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

Merge "Revert "Revert "Removal of legacy zbin_extra / zbin_oq_value."""

Browse Source
This commit is contained in:
Jim Bankoski
2014-12-22 13:43:56 -08:00
committed by Gerrit Code Review
parent fba0ead543 d0f2377027
commit d6d431c476
14 changed files with 91 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
test/vp9_quantize_test.cc
View File

@@ -34,7 +34,7 @@ typedef void (*QuantizeFunc)(const tran_low_t *coeff, intptr_t count,
const int16_t *round, const int16_t *quant, const int16_t *round, const int16_t *quant,
const int16_t *quant_shift, const int16_t *quant_shift,
tran_low_t *qcoeff, tran_low_t *dqcoeff, tran_low_t *qcoeff, tran_low_t *dqcoeff,
const int16_t *dequant, int zbin_oq_value, const int16_t *dequant,
uint16_t *eob, const int16_t *scan, uint16_t *eob, const int16_t *scan,
const int16_t *iscan); const int16_t *iscan);
typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, vpx_bit_depth_t> typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, vpx_bit_depth_t>
@@ -80,7 +80,6 @@ class VP9Quantize32Test : public ::testing::TestWithParam<QuantizeParam> {
TEST_P(VP9QuantizeTest, OperationCheck) { TEST_P(VP9QuantizeTest, OperationCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed()); ACMRandom rnd(ACMRandom::DeterministicSeed());
int zbin_oq_value = 0;
DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 256); DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 256);
DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2);
DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2);
@@ -116,13 +115,12 @@ TEST_P(VP9QuantizeTest, OperationCheck) {
} }
ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
quant_ptr, quant_shift_ptr, ref_qcoeff_ptr, quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
ref_dqcoeff_ptr, dequant_ptr, zbin_oq_value, ref_dqcoeff_ptr, dequant_ptr,
ref_eob_ptr, scan_order->scan, scan_order->iscan); ref_eob_ptr, scan_order->scan, scan_order->iscan);
ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block, ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
zbin_ptr, round_ptr, quant_ptr, zbin_ptr, round_ptr, quant_ptr,
quant_shift_ptr, qcoeff_ptr, quant_shift_ptr, qcoeff_ptr,
dqcoeff_ptr, dequant_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr,
zbin_oq_value, eob_ptr,
scan_order->scan, scan_order->iscan)); scan_order->scan, scan_order->iscan));
for (int j = 0; j < sz; ++j) { for (int j = 0; j < sz; ++j) {
err_count += (ref_qcoeff_ptr[j] != qcoeff_ptr[j]) | err_count += (ref_qcoeff_ptr[j] != qcoeff_ptr[j]) |
@@ -141,7 +139,6 @@ TEST_P(VP9QuantizeTest, OperationCheck) {
TEST_P(VP9Quantize32Test, OperationCheck) { TEST_P(VP9Quantize32Test, OperationCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed()); ACMRandom rnd(ACMRandom::DeterministicSeed());
int zbin_oq_value = 0;
DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 1024); DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 1024);
DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2);
DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2);
@@ -177,13 +174,12 @@ TEST_P(VP9Quantize32Test, OperationCheck) {
} }
ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
quant_ptr, quant_shift_ptr, ref_qcoeff_ptr, quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
ref_dqcoeff_ptr, dequant_ptr, zbin_oq_value, ref_dqcoeff_ptr, dequant_ptr,
ref_eob_ptr, scan_order->scan, scan_order->iscan); ref_eob_ptr, scan_order->scan, scan_order->iscan);
ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block, ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
zbin_ptr, round_ptr, quant_ptr, zbin_ptr, round_ptr, quant_ptr,
quant_shift_ptr, qcoeff_ptr, quant_shift_ptr, qcoeff_ptr,
dqcoeff_ptr, dequant_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr,
zbin_oq_value, eob_ptr,
scan_order->scan, scan_order->iscan)); scan_order->scan, scan_order->iscan));
for (int j = 0; j < sz; ++j) { for (int j = 0; j < sz; ++j) {
err_count += (ref_qcoeff_ptr[j] != qcoeff_ptr[j]) | err_count += (ref_qcoeff_ptr[j] != qcoeff_ptr[j]) |
@@ -202,7 +198,6 @@ TEST_P(VP9Quantize32Test, OperationCheck) {
TEST_P(VP9QuantizeTest, EOBCheck) { TEST_P(VP9QuantizeTest, EOBCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed()); ACMRandom rnd(ACMRandom::DeterministicSeed());
int zbin_oq_value = 0;
DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 256); DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 256);
DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2);
DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2);
@@ -242,13 +237,12 @@ TEST_P(VP9QuantizeTest, EOBCheck) {
ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
quant_ptr, quant_shift_ptr, ref_qcoeff_ptr, quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
ref_dqcoeff_ptr, dequant_ptr, zbin_oq_value, ref_dqcoeff_ptr, dequant_ptr,
ref_eob_ptr, scan_order->scan, scan_order->iscan); ref_eob_ptr, scan_order->scan, scan_order->iscan);
ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block, ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
zbin_ptr, round_ptr, quant_ptr, zbin_ptr, round_ptr, quant_ptr,
quant_shift_ptr, qcoeff_ptr, quant_shift_ptr, qcoeff_ptr,
dqcoeff_ptr, dequant_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr,
zbin_oq_value, eob_ptr,
scan_order->scan, scan_order->iscan)); scan_order->scan, scan_order->iscan));
for (int j = 0; j < sz; ++j) { for (int j = 0; j < sz; ++j) {
@@ -268,7 +262,6 @@ TEST_P(VP9QuantizeTest, EOBCheck) {
TEST_P(VP9Quantize32Test, EOBCheck) { TEST_P(VP9Quantize32Test, EOBCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed()); ACMRandom rnd(ACMRandom::DeterministicSeed());
int zbin_oq_value = 0;
DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 1024); DECLARE_ALIGNED_ARRAY(16, tran_low_t, coeff_ptr, 1024);
DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, zbin_ptr, 2);
DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2); DECLARE_ALIGNED_ARRAY(16, int16_t, round_ptr, 2);
@@ -308,13 +301,12 @@ TEST_P(VP9Quantize32Test, EOBCheck) {
ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
quant_ptr, quant_shift_ptr, ref_qcoeff_ptr, quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
ref_dqcoeff_ptr, dequant_ptr, zbin_oq_value, ref_dqcoeff_ptr, dequant_ptr,
ref_eob_ptr, scan_order->scan, scan_order->iscan); ref_eob_ptr, scan_order->scan, scan_order->iscan);
ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block, ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
zbin_ptr, round_ptr, quant_ptr, zbin_ptr, round_ptr, quant_ptr,
quant_shift_ptr, qcoeff_ptr, quant_shift_ptr, qcoeff_ptr,
dqcoeff_ptr, dequant_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr,
zbin_oq_value, eob_ptr,
scan_order->scan, scan_order->iscan)); scan_order->scan, scan_order->iscan));
for (int j = 0; j < sz; ++j) { for (int j = 0; j < sz; ++j) {

