ref: f65473c036a1270f360e1c1d323786257fae9f0e
parent: b0e6811ace3425a5ca7951448eb939c357acfb85
parent: 38f1fbbb759518356c6d8ef6f27bab4e6e263a1a
author: Yunqing Wang <yunqingwang@google.com>
date: Mon Jul 20 12:20:06 EDT 2015
Merge "Migrate quantization functions from vp9/ to vpx_dsp/"
--- a/test/vp9_quantize_test.cc
+++ b/test/vp9_quantize_test.cc
@@ -19,7 +19,7 @@
#include "test/register_state_check.h"
#include "test/util.h"
#include "./vpx_config.h"
-#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_scan.h"
#include "vpx/vpx_codec.h"
--- a/vp9/common/vp9_common.h
+++ b/vp9/common/vp9_common.h
@@ -56,20 +56,6 @@
return (uint16_t)clamp(val, 0, 4095);
}
}
-
-// Note:
-// tran_low_t is the datatype used for final transform coefficients.
-// tran_high_t is the datatype used for intermediate transform stages.
-typedef int64_t tran_high_t;
-typedef int32_t tran_low_t;
-
-#else
-
-// Note:
-// tran_low_t is the datatype used for final transform coefficients.
-// tran_high_t is the datatype used for intermediate transform stages.
-typedef int32_t tran_high_t;
-typedef int16_t tran_low_t;
#endif // CONFIG_VP9_HIGHBITDEPTH
#if CONFIG_DEBUG
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -781,12 +781,6 @@
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/;
- 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/;
-
- 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/;
-
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/;
} else {@@ -802,12 +796,6 @@
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/, "$ssse3_x86_64_x86inc";
- add_proto qw/void vp9_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/vp9_quantize_b sse2/, "$ssse3_x86_64_x86inc";
-
- 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_x86inc";
-
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 neon/;
}
@@ -934,12 +922,6 @@
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/;
-
- 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/;
-
- 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/;
#
# Structured Similarity (SSIM)
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -13,6 +13,7 @@
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/quantize.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
@@ -23,7 +24,6 @@
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/encoder/vp9_encodemb.h"
-#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_tokenize.h"
--- a/vp9/encoder/vp9_quantize.c
+++ b/vp9/encoder/vp9_quantize.c
@@ -9,7 +9,7 @@
*/
#include <math.h>
-
+#include "./vpx_dsp_rtcd.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
@@ -20,113 +20,6 @@
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_rd.h"
-void vp9_quantize_dc(const tran_low_t *coeff_ptr,
- int n_coeffs, int skip_block,
- const int16_t *round_ptr, const int16_t quant,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr) {- const int rc = 0;
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- int tmp, eob = -1;
-
- memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
-
- if (!skip_block) {- tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
- tmp = (tmp * quant) >> 16;
- qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
- if (tmp)
- eob = 0;
- }
- *eob_ptr = eob + 1;
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
- int n_coeffs, int skip_block,
- const int16_t *round_ptr, const int16_t quant,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr) {- int eob = -1;
-
- memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
-
- if (!skip_block) {- const int coeff = coeff_ptr[0];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- const int64_t tmp = abs_coeff + round_ptr[0];
- const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 16);
- qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
- dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr;
- if (abs_qcoeff)
- eob = 0;
- }
- *eob_ptr = eob + 1;
-}
-#endif
-
-void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
- const int16_t *round_ptr, const int16_t quant,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr) {- const int n_coeffs = 1024;
- const int rc = 0;
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- int tmp, eob = -1;
-
- memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
-
- if (!skip_block) {- tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
- INT16_MIN, INT16_MAX);
- tmp = (tmp * quant) >> 15;
- qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
- if (tmp)
- eob = 0;
- }
- *eob_ptr = eob + 1;
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
- int skip_block,
- const int16_t *round_ptr,
- const int16_t quant,
- tran_low_t *qcoeff_ptr,
- tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr,
- uint16_t *eob_ptr) {- const int n_coeffs = 1024;
- int eob = -1;
-
- memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
-
- if (!skip_block) {- const int coeff = coeff_ptr[0];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
- const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 15);
- qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
- dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 2;
- if (abs_qcoeff)
- eob = 0;
- }
- *eob_ptr = eob + 1;
-}
-#endif
-
void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block,
const int16_t *zbin_ptr, const int16_t *round_ptr,
@@ -292,224 +185,6 @@
if (abs_qcoeff)
eob = i;
- }
- }
- *eob_ptr = eob + 1;
-}
-#endif
-
-void vp9_quantize_b_c(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) {- int i, non_zero_count = (int)n_coeffs, eob = -1;
- const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};- const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};- (void)iscan;
-
- 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 = (int)n_coeffs - 1; i >= 0; i--) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
-
- if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
- non_zero_count--;
- else
- break;
- }
-
- // Quantization pass: All coefficients with index >= zero_flag are
- // skippable. Note: zero_flag can be zero.
