx_dsp/x86/highbd_quantize_intrin_sse2.c
View file @ c02fdd02
......@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
#include <assert.h>
#include <emmintrin.h>
#include "vpx_dsp/vpx_dsp_common.h"
......@@ -37,54 +38,54 @@ void vpx_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
(void)scan;
(void)skip_block;
assert(!skip_block);
memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
if (!skip_block) {
// Pre-scan pass
for (i = ((int)count / 4) - 1; i >= 0; i--) {
__m128i coeffs, cmp1, cmp2;
int test;
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
cmp1 = _mm_and_si128(cmp1, cmp2);
test = _mm_movemask_epi8(cmp1);
if (test == 0xffff)
non_zero_regs--;
else
break;
}
// Pre-scan pass
for (i = ((int)count / 4) - 1; i >= 0; i--) {
__m128i coeffs, cmp1, cmp2;
int test;
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
cmp1 = _mm_and_si128(cmp1, cmp2);
test = _mm_movemask_epi8(cmp1);
if (test == 0xffff)
non_zero_regs--;
else
break;
}
// Quantization pass:
for (i = 0; i < non_zero_regs; i++) {
__m128i coeffs, coeffs_sign, tmp1, tmp2;
int test;
int abs_coeff[4];
int coeff_sign[4];
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
coeffs_sign = _mm_srai_epi32(coeffs, 31);
coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
tmp1 = _mm_or_si128(tmp1, tmp2);
test = _mm_movemask_epi8(tmp1);
_mm_storeu_si128((__m128i *)abs_coeff, coeffs);
_mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
for (j = 0; j < 4; j++) {
if (test & (1 << (4 * j))) {
int k = 4 * i + j;
const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
const uint32_t abs_qcoeff =
(uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
}
// Quantization pass:
for (i = 0; i < non_zero_regs; i++) {
__m128i coeffs, coeffs_sign, tmp1, tmp2;
int test;
int abs_coeff[4];
int coeff_sign[4];
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
coeffs_sign = _mm_srai_epi32(coeffs, 31);
coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
tmp1 = _mm_or_si128(tmp1, tmp2);
test = _mm_movemask_epi8(tmp1);
_mm_storeu_si128((__m128i *)abs_coeff, coeffs);
_mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
for (j = 0; j < 4; j++) {
if (test & (1 << (4 * j))) {
int k = 4 * i + j;
const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
const uint32_t abs_qcoeff =
(uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
}
}
}
......@@ -105,6 +106,9 @@ void vpx_highbd_quantize_b_32x32_sse2(
const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
(void)scan;
(void)skip_block;
assert(!skip_block);
zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
zbins[1] = _mm_set1_epi32(zbin1_tmp);
......@@ -116,38 +120,35 @@ void vpx_highbd_quantize_b_32x32_sse2(
memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
if (!skip_block) {
// Pre-scan pass
for (i = 0; i < n_coeffs / 4; i++) {
__m128i coeffs, cmp1, cmp2;
int test;
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
cmp1 = _mm_and_si128(cmp1, cmp2);
test = _mm_movemask_epi8(cmp1);
if (!(test & 0xf)) idx_arr[idx++] = i * 4;
if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
}
// Pre-scan pass
for (i = 0; i < n_coeffs / 4; i++) {
__m128i coeffs, cmp1, cmp2;
int test;
coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
cmp1 = _mm_and_si128(cmp1, cmp2);
test = _mm_movemask_epi8(cmp1);
if (!(test & 0xf)) idx_arr[idx++] = i * 4;
if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
}
// Quantization pass: only process the coefficients selected in
// pre-scan pass. Note: idx can be zero.
for (i = 0; i < idx; i++) {
const int rc = idx_arr[i];
const int coeff = coeff_ptr[rc];
const int coeff_sign = (coeff >> 31);
const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
const int64_t tmp1 =
abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
const uint32_t abs_qcoeff =
(uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
}
// Quantization pass: only process the coefficients selected in
// pre-scan pass. Note: idx can be zero.
for (i = 0; i < idx; i++) {
const int rc = idx_arr[i];
const int coeff = coeff_ptr[rc];
const int coeff_sign = (coeff >> 31);
const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
const uint32_t abs_qcoeff =
(uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
}
*eob_ptr = eob + 1;
}
......
vpx_dsp/x86/quantize_avx_x86_64.asm
View file @ c02fdd02
......@@ -19,10 +19,6 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
vzeroupper
; If we can skip this block, then just zero the output
cmp skipmp, 0
jne .blank
%ifnidn %1, b_32x32
; Special case for ncoeff == 16, as it is frequent and we can save on
......@@ -493,48 +489,6 @@ DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
mov [r2], ax
vzeroupper
RET
; Skip-block, i.e. just write all zeroes
.blank:
DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
qcoeff, dqcoeff, dequant, eob, scan, iscan
mov r0, dqcoeffmp
movifnidn ncoeffq, ncoeffmp
mov r2, qcoeffmp
mov r3, eobmp
DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
%if CONFIG_VP9_HIGHBITDEPTH
lea dqcoeffq, [dqcoeffq+ncoeffq*4]
lea qcoeffq, [ qcoeffq+ncoeffq*4]
%else
lea dqcoeffq, [dqcoeffq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2]
%endif
neg ncoeffq
pxor m7, m7
.blank_loop:
%if CONFIG_VP9_HIGHBITDEPTH
mova [dqcoeffq+ncoeffq*4+ 0], ymm7
mova [dqcoeffq+ncoeffq*4+32], ymm7
mova [qcoeffq+ncoeffq*4+ 0], ymm7
mova [qcoeffq+ncoeffq*4+32], ymm7
%else
mova [dqcoeffq+ncoeffq*2+ 0], ymm7
mova [qcoeffq+ncoeffq*2+ 0], ymm7
%endif
add ncoeffq, mmsize
jl .blank_loop
mov [eobq], word 0
vzeroupper
RET
%endmacro
INIT_XMM avx
......
vpx_dsp/x86/quantize_sse2.c
View file @ c02fdd02
......@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
#include <assert.h>
#include <emmintrin.h>
#include <xmmintrin.h>
......@@ -23,7 +24,12 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) {
__m128i zero;
__m128i eob;
__m128i zbin;
__m128i round, quant, dequant, shift;
(void)scan_ptr;
(void)skip_block;
assert(!skip_block);
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
......@@ -31,193 +37,179 @@ void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
if (!skip_block) {
__m128i eob;
__m128i zbin;
__m128i round, quant, dequant, shift;
{
__m128i coeff0, coeff1;
// Setup global values
{
__m128i pw_1;
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);
}
{
__m128i coeff0, coeff1;
// Setup global values
{
__m128i pw_1;
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);
}
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
// Do DC and first 15 AC
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);
zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
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
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);
dequant = _mm_unpackhi_epi64(dequant, 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 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);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
// Do DC and first 15 AC
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);
zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
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
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);
dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
// AC only loop
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;
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i 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);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
}
// Accumulate EOB
// AC only loop
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
eob = _mm_max_epi16(eob, eob_shuffled);
*eob_ptr = _mm_extract_epi16(eob, 1);
__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);
}
} else {
do {
store_tran_low(zero, dqcoeff_ptr + n_coeffs);
store_tran_low(zero, dqcoeff_ptr + n_coeffs + 8);
store_tran_low(zero, qcoeff_ptr + n_coeffs);