28
vp9/common/vp9_rtcd_defs.pl
View File

@@ -1125,37 +1125,37 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz"; add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
specialize qw/vp9_block_error/; specialize qw/vp9_block_error/;
add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_fp/; specialize qw/vp9_quantize_fp/;
add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_fp_32x32/; specialize qw/vp9_quantize_fp_32x32/;
add_proto qw/void vp9_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_b/; specialize qw/vp9_quantize_b/;
add_proto qw/void vp9_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_b_32x32/; specialize qw/vp9_quantize_b_32x32/;
add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_fdct8x8_quant/; specialize qw/vp9_fdct8x8_quant/;
} else { } else {
add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz"; add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
specialize qw/vp9_block_error avx2/, "$sse2_x86inc"; specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_fp neon sse2/, "$ssse3_x86_64"; specialize qw/vp9_quantize_fp neon sse2/, "$ssse3_x86_64";
add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_fp_32x32/, "$ssse3_x86_64"; specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64";
add_proto qw/void vp9_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_b sse2/, "$ssse3_x86_64"; specialize qw/vp9_quantize_b sse2/, "$ssse3_x86_64";
add_proto qw/void vp9_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_quantize_b_32x32/, "$ssse3_x86_64"; specialize qw/vp9_quantize_b_32x32/, "$ssse3_x86_64";
add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_fdct8x8_quant sse2 ssse3/; specialize qw/vp9_fdct8x8_quant sse2 ssse3/;
} }
@@ -1850,16 +1850,16 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/void vp9_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd"; add_proto qw/void vp9_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
specialize qw/vp9_highbd_subtract_block/; specialize qw/vp9_highbd_subtract_block/;
add_proto qw/void vp9_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_highbd_quantize_fp/; specialize qw/vp9_highbd_quantize_fp/;
add_proto qw/void vp9_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_highbd_quantize_fp_32x32/; specialize qw/vp9_highbd_quantize_fp_32x32/;
add_proto qw/void vp9_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_highbd_quantize_b sse2/; specialize qw/vp9_highbd_quantize_b sse2/;
add_proto qw/void vp9_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; add_proto qw/void vp9_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vp9_highbd_quantize_b_32x32 sse2/; specialize qw/vp9_highbd_quantize_b_32x32 sse2/;
# #