- for (i = 0; i < non_zero_count; i++) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
-
- if (abs_coeff >= zbins[rc != 0]) {- int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
- tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
- quant_shift_ptr[rc != 0]) >> 16; // quantization
- qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
-
- if (tmp)
- eob = i;
- }
- }
- }
- *eob_ptr = eob + 1;
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-void vp9_highbd_quantize_b_c(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) {- int i, non_zero_count = (int)n_coeffs, eob = -1;
- const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};- const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};- (void)iscan;
-
- 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 = (int)n_coeffs - 1; i >= 0; i--) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
-
- if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
- non_zero_count--;
- else
- break;
- }
-
- // Quantization pass: All coefficients with index >= zero_flag are
- // skippable. Note: zero_flag can be zero.
- for (i = 0; i < non_zero_count; i++) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
-
- if (abs_coeff >= zbins[rc != 0]) {- const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
- const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
- const uint32_t abs_qcoeff =
- (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
- qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
- if (abs_qcoeff)
- eob = i;
- }
- }
- }
- *eob_ptr = eob + 1;
-}
-#endif
-
-void vp9_quantize_b_32x32_c(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) {- const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),- ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
- const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};-
- int idx = 0;
- int idx_arr[1024];
- int i, eob = -1;
- (void)iscan;
-
- 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; i++) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
-
- // If the coefficient is out of the base ZBIN range, keep it for
- // quantization.
- if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
- idx_arr[idx++] = i;
- }
-
- // 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 = scan[idx_arr[i]];
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- int tmp;
- int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
- abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
- tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
- quant_shift_ptr[rc != 0]) >> 15;
-
- qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
-
- if (tmp)
- eob = idx_arr[i];
- }
- }
- *eob_ptr = eob + 1;
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-void vp9_highbd_quantize_b_32x32_c(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) {- const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),- ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
- const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};-
- int idx = 0;
- int idx_arr[1024];
- int i, eob = -1;
- (void)iscan;
-
- 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; i++) {- const int rc = scan[i];
- const int coeff = coeff_ptr[rc];
-
- // If the coefficient is out of the base ZBIN range, keep it for
- // quantization.
