Add SIMD support for ext-quant experiment
Change-Id: I2c46a098af9f70bd2668e07acc28c0bac01007ab
diff --git a/aom_dsp/aom_dsp_common.h b/aom_dsp/aom_dsp_common.h
index d3e2c6b..ee385b7 100644
--- a/aom_dsp/aom_dsp_common.h
+++ b/aom_dsp/aom_dsp_common.h
@@ -57,6 +57,7 @@
#if CONFIG_EXTQUANT
#define QUANT_TABLE_BITS 3
+#define QUANT_FP_BITS 4
#else
#define QUANT_TABLE_BITS 0
#endif // CONFIG_EXTQUANT
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 6e102e3..9ad5d4b 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -551,11 +551,19 @@
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
if (aom_config("CONFIG_EXTQUANT") eq "yes") {
add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- add_proto qw/void aom_highbd_quantize_b_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- add_proto qw/void aom_highbd_quantize_b_32x32_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- add_proto qw/void aom_highbd_quantize_b_64x64_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+ specialize qw/aom_highbd_quantize_b sse2 avx2/;
+
+ add_proto qw/void aom_highbd_quantize_b_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+
+ add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+ specialize qw/aom_highbd_quantize_b_32x32 sse2/;
+
+ add_proto qw/void aom_highbd_quantize_b_32x32_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+
+ add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+ specialize qw/aom_highbd_quantize_b_64x64 sse2/;
+
+ add_proto qw/void aom_highbd_quantize_b_64x64_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
} else {
add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/aom_highbd_quantize_b sse2 avx2/;
diff --git a/aom_dsp/x86/highbd_quantize_intrin_avx2.c b/aom_dsp/x86/highbd_quantize_intrin_avx2.c
index 469149c..c816b5f 100644
--- a/aom_dsp/x86/highbd_quantize_intrin_avx2.c
+++ b/aom_dsp/x86/highbd_quantize_intrin_avx2.c
@@ -15,11 +15,21 @@
#include "aom/aom_integer.h"
+#if CONFIG_EXTQUANT
+static INLINE void init_one_qp(const int32_t *p, __m256i *qp) {
+#else
static INLINE void init_one_qp(const __m128i *p, __m256i *qp) {
+#endif
+#if CONFIG_EXTQUANT
+ const int p1 = p[1];
+ const int p0 = p[0];
+ *qp = _mm256_set_epi32(p1, p1, p1, p1, p1, p1, p1, p0);
+#else
const __m128i sign = _mm_srai_epi16(*p, 15);
const __m128i dc = _mm_unpacklo_epi16(*p, sign);
const __m128i ac = _mm_unpackhi_epi16(*p, sign);
*qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1);
+#endif
}
static INLINE void update_qp(__m256i *qp) {
@@ -38,6 +48,13 @@
const int16_t *quant_ptr, const int16_t *dequant_ptr,
const int16_t *quant_shift_ptr, __m256i *qp) {
#endif
+#if CONFIG_EXTQUANT
+ init_one_qp(zbin_ptr, &qp[0]);
+ init_one_qp(round_ptr, &qp[1]);
+ init_one_qp(quant_ptr, &qp[2]);
+ init_one_qp(dequant_ptr, &qp[3]);
+ init_one_qp(quant_shift_ptr, &qp[4]);
+#else
const __m128i zbin = _mm_loadu_si128((const __m128i *)zbin_ptr);
const __m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
@@ -48,6 +65,7 @@
init_one_qp(&quant, &qp[2]);
init_one_qp(&dequant, &qp[3]);
init_one_qp(&quant_shift, &qp[4]);
+#endif
}
// Note:
@@ -77,7 +95,9 @@
__m256i flag2 = _mm256_cmpeq_epi32(abs, qp[0]);
flag2 = _mm256_or_si256(flag1, flag2);
const int32_t nzflag = _mm256_movemask_epi8(flag2);
-
+#if CONFIG_EXTQUANT
+ const __m256i offset = _mm256_set1_epi32((1 << QUANT_TABLE_BITS) >> 1);