3
vp9/encoder/arm/neon/vp9_quantize_neon.c
View File

@@ -26,13 +26,12 @@ void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
// TODO(jingning) Decide the need of these arguments after the // TODO(jingning) Decide the need of these arguments after the
// quantization process is completed. // quantization process is completed.
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)scan; (void)scan;
if (!skip_block) { if (!skip_block) {

2
vp9/encoder/vp9_block.h
View File

@@ -40,8 +40,6 @@ struct macroblock_plane {
int16_t *round; int16_t *round;
int64_t quant_thred[2]; int64_t quant_thred[2];
// Zbin Over Quant value
int16_t zbin_extra;
}; };
/* The [2] dimension is for whether we skip the EOB node (i.e. if previous /* The [2] dimension is for whether we skip the EOB node (i.e. if previous

3
vp9/encoder/vp9_dct.c
View File

@@ -339,7 +339,7 @@ void vp9_fdct8x8_quant_c(const int16_t *input, int stride,
const int16_t *quant_shift_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
int eob = -1; int eob = -1;
@@ -416,7 +416,6 @@ void vp9_fdct8x8_quant_c(const int16_t *input, int stride,
// quantization process is completed. // quantization process is completed.
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));

2
vp9/encoder/vp9_encodeframe.c
View File

@@ -3893,8 +3893,6 @@ static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
vp9_update_zbin_extra(x);
if (!is_inter_block(mbmi)) { if (!is_inter_block(mbmi)) {
int plane; int plane;
mbmi->skip = 1; mbmi->skip = 1;