- if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
- idx_arr[idx++] = i;
- }
-
- // 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 = scan[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] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
- if (abs_qcoeff)
- eob = idx_arr[i];
}
}
*eob_ptr = eob + 1;
--- a/vp9/encoder/vp9_quantize.h
+++ b/vp9/encoder/vp9_quantize.h
@@ -37,33 +37,8 @@
DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
} QUANTS;
-void vp9_quantize_dc(const tran_low_t *coeff_ptr,
- int n_coeffs, int skip_block,
- const int16_t *round_ptr, const int16_t quant_ptr,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr);
-void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
- const int16_t *round_ptr, const int16_t quant_ptr,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr);
void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
const int16_t *scan, const int16_t *iscan);
-
-#if CONFIG_VP9_HIGHBITDEPTH
-void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
- int n_coeffs, int skip_block,
- const int16_t *round_ptr, const int16_t quant_ptr,
- tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr, uint16_t *eob_ptr);
-void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
- int skip_block,
- const int16_t *round_ptr,
- const int16_t quant_ptr,
- tran_low_t *qcoeff_ptr,
- tran_low_t *dqcoeff_ptr,
- const int16_t dequant_ptr,
- uint16_t *eob_ptr);
-#endif
struct VP9_COMP;
struct VP9Common;
--- a/vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c
+++ /dev/null
@@ -1,179 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <emmintrin.h>
-
-#include "vpx_ports/mem.h"
-#include "vp9/common/vp9_common.h"
-
-#if CONFIG_VP9_HIGHBITDEPTH
-// from vp9_idct.h: typedef int32_t tran_low_t;
-void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
- intptr_t count,
- 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) {- int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
- __m128i zbins[2];
- __m128i nzbins[2];
-
- zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
- (int)zbin_ptr[1],
- (int)zbin_ptr[1],
- (int)zbin_ptr[0]);
- zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
-
- nzbins[0] = _mm_setzero_si128();
- nzbins[1] = _mm_setzero_si128();
- nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
- nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
-
- (void)scan;
-
- 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;
- }
-
- // 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 tmp1 = abs_coeff[j] + round_ptr[k != 0];
- const int64_t tmp2 = ((tmp1 * quant_ptr[k != 0]) >> 16) + tmp1;
- const uint32_t abs_qcoeff =
- (uint32_t)((tmp2 * 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;
- }
- }
- }
- }
- *eob_ptr = eob_i + 1;
-}
-
-
-void vp9_highbd_quantize_b_32x32_sse2(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) {- __m128i zbins[2];
- __m128i nzbins[2];
- int idx = 0;
- int idx_arr[1024];
- int i, eob = -1;
- 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;
- zbins[0] = _mm_set_epi32(zbin1_tmp,
- zbin1_tmp,
- zbin1_tmp,
- zbin0_tmp);
- zbins[1] = _mm_set1_epi32(zbin1_tmp);
-
- nzbins[0] = _mm_setzero_si128();
- nzbins[1] = _mm_setzero_si128();
- nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
- nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
-
- 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;
- }
-
- // 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;
-}
-#endif
--- a/vp9/encoder/x86/vp9_quantize_sse2.c
+++ b/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -14,214 +14,6 @@
#include "./vp9_rtcd.h"
#include "vpx/vpx_integer.h"
-void vp9_quantize_b_sse2(const int16_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, int16_t* qcoeff_ptr,
- int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
- uint16_t* eob_ptr,
- const int16_t* scan_ptr,
- const int16_t* iscan_ptr) {- __m128i zero;
- (void)scan_ptr;
-
- coeff_ptr += n_coeffs;
- iscan_ptr += n_coeffs;
- qcoeff_ptr += 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_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
- __m128i cmp_mask0, cmp_mask1;
- // Do DC and first 15 AC
- coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
- coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
-
- // 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);
-
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
-
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- dequant = _mm_unpackhi_epi64(dequant, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
- }
-
- {- // 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;
- }
-
- // 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 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
- coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
-
- // 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);
-
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
-
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
- }
-
- {- // Scan for eob
- __m128i zero_coeff0, zero_coeff1;
- __m128i nzero_coeff0, nzero_coeff1;
- __m128i iscan0, iscan1;
- __m128i eob0, eob1;
- zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
- zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
- nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
- nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
- iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
- iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
- // Add one to convert from indices to counts
- iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
- iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
- eob0 = _mm_and_si128(iscan0, nzero_coeff0);
- eob1 = _mm_and_si128(iscan1, nzero_coeff1);
- eob0 = _mm_max_epi16(eob0, eob1);
- eob = _mm_max_epi16(eob, eob0);
- }
- n_coeffs += 8 * 2;
- }
-
- // Accumulate EOB
- {- __m128i eob_shuffled;
- 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);
- }
- } else {- do {- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
- _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
- _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
- n_coeffs += 8 * 2;
- } while (n_coeffs < 0);
- *eob_ptr = 0;
- }
-}
-
void vp9_quantize_fp_sse2(const int16_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,
--- a/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
+++ b/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -15,206 +15,6 @@
SECTION .text
-; TODO(yunqingwang)fix quantize_b code for skip=1 case.