+#endif
if (LIKELY(nzflag)) {
__m256i q = _mm256_add_epi32(abs, qp[1]);
__m256i tmp;
@@ -86,6 +106,10 @@
mm256_mul_shift_epi32(&q, &qp[4], &q);
__m256i dq = _mm256_mullo_epi32(q, qp[3]);
+#if CONFIG_EXTQUANT
+ dq = _mm256_add_epi32(dq, offset);
+ dq = _mm256_srli_epi32(dq, QUANT_TABLE_BITS);
+#endif
q = _mm256_sign_epi32(q, *c);
dq = _mm256_sign_epi32(dq, *c);
diff --git a/aom_dsp/x86/highbd_quantize_intrin_sse2.c b/aom_dsp/x86/highbd_quantize_intrin_sse2.c
index b2fe741..4f48663 100644
--- a/aom_dsp/x86/highbd_quantize_intrin_sse2.c
+++ b/aom_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -15,6 +15,16 @@
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"
+#if CONFIG_EXTQUANT
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
+ const int32_t *zbin_ptr,
+ const int32_t *round_ptr,
+ const int32_t *quant_ptr,
+ const int32_t *quant_shift_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int32_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+#else
void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
const int16_t *zbin_ptr,
const int16_t *round_ptr,
@@ -23,6 +33,7 @@
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) {
+#endif
int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
__m128i zbins[2];
__m128i nzbins[2];
@@ -82,8 +93,10 @@
(uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
#if CONFIG_EXTQUANT
+ const tran_low_t abs_dqcoeff = (tran_low_t)ROUND_POWER_OF_TWO_64(
+ abs_qcoeff * dequant_ptr[k != 0], QUANT_TABLE_BITS);
dqcoeff_ptr[k] =
- qcoeff_ptr[k] * dequant_ptr[k != 0] / (1 << QUANT_TABLE_BITS);
+ (tran_low_t)((abs_dqcoeff ^ coeff_sign[j]) - coeff_sign[j]);
#else
dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
#endif
@@ -94,12 +107,21 @@
*eob_ptr = eob_i + 1;
}
+#if CONFIG_EXTQUANT
+void aom_highbd_quantize_b_32x32_sse2(
+ const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr,
+ const int32_t *round_ptr, const int32_t *quant_ptr,
+ const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+#else
void aom_highbd_quantize_b_32x32_sse2(
const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) {
+#endif // CONFIG_EXTQUANT
__m128i zbins[2];
__m128i nzbins[2];
int idx = 0;
@@ -143,12 +165,19 @@
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;
+#if CONFIG_EXTQUANT
+ const int32_t abs_qcoeff =
+#else
const uint32_t abs_qcoeff =
+#endif
(uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
#if CONFIG_EXTQUANT
- dqcoeff_ptr[rc] =
- qcoeff_ptr[rc] * dequant_ptr[rc != 0] / (2 << QUANT_TABLE_BITS);
+ const tran_low_t abs_dqcoeff =
+ (tran_low_t)ROUND_POWER_OF_TWO_64(abs_qcoeff * dequant_ptr[rc != 0],
+ QUANT_TABLE_BITS) >>
+ 1;
+ dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
#else
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
#endif
@@ -157,12 +186,21 @@
*eob_ptr = eob + 1;
}
+#if CONFIG_EXTQUANT
+void aom_highbd_quantize_b_64x64_sse2(
+ const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr,
+ const int32_t *round_ptr, const int32_t *quant_ptr,
+ const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+#else
void aom_highbd_quantize_b_64x64_sse2(
const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) {
+#endif // CONFIG_EXTQUANT
__m128i zbins[2];
__m128i nzbins[2];
int idx = 0;
@@ -206,12 +244,19 @@
const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2);