49
vp9/encoder/vp9_encodemb.c
View File

@@ -386,28 +386,28 @@ void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
vp9_highbd_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin, vp9_highbd_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin,
p->round_fp, p->quant_fp, p->quant_shift, p->round_fp, p->quant_fp, p->quant_shift,
qcoeff, dqcoeff, pd->dequant, qcoeff, dqcoeff, pd->dequant,
p->zbin_extra, eob, scan_order->scan, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
break; break;
case TX_16X16: case TX_16X16:
vp9_highbd_fdct16x16(src_diff, coeff, diff_stride); vp9_highbd_fdct16x16(src_diff, coeff, diff_stride);
vp9_highbd_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp, vp9_highbd_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_8X8: case TX_8X8:
vp9_highbd_fdct8x8(src_diff, coeff, diff_stride); vp9_highbd_fdct8x8(src_diff, coeff, diff_stride);
vp9_highbd_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp, vp9_highbd_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_4X4: case TX_4X4:
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_highbd_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp, vp9_highbd_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
default: default:
@@ -422,28 +422,28 @@ void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride); fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin, p->round_fp, vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
break; break;
case TX_16X16: case TX_16X16:
vp9_fdct16x16(src_diff, coeff, diff_stride); vp9_fdct16x16(src_diff, coeff, diff_stride);
vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp, vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_8X8: case TX_8X8:
vp9_fdct8x8_quant(src_diff, diff_stride, coeff, 64, vp9_fdct8x8_quant(src_diff, diff_stride, coeff, 64,
x->skip_block, p->zbin, p->round_fp, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_4X4: case TX_4X4:
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp, vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
p->quant_fp, p->quant_shift, qcoeff, dqcoeff, p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
default: default:
@@ -556,28 +556,28 @@ void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride); highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, vp9_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin,
p->round, p->quant, p->quant_shift, qcoeff, p->round, p->quant, p->quant_shift, qcoeff,
dqcoeff, pd->dequant, p->zbin_extra, eob, dqcoeff, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_16X16: case TX_16X16:
vp9_highbd_fdct16x16(src_diff, coeff, diff_stride); vp9_highbd_fdct16x16(src_diff, coeff, diff_stride);
vp9_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_8X8: case TX_8X8:
vp9_highbd_fdct8x8(src_diff, coeff, diff_stride); vp9_highbd_fdct8x8(src_diff, coeff, diff_stride);
vp9_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_4X4: case TX_4X4:
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
default: default:
@@ -592,28 +592,28 @@ void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride); fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round, vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
break; break;
case TX_16X16: case TX_16X16:
vp9_fdct16x16(src_diff, coeff, diff_stride); vp9_fdct16x16(src_diff, coeff, diff_stride);
vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_8X8: case TX_8X8:
vp9_fdct8x8(src_diff, coeff, diff_stride); vp9_fdct8x8(src_diff, coeff, diff_stride);
vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
case TX_4X4: case TX_4X4:
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
break; break;
default: default:
@@ -844,8 +844,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride); highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, vp9_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin,
p->round, p->quant, p->quant_shift, p->round, p->quant, p->quant_shift,
qcoeff, dqcoeff, pd->dequant, qcoeff, dqcoeff, pd->dequant, eob,
p->zbin_extra, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
} }
if (!x->skip_encode && *eob) { if (!x->skip_encode && *eob) {
@@ -866,7 +865,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
vp9_highbd_fht16x16(src_diff, coeff, diff_stride, tx_type); vp9_highbd_fht16x16(src_diff, coeff, diff_stride, tx_type);
vp9_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
} }
if (!x->skip_encode && *eob) { if (!x->skip_encode && *eob) {
@@ -888,7 +887,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
vp9_highbd_fht8x8(src_diff, coeff, diff_stride, tx_type); vp9_highbd_fht8x8(src_diff, coeff, diff_stride, tx_type);
vp9_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
} }
if (!x->skip_encode && *eob) { if (!x->skip_encode && *eob) {
@@ -914,7 +913,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, vp9_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, pd->dequant, eob,
scan_order->scan, scan_order->iscan); scan_order->scan, scan_order->iscan);
} }
@@ -953,7 +952,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride); fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round, vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
} }
if (!x->skip_encode && *eob) if (!x->skip_encode && *eob)
@@ -973,7 +972,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
vp9_fht16x16(src_diff, coeff, diff_stride, tx_type); vp9_fht16x16(src_diff, coeff, diff_stride, tx_type);
vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff, p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
} }
if (!x->skip_encode && *eob) if (!x->skip_encode && *eob)
@@ -993,7 +992,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
vp9_fht8x8(src_diff, coeff, diff_stride, tx_type); vp9_fht8x8(src_diff, coeff, diff_stride, tx_type);
vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant, vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
p->quant_shift, qcoeff, dqcoeff, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
} }
if (!x->skip_encode && *eob) if (!x->skip_encode && *eob)
@@ -1017,7 +1016,7 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
x->fwd_txm4x4(src_diff, coeff, diff_stride); x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant, vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
p->quant_shift, qcoeff, dqcoeff, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan, pd->dequant, eob, scan_order->scan,
scan_order->iscan); scan_order->iscan);
} }