-%macro QUANTIZE_FN 2
-cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
- shift, qcoeff, dqcoeff, dequant, \
- eob, scan, iscan
- cmp dword skipm, 0
- jne .blank
-
- ; actual quantize loop - setup pointers, rounders, etc.
- movifnidn coeffq, coeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, dequantmp
- movifnidn zbinq, zbinmp
- movifnidn roundq, roundmp
- movifnidn quantq, quantmp
- mova m0, [zbinq] ; m0 = zbin
- mova m1, [roundq] ; m1 = round
- mova m2, [quantq] ; m2 = quant
-%ifidn %1, b_32x32
- pcmpeqw m5, m5
- psrlw m5, 15
- paddw m0, m5
- paddw m1, m5
- psrlw m0, 1 ; m0 = (m0 + 1) / 2
- psrlw m1, 1 ; m1 = (m1 + 1) / 2
-%endif
- mova m3, [r2q] ; m3 = dequant
- psubw m0, [pw_1]
- mov r2, shiftmp
- mov r3, qcoeffmp
- mova m4, [r2] ; m4 = shift
- mov r4, dqcoeffmp
- mov r5, iscanmp
-%ifidn %1, b_32x32
- psllw m4, 1
-%endif
- pxor m5, m5 ; m5 = dedicated zero
- DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
- lea coeffq, [ coeffq+ncoeffq*2]
- lea iscanq, [ iscanq+ncoeffq*2]
- lea qcoeffq, [ qcoeffq+ncoeffq*2]
- lea dqcoeffq, [dqcoeffq+ncoeffq*2]
- neg ncoeffq
-
- ; get DC and first 15 AC coeffs
- mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
- mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
- pabsw m6, m9 ; m6 = abs(m9)
- pabsw m11, m10 ; m11 = abs(m10)
- pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
- punpckhqdq m0, m0
- pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
- paddsw m6, m1 ; m6 += round
- punpckhqdq m1, m1
- paddsw m11, m1 ; m11 += round
- pmulhw m8, m6, m2 ; m8 = m6*q>>16
- punpckhqdq m2, m2
- pmulhw m13, m11, m2 ; m13 = m11*q>>16
- paddw m8, m6 ; m8 += m6
- paddw m13, m11 ; m13 += m11
- pmulhw m8, m4 ; m8 = m8*qsh>>16
- punpckhqdq m4, m4
- pmulhw m13, m4 ; m13 = m13*qsh>>16
- psignw m8, m9 ; m8 = reinsert sign
- psignw m13, m10 ; m13 = reinsert sign
- pand m8, m7
- pand m13, m12
- mova [qcoeffq+ncoeffq*2+ 0], m8
- mova [qcoeffq+ncoeffq*2+16], m13
-%ifidn %1, b_32x32
- pabsw m8, m8
- pabsw m13, m13
-%endif
- pmullw m8, m3 ; dqc[i] = qc[i] * q
- punpckhqdq m3, m3
- pmullw m13, m3 ; dqc[i] = qc[i] * q
-%ifidn %1, b_32x32
- psrlw m8, 1
- psrlw m13, 1
- psignw m8, m9
- psignw m13, m10
-%endif
- mova [dqcoeffq+ncoeffq*2+ 0], m8
- mova [dqcoeffq+ncoeffq*2+16], m13
- pcmpeqw m8, m5 ; m8 = c[i] == 0
- pcmpeqw m13, m5 ; m13 = c[i] == 0
- mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
- mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
- psubw m6, m7 ; m6 = scan[i] + 1
- psubw m11, m12 ; m11 = scan[i] + 1
- pandn m8, m6 ; m8 = max(eob)
- pandn m13, m11 ; m13 = max(eob)
- pmaxsw m8, m13
- add ncoeffq, mmsize
- jz .accumulate_eob
-
-.ac_only_loop:
- mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
- mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
- pabsw m6, m9 ; m6 = abs(m9)
- pabsw m11, m10 ; m11 = abs(m10)
- pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
- pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
-%ifidn %1, b_32x32
- pmovmskb r6d, m7
- pmovmskb r2d, m12
- or r6, r2
- jz .skip_iter
-%endif
- paddsw m6, m1 ; m6 += round
- paddsw m11, m1 ; m11 += round
- pmulhw m14, m6, m2 ; m14 = m6*q>>16
- pmulhw m13, m11, m2 ; m13 = m11*q>>16
- paddw m14, m6 ; m14 += m6
- paddw m13, m11 ; m13 += m11
- pmulhw m14, m4 ; m14 = m14*qsh>>16
- pmulhw m13, m4 ; m13 = m13*qsh>>16
- psignw m14, m9 ; m14 = reinsert sign
- psignw m13, m10 ; m13 = reinsert sign
- pand m14, m7
- pand m13, m12
- mova [qcoeffq+ncoeffq*2+ 0], m14
- mova [qcoeffq+ncoeffq*2+16], m13
-%ifidn %1, b_32x32
- pabsw m14, m14
- pabsw m13, m13
-%endif
- pmullw m14, m3 ; dqc[i] = qc[i] * q
- pmullw m13, m3 ; dqc[i] = qc[i] * q
-%ifidn %1, b_32x32
- psrlw m14, 1
- psrlw m13, 1
- psignw m14, m9
- psignw m13, m10
-%endif
- mova [dqcoeffq+ncoeffq*2+ 0], m14
- mova [dqcoeffq+ncoeffq*2+16], m13
- pcmpeqw m14, m5 ; m14 = c[i] == 0
- pcmpeqw m13, m5 ; m13 = c[i] == 0
- mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
- mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
- psubw m6, m7 ; m6 = scan[i] + 1
- psubw m11, m12 ; m11 = scan[i] + 1
- pandn m14, m6 ; m14 = max(eob)
- pandn m13, m11 ; m13 = max(eob)
- pmaxsw m8, m14
- pmaxsw m8, m13
- add ncoeffq, mmsize
- jl .ac_only_loop
-
-%ifidn %1, b_32x32
- jmp .accumulate_eob
-.skip_iter:
- mova [qcoeffq+ncoeffq*2+ 0], m5
- mova [qcoeffq+ncoeffq*2+16], m5
- mova [dqcoeffq+ncoeffq*2+ 0], m5
- mova [dqcoeffq+ncoeffq*2+16], m5
- add ncoeffq, mmsize
- jl .ac_only_loop
-%endif
-
-.accumulate_eob:
- ; horizontally accumulate/max eobs and write into [eob] memory pointer
- mov r2, eobmp
- pshufd m7, m8, 0xe
- pmaxsw m8, m7
- pshuflw m7, m8, 0xe
- pmaxsw m8, m7
- pshuflw m7, m8, 0x1
- pmaxsw m8, m7
- pextrw r6, m8, 0
- mov [r2], r6
- RET
-
- ; skip-block, i.e. just write all zeroes
-.blank:
- mov r0, dqcoeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, qcoeffmp
- mov r3, eobmp
- DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
- lea dqcoeffq, [dqcoeffq+ncoeffq*2]
- lea qcoeffq, [ qcoeffq+ncoeffq*2]
- neg ncoeffq
- pxor m7, m7
-.blank_loop:
- mova [dqcoeffq+ncoeffq*2+ 0], m7
- mova [dqcoeffq+ncoeffq*2+16], m7
- mova [qcoeffq+ncoeffq*2+ 0], m7
- mova [qcoeffq+ncoeffq*2+16], m7
- add ncoeffq, mmsize
- jl .blank_loop
- mov word [eobq], 0
- RET
-%endmacro
-
-INIT_XMM ssse3
-QUANTIZE_FN b, 7
-QUANTIZE_FN b_32x32, 7
-
%macro QUANTIZE_FP 2
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, \
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -104,7 +104,6 @@
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
-VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_quantize_intrin_sse2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_block_error_intrin_sse2.c
endif
--- /dev/null
+++ b/vpx_dsp/quantize.c
@@ -1,0 +1,337 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/quantize.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp9_quantize_dc(const tran_low_t *coeff_ptr,
+ int n_coeffs, int skip_block,
+ const int16_t *round_ptr, const int16_t quant,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr) {+ const int rc = 0;
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ int tmp, eob = -1;
+
+ memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+ if (!skip_block) {+ tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+ tmp = (tmp * quant) >> 16;
+ qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+ if (tmp)
+ eob = 0;
+ }
+ *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+ int n_coeffs, int skip_block,
+ const int16_t *round_ptr, const int16_t quant,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr) {+ int eob = -1;
+
+ memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+ if (!skip_block) {+ const int coeff = coeff_ptr[0];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ const int64_t tmp = abs_coeff + round_ptr[0];
+ const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 16);
+ qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr;
+ if (abs_qcoeff)
+ eob = 0;
+ }
+ *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+ const int16_t *round_ptr, const int16_t quant,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr) {+ const int n_coeffs = 1024;
+ const int rc = 0;
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ int tmp, eob = -1;
+
+ memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+ if (!skip_block) {+ tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+ INT16_MIN, INT16_MAX);
+ tmp = (tmp * quant) >> 15;
+ qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+ if (tmp)
+ eob = 0;
+ }
+ *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+ int skip_block,
+ const int16_t *round_ptr,
+ const int16_t quant,
+ tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr,
+ uint16_t *eob_ptr) {+ const int n_coeffs = 1024;
+ int eob = -1;
+
+ memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+ memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+ if (!skip_block) {+ const int coeff = coeff_ptr[0];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