const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+#if CONFIG_EXTQUANT
+ const int32_t abs_qcoeff =
+#else
const uint32_t abs_qcoeff =
+#endif
(uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 14);
qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
#if CONFIG_EXTQUANT
- dqcoeff_ptr[rc] =
- qcoeff_ptr[rc] * dequant_ptr[rc != 0] / (4 << QUANT_TABLE_BITS);
+ const tran_low_t abs_dqcoeff =
+ (tran_low_t)ROUND_POWER_OF_TWO_64(abs_qcoeff * dequant_ptr[rc != 0],
+ QUANT_TABLE_BITS) >>
+ 2;
+ dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
#else
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 4;
#endif
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index b61b431..b3d533f 100644
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -371,6 +371,7 @@
if (aom_config("CONFIG_EXTQUANT") eq "yes") {
add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int32_t *zbin_ptr, const int32_t *round_ptr, const int32_t *quant_ptr, const int32_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int32_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
+ specialize qw/av1_highbd_quantize_fp sse4_1 avx2/;
} else {
add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
specialize qw/av1_highbd_quantize_fp sse4_1 avx2/;
diff --git a/av1/encoder/av1_quantize.h b/av1/encoder/av1_quantize.h
index 1d1d0c0..314d4ae 100644
--- a/av1/encoder/av1_quantize.h
+++ b/av1/encoder/av1_quantize.h
@@ -43,7 +43,6 @@
const QUANT_PARAM *qparam);
#if CONFIG_EXTQUANT
-#define QUANT_FP_BITS 4
static const int qindex_10b_offset[] = {
0,
48,
diff --git a/av1/encoder/encodemb.c b/av1/encoder/encodemb.c
index d8809f1..285a106 100644
--- a/av1/encoder/encodemb.c
+++ b/av1/encoder/encodemb.c
@@ -297,8 +297,13 @@
if (qparam->xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
const int n_coeffs = av1_get_max_eob(txfm_param->tx_size);
if (LIKELY(!x->seg_skip_block)) {
+#if CONFIG_EXTQUANT
+ quant_func_list[qparam->xform_quant_idx][1](
+ coeff, n_coeffs, p, qcoeff, dqcoeff, eob, scan_order, qparam);
+#else
quant_func_list[qparam->xform_quant_idx][txfm_param->is_hbd](
coeff, n_coeffs, p, qcoeff, dqcoeff, eob, scan_order, qparam);
+#endif
} else {
av1_quantize_skip(n_coeffs, qcoeff, dqcoeff, eob);
}
diff --git a/av1/encoder/x86/av1_highbd_quantize_avx2.c b/av1/encoder/x86/av1_highbd_quantize_avx2.c
index cc8d6e2..bed144e 100644
--- a/av1/encoder/x86/av1_highbd_quantize_avx2.c
+++ b/av1/encoder/x86/av1_highbd_quantize_avx2.c
@@ -38,6 +38,22 @@
const int16_t *dequant_ptr, int log_scale,
__m256i *qp) {
#endif
+#if CONFIG_EXTQUANT
+ const int round1 = ROUND_POWER_OF_TWO(round_ptr[1], log_scale);
+ const int round0 = ROUND_POWER_OF_TWO(round_ptr[0], log_scale);
+ qp[0] = _mm256_set_epi32(round1, round1, round1, round1, round1, round1,
+ round1, round0);
+
+ const int quant1 = quant_ptr[1];
+ const int quant0 = quant_ptr[0];
+ qp[1] = _mm256_set_epi32(quant1, quant1, quant1, quant1, quant1, quant1,
+ quant1, quant0);
+
+ const int dequant1 = dequant_ptr[1];
+ const int dequant0 = dequant_ptr[0];
+ qp[2] = _mm256_set_epi32(dequant1, dequant1, dequant1, dequant1, dequant1,
+ dequant1, dequant1, dequant0);
+#else
__m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
if (log_scale) {
const __m128i round_scale = _mm_set1_epi16(1 << (15 - log_scale));
@@ -49,6 +65,7 @@
init_one_qp(&round, &qp[0]);
init_one_qp(&quant, &qp[1]);