70
vp9/encoder/vp9_quantize.c
View File

@@ -122,14 +122,13 @@ void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
int i, eob = -1; int i, eob = -1;
// TODO(jingning) Decide the need of these arguments after the // TODO(jingning) Decide the need of these arguments after the
// quantization process is completed. // quantization process is completed.
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
@@ -168,7 +167,6 @@ void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value,
uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *scan,
const int16_t *iscan) { const int16_t *iscan) {
@@ -178,7 +176,6 @@ void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr,
// quantization process is completed. // quantization process is completed.
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
@@ -217,12 +214,11 @@ void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *quant_shift_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
int i, eob = -1; int i, eob = -1;
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
@@ -261,12 +257,11 @@ void vp9_highbd_quantize_fp_32x32_c(const tran_low_t *coeff_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
int i, eob = -1; int i, eob = -1;
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
@@ -302,13 +297,11 @@ void vp9_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
int i, non_zero_count = (int)n_coeffs, eob = -1; int i, non_zero_count = (int)n_coeffs, eob = -1;
const int zbins[2] = { zbin_ptr[0] + zbin_oq_value, const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
zbin_ptr[1] + zbin_oq_value }; const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
const int nzbins[2] = { zbins[0] * -1,
zbins[1] * -1 };
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
@@ -355,14 +348,12 @@ void vp9_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, int zbin_oq_value, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan, uint16_t *eob_ptr, const int16_t *scan,
const int16_t *iscan) { const int16_t *iscan) {
int i, non_zero_count = (int)n_coeffs, eob = -1; int i, non_zero_count = (int)n_coeffs, eob = -1;
const int zbins[2] = { zbin_ptr[0] + zbin_oq_value, const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
zbin_ptr[1] + zbin_oq_value }; const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
const int nzbins[2] = { zbins[0] * -1,
zbins[1] * -1 };
(void)iscan; (void)iscan;
vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); vpx_memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
@@ -412,10 +403,10 @@ void vp9_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *quant_shift_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0] + zbin_oq_value, 1), const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
ROUND_POWER_OF_TWO(zbin_ptr[1] + zbin_oq_value, 1) }; ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1}; const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
int idx = 0; int idx = 0;
@@ -471,10 +462,10 @@ void vp9_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value, uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) { const int16_t *scan, const int16_t *iscan) {
const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0] + zbin_oq_value, 1), const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
ROUND_POWER_OF_TWO(zbin_ptr[1] + zbin_oq_value, 1) }; ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1}; const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
int idx = 0; int idx = 0;
@@ -534,7 +525,7 @@ void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
p->zbin, p->round, p->quant, p->quant_shift, p->zbin, p->round, p->quant, p->quant_shift,
BLOCK_OFFSET(p->qcoeff, block), BLOCK_OFFSET(p->qcoeff, block),
BLOCK_OFFSET(pd->dqcoeff, block), BLOCK_OFFSET(pd->dqcoeff, block),
pd->dequant, p->zbin_extra, &p->eobs[block], pd->dequant, &p->eobs[block],
scan, iscan); scan, iscan);
return; return;
} }
@@ -544,7 +535,7 @@ void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
p->zbin, p->round, p->quant, p->quant_shift, p->zbin, p->round, p->quant, p->quant_shift,
BLOCK_OFFSET(p->qcoeff, block), BLOCK_OFFSET(p->qcoeff, block),
BLOCK_OFFSET(pd->dqcoeff, block), BLOCK_OFFSET(pd->dqcoeff, block),
pd->dequant, p->zbin_extra, &p->eobs[block], scan, iscan); pd->dequant, &p->eobs[block], scan, iscan);
} }
static void invert_quant(int16_t *quant, int16_t *shift, int d) { static void invert_quant(int16_t *quant, int16_t *shift, int d) {
@@ -641,8 +632,6 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
const int segment_id = xd->mi[0].src_mi->mbmi.segment_id; const int segment_id = xd->mi[0].src_mi->mbmi.segment_id;
const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex); const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q); const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
// TODO(paulwilkins): 0 value for zbin for now pending follow on patch.
const int zbin = 0;
int i; int i;
// Y // Y
@@ -652,13 +641,10 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
x->plane[0].quant_shift = quants->y_quant_shift[qindex]; x->plane[0].quant_shift = quants->y_quant_shift[qindex];
x->plane[0].zbin = quants->y_zbin[qindex]; x->plane[0].zbin = quants->y_zbin[qindex];
x->plane[0].round = quants->y_round[qindex]; x->plane[0].round = quants->y_round[qindex];
x->plane[0].zbin_extra = (int16_t)((cm->y_dequant[qindex][1] * zbin) >> 7);
xd->plane[0].dequant = cm->y_dequant[qindex]; xd->plane[0].dequant = cm->y_dequant[qindex];
x->plane[0].quant_thred[0] = (x->plane[0].zbin[0] + x->plane[0].zbin_extra) * x->plane[0].quant_thred[0] = x->plane[0].zbin[0] * x->plane[0].zbin[0];
(x->plane[0].zbin[0] + x->plane[0].zbin_extra); x->plane[0].quant_thred[1] = x->plane[0].zbin[1] * x->plane[0].zbin[1];
x->plane[0].quant_thred[1] = (x->plane[0].zbin[1] + x->plane[0].zbin_extra) *
(x->plane[0].zbin[1] + x->plane[0].zbin_extra);
// UV // UV
for (i = 1; i < 3; i++) { for (i = 1; i < 3; i++) {
@@ -668,15 +654,10 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
x->plane[i].quant_shift = quants->uv_quant_shift[qindex]; x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
x->plane[i].zbin = quants->uv_zbin[qindex]; x->plane[i].zbin = quants->uv_zbin[qindex];
x->plane[i].round = quants->uv_round[qindex]; x->plane[i].round = quants->uv_round[qindex];
x->plane[i].zbin_extra = (int16_t)((cm->uv_dequant[qindex][1] * zbin) >> 7);
xd->plane[i].dequant = cm->uv_dequant[qindex]; xd->plane[i].dequant = cm->uv_dequant[qindex];
x->plane[i].quant_thred[0] = x->plane[i].quant_thred[0] = x->plane[i].zbin[0] * x->plane[i].zbin[0];
(x->plane[i].zbin[0] + x->plane[i].zbin_extra) * x->plane[i].quant_thred[1] = x->plane[i].zbin[1] * x->plane[i].zbin[1];
(x->plane[i].zbin[0] + x->plane[i].zbin_extra);
x->plane[i].quant_thred[1] =
(x->plane[i].zbin[1] + x->plane[i].zbin_extra) *
(x->plane[i].zbin[1] + x->plane[i].zbin_extra);
} }
x->skip_block = vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP); x->skip_block = vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
@@ -688,15 +669,6 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
vp9_initialize_me_consts(cpi, x->q_index); vp9_initialize_me_consts(cpi, x->q_index);
} }
void vp9_update_zbin_extra(MACROBLOCK *x) {
const int y_zbin_extra = 0;
const int uv_zbin_extra = 0;
x->plane[0].zbin_extra = (int16_t)y_zbin_extra;
x->plane[1].zbin_extra = (int16_t)uv_zbin_extra;
x->plane[2].zbin_extra = (int16_t)uv_zbin_extra;
}
void vp9_frame_init_quantizer(VP9_COMP *cpi) { void vp9_frame_init_quantizer(VP9_COMP *cpi) {
vp9_init_plane_quantizers(cpi, &cpi->td.mb); vp9_init_plane_quantizers(cpi, &cpi->td.mb);
} }

2
vp9/encoder/vp9_quantize.h
View File

@@ -68,8 +68,6 @@ struct VP9Common;
void vp9_frame_init_quantizer(struct VP9_COMP *cpi); void vp9_frame_init_quantizer(struct VP9_COMP *cpi);
void vp9_update_zbin_extra(MACROBLOCK *x);
void vp9_init_plane_quantizers(struct VP9_COMP *cpi, MACROBLOCK *x); void vp9_init_plane_quantizers(struct VP9_COMP *cpi, MACROBLOCK *x);
void vp9_init_quantizer(struct VP9_COMP *cpi); void vp9_init_quantizer(struct VP9_COMP *cpi);

3
vp9/encoder/x86/vp9_dct_sse2.c
View File

@@ -254,7 +254,7 @@ void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
const int16_t* round_ptr, const int16_t* quant_ptr, const int16_t* round_ptr, const int16_t* quant_ptr,
const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr, const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
int16_t* dqcoeff_ptr, const int16_t* dequant_ptr, int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
int zbin_oq_value, uint16_t* eob_ptr, uint16_t* eob_ptr,
const int16_t* scan_ptr, const int16_t* scan_ptr,
const int16_t* iscan_ptr) { const int16_t* iscan_ptr) {
__m128i zero; __m128i zero;
@@ -287,7 +287,6 @@ void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
(void)scan_ptr; (void)scan_ptr;
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)coeff_ptr; (void)coeff_ptr;
// Pre-condition input (shift by two) // Pre-condition input (shift by two)