init_one_qp(&dequant, &qp[2]);
+#endif
}
static INLINE void quantize(const __m256i *qp, __m256i *c,
@@ -56,14 +73,27 @@
tran_low_t *qcoeff, tran_low_t *dqcoeff,
__m256i *eob) {
const __m256i abs_coeff = _mm256_abs_epi32(*c);
+#if CONFIG_EXTQUANT
+ const __m256i round = _mm256_set1_epi64x((1 << QUANT_TABLE_BITS) >> 1);
+#endif
__m256i q = _mm256_add_epi32(abs_coeff, qp[0]);
__m256i q_lo = _mm256_mul_epi32(q, qp[1]);
__m256i q_hi = _mm256_srli_epi64(q, 32);
const __m256i qp_hi = _mm256_srli_epi64(qp[1], 32);
q_hi = _mm256_mul_epi32(q_hi, qp_hi);
+#if CONFIG_EXTQUANT
+ q_lo = _mm256_srli_epi64(q_lo, 16 - log_scale + QUANT_FP_BITS);
+ q_hi = _mm256_srli_epi64(q_hi, 16 - log_scale + QUANT_FP_BITS);
+#else
q_lo = _mm256_srli_epi64(q_lo, 16 - log_scale);
q_hi = _mm256_srli_epi64(q_hi, 16 - log_scale);
+#endif
+
+#if CONFIG_EXTQUANT
+ log_scale += QUANT_TABLE_BITS;
+#endif
+
q_hi = _mm256_slli_epi64(q_hi, 32);
q = _mm256_or_si256(q_lo, q_hi);
const __m256i abs_s = _mm256_slli_epi32(abs_coeff, 1 + log_scale);
@@ -71,6 +101,9 @@
q = _mm256_andnot_si256(mask, q);
__m256i dq = _mm256_mullo_epi32(q, qp[2]);
+#if CONFIG_EXTQUANT
+ dq = _mm256_add_epi64(dq, round);
+#endif
dq = _mm256_srai_epi32(dq, log_scale);
q = _mm256_sign_epi32(q, *c);
dq = _mm256_sign_epi32(dq, *c);
diff --git a/av1/encoder/x86/av1_highbd_quantize_sse4.c b/av1/encoder/x86/av1_highbd_quantize_sse4.c
index c9c2df1..bd9af6d 100644
--- a/av1/encoder/x86/av1_highbd_quantize_sse4.c
+++ b/av1/encoder/x86/av1_highbd_quantize_sse4.c
@@ -25,6 +25,9 @@
__m128i *sign) {
const __m128i zero = _mm_setzero_si128();
const __m128i one = _mm_set1_epi32(1);
+#if CONFIG_EXTQUANT
+ const __m128i round = _mm_set1_epi64x((1 << QUANT_TABLE_BITS) >> 1);
+#endif
*sign = _mm_cmplt_epi32(*coeff, zero);
*sign = _mm_or_si128(*sign, one);
@@ -37,6 +40,9 @@
qcoeff[0] = _mm_mul_epi32(qcoeff[0], param[1]);
qcoeff[0] = _mm_srli_epi64(qcoeff[0], shift);
dquan[0] = _mm_mul_epi32(qcoeff[0], param[2]);
+#if CONFIG_EXTQUANT
+ dquan[0] = _mm_add_epi64(dquan[0], round);
+#endif
dquan[0] = _mm_srli_epi64(dquan[0], scale);
const __m128i abs_s = _mm_slli_epi32(*coeff, 1 + scale);
qcoeff[2] = _mm_cmplt_epi32(abs_s, param[3]);
@@ -50,10 +56,16 @@
tran_low_t *dqAddr) {
__m128i mask0L = _mm_set_epi32(-1, -1, 0, 0);
__m128i mask0H = _mm_set_epi32(0, 0, -1, -1);
+#if CONFIG_EXTQUANT
+ const __m128i round = _mm_set1_epi64x((1 << QUANT_TABLE_BITS) >> 1);
+#endif
qcoeff[1] = _mm_mul_epi32(qcoeff[1], param[1]);
qcoeff[1] = _mm_srli_epi64(qcoeff[1], shift);
dquan[1] = _mm_mul_epi32(qcoeff[1], param[2]);
+#if CONFIG_EXTQUANT
+ dquan[1] = _mm_add_epi64(dquan[1], round);
+#endif
dquan[1] = _mm_srli_epi64(dquan[1], scale);
// combine L&H
@@ -132,7 +144,11 @@
const tran_low_t *src = coeff_ptr;
tran_low_t *quanAddr = qcoeff_ptr;
tran_low_t *dquanAddr = dqcoeff_ptr;
+#if CONFIG_EXTQUANT
+ const int shift = 16 - log_scale + QUANT_FP_BITS;
+#else
const int shift = 16 - log_scale;
+#endif
const int coeff_stride = 4;
const int quan_stride = coeff_stride;
(void)zbin_ptr;
@@ -151,6 +167,9 @@
qparam[2] = xx_set_64_from_32i(dequant_ptr[1], dequant_ptr[0]);
qparam[3] = _mm_set_epi32(dequant_ptr[1], dequant_ptr[1], dequant_ptr[1],
dequant_ptr[0]);
+#if CONFIG_EXTQUANT
+ log_scale += QUANT_TABLE_BITS;
+#endif
// DC and first 3 AC
quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff, dequant,