3
vp9/encoder/x86/vp9_dct_ssse3.c
View File

@@ -23,7 +23,7 @@ void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
const int16_t* quant_shift_ptr, const int16_t* quant_shift_ptr,
int16_t* qcoeff_ptr, int16_t* qcoeff_ptr,
int16_t* dqcoeff_ptr, const int16_t* dequant_ptr, int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
int zbin_oq_value, uint16_t* eob_ptr, uint16_t* eob_ptr,
const int16_t* scan_ptr, const int16_t* scan_ptr,
const int16_t* iscan_ptr) { const int16_t* iscan_ptr) {
__m128i zero; __m128i zero;
@@ -57,7 +57,6 @@ void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
(void)scan_ptr; (void)scan_ptr;
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
(void)coeff_ptr; (void)coeff_ptr;
// Pre-condition input (shift by two) // Pre-condition input (shift by two)

26
vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c
View File

@@ -24,7 +24,6 @@ void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value,
uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *scan,
const int16_t *iscan) { const int16_t *iscan) {
@@ -32,11 +31,11 @@ void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
__m128i zbins[2]; __m128i zbins[2];
__m128i nzbins[2]; __m128i nzbins[2];
zbins[0] = _mm_set_epi32((int)(zbin_ptr[1] + zbin_oq_value), zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
(int)(zbin_ptr[1] + zbin_oq_value), (int)zbin_ptr[1],
(int)(zbin_ptr[1] + zbin_oq_value), (int)zbin_ptr[1],
(int)(zbin_ptr[0] + zbin_oq_value)); (int)zbin_ptr[0]);
zbins[1] = _mm_set1_epi32((int)(zbin_ptr[1] + zbin_oq_value)); zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
nzbins[0] = _mm_setzero_si128(); nzbins[0] = _mm_setzero_si128();
nzbins[1] = _mm_setzero_si128(); nzbins[1] = _mm_setzero_si128();
@@ -111,7 +110,6 @@ void vp9_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, const int16_t *dequant_ptr,
int zbin_oq_value,
uint16_t *eob_ptr, uint16_t *eob_ptr,
const int16_t *scan, const int16_t *scan,
const int16_t *iscan) { const int16_t *iscan) {
@@ -120,14 +118,14 @@ void vp9_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr,
int idx = 0; int idx = 0;
int idx_arr[1024]; int idx_arr[1024];
int i, eob = -1; int i, eob = -1;
const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0] + zbin_oq_value, 1); const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1] + zbin_oq_value, 1); const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
(void)scan; (void)scan;
zbins[0] = _mm_set_epi32((zbin1_tmp + zbin_oq_value), zbins[0] = _mm_set_epi32(zbin1_tmp,
(zbin1_tmp + zbin_oq_value), zbin1_tmp,
(zbin1_tmp + zbin_oq_value), zbin1_tmp,
(zbin0_tmp + zbin_oq_value)); zbin0_tmp);
zbins[1] = _mm_set1_epi32((zbin1_tmp + zbin_oq_value)); zbins[1] = _mm_set1_epi32(zbin1_tmp);
nzbins[0] = _mm_setzero_si128(); nzbins[0] = _mm_setzero_si128();
nzbins[1] = _mm_setzero_si128(); nzbins[1] = _mm_setzero_si128();

8
vp9/encoder/x86/vp9_quantize_sse2.c
View File

@@ -18,7 +18,7 @@ void vp9_quantize_b_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
const int16_t* round_ptr, const int16_t* quant_ptr, const int16_t* round_ptr, const int16_t* quant_ptr,
const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr, const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
int16_t* dqcoeff_ptr, const int16_t* dequant_ptr, int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
int zbin_oq_value, uint16_t* eob_ptr, uint16_t* eob_ptr,
const int16_t* scan_ptr, const int16_t* scan_ptr,
const int16_t* iscan_ptr) { const int16_t* iscan_ptr) {
__m128i zero; __m128i zero;
@@ -39,13 +39,10 @@ void vp9_quantize_b_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
// Setup global values // Setup global values
{ {
__m128i zbin_oq;
__m128i pw_1; __m128i pw_1;
zbin_oq = _mm_set1_epi16(zbin_oq_value);
zbin = _mm_load_si128((const __m128i*)zbin_ptr); zbin = _mm_load_si128((const __m128i*)zbin_ptr);
round = _mm_load_si128((const __m128i*)round_ptr); round = _mm_load_si128((const __m128i*)round_ptr);
quant = _mm_load_si128((const __m128i*)quant_ptr); quant = _mm_load_si128((const __m128i*)quant_ptr);
zbin = _mm_add_epi16(zbin, zbin_oq);
pw_1 = _mm_set1_epi16(1); pw_1 = _mm_set1_epi16(1);
zbin = _mm_sub_epi16(zbin, pw_1); zbin = _mm_sub_epi16(zbin, pw_1);
dequant = _mm_load_si128((const __m128i*)dequant_ptr); dequant = _mm_load_si128((const __m128i*)dequant_ptr);
@@ -229,14 +226,13 @@ void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
const int16_t* round_ptr, const int16_t* quant_ptr, const int16_t* round_ptr, const int16_t* quant_ptr,
const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr, const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
int16_t* dqcoeff_ptr, const int16_t* dequant_ptr, int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
int zbin_oq_value, uint16_t* eob_ptr, uint16_t* eob_ptr,
const int16_t* scan_ptr, const int16_t* scan_ptr,
const int16_t* iscan_ptr) { const int16_t* iscan_ptr) {
__m128i zero; __m128i zero;
(void)scan_ptr; (void)scan_ptr;
(void)zbin_ptr; (void)zbin_ptr;
(void)quant_shift_ptr; (void)quant_shift_ptr;
(void)zbin_oq_value;
coeff_ptr += n_coeffs; coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs; iscan_ptr += n_coeffs;

12
vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
View File

@@ -17,7 +17,7 @@ SECTION .text
%macro QUANTIZE_FN 2 %macro QUANTIZE_FN 2
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, zbin_oq, \ shift, qcoeff, dqcoeff, dequant, \
eob, scan, iscan eob, scan, iscan
cmp dword skipm, 0 cmp dword skipm, 0
jne .blank jne .blank
@@ -29,13 +29,9 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
movifnidn zbinq, zbinmp movifnidn zbinq, zbinmp
movifnidn roundq, roundmp movifnidn roundq, roundmp
movifnidn quantq, quantmp movifnidn quantq, quantmp
movd m4, dword zbin_oqm ; m4 = zbin_oq
mova m0, [zbinq] ; m0 = zbin mova m0, [zbinq] ; m0 = zbin
punpcklwd m4, m4
mova m1, [roundq] ; m1 = round mova m1, [roundq] ; m1 = round
pshufd m4, m4, 0
mova m2, [quantq] ; m2 = quant mova m2, [quantq] ; m2 = quant
paddw m0, m4 ; m0 = zbin + zbin_oq
%ifidn %1, b_32x32 %ifidn %1, b_32x32
pcmpeqw m5, m5 pcmpeqw m5, m5
psrlw m5, 15 psrlw m5, 15
@@ -55,7 +51,7 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
psllw m4, 1 psllw m4, 1
%endif %endif
pxor m5, m5 ; m5 = dedicated zero pxor m5, m5 ; m5 = dedicated zero
DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, d6, eob DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
lea coeffq, [ coeffq+ncoeffq*2] lea coeffq, [ coeffq+ncoeffq*2]
lea iscanq, [ iscanq+ncoeffq*2] lea iscanq, [ iscanq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2] lea qcoeffq, [ qcoeffq+ncoeffq*2]
@@ -220,7 +216,7 @@ QUANTIZE_FN b_32x32, 7
%macro QUANTIZE_FP 2 %macro QUANTIZE_FP 2
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, zbin_oq, \ shift, qcoeff, dqcoeff, dequant, \
eob, scan, iscan eob, scan, iscan
cmp dword skipm, 0 cmp dword skipm, 0
jne .blank jne .blank
@@ -248,7 +244,7 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
psllw m2, 1 psllw m2, 1
%endif %endif
pxor m5, m5 ; m5 = dedicated zero pxor m5, m5 ; m5 = dedicated zero
DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, d6, eob DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
lea coeffq, [ coeffq+ncoeffq*2] lea coeffq, [ coeffq+ncoeffq*2]
lea iscanq, [ iscanq+ncoeffq*2] lea iscanq, [ iscanq+ncoeffq*2]
lea qcoeffq, [ qcoeffq+ncoeffq*2] lea qcoeffq, [ qcoeffq+ncoeffq*2]