+ const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 15);
+ qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 2;
+ if (abs_qcoeff)
+ eob = 0;
+ }
+ *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_b_c(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) {+ int i, non_zero_count = (int)n_coeffs, eob = -1;
+ const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};+ const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};+ (void)iscan;
+
+ 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 = (int)n_coeffs - 1; i >= 0; i--) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+
+ if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+ non_zero_count--;
+ else
+ break;
+ }
+
+ // Quantization pass: All coefficients with index >= zero_flag are
+ // skippable. Note: zero_flag can be zero.
+ for (i = 0; i < non_zero_count; i++) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+ if (abs_coeff >= zbins[rc != 0]) {+ int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+ tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+ quant_shift_ptr[rc != 0]) >> 16; // quantization
+ qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+ if (tmp)
+ eob = i;
+ }
+ }
+ }
+ *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_b_c(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) {+ int i, non_zero_count = (int)n_coeffs, eob = -1;
+ const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};+ const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};+ (void)iscan;
+
+ 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 = (int)n_coeffs - 1; i >= 0; i--) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+
+ if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+ non_zero_count--;
+ else
+ break;
+ }
+
+ // Quantization pass: All coefficients with index >= zero_flag are
+ // skippable. Note: zero_flag can be zero.
+ for (i = 0; i < non_zero_count; i++) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+ if (abs_coeff >= zbins[rc != 0]) {+ const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
+ const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
+ qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+ if (abs_qcoeff)
+ eob = i;
+ }
+ }
+ }
+ *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_b_32x32_c(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) {+ const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),+ ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+ const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};+
+ int idx = 0;
+ int idx_arr[1024];
+ int i, eob = -1;
+ (void)iscan;
+
+ 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; i++) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+
+ // If the coefficient is out of the base ZBIN range, keep it for
+ // quantization.
+ if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+ idx_arr[idx++] = i;
+ }
+
+ // 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 = scan[idx_arr[i]];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ int tmp;
+ int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+ abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+ tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
+ quant_shift_ptr[rc != 0]) >> 15;
+
+ qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+ if (tmp)
+ eob = idx_arr[i];
+ }
+ }
+ *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_b_32x32_c(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) {+ const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),+ ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+ const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};+
+ int idx = 0;
+ int idx_arr[1024];
+ int i, eob = -1;
+ (void)iscan;
+
+ 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; i++) {+ const int rc = scan[i];
+ const int coeff = coeff_ptr[rc];
+
+ // If the coefficient is out of the base ZBIN range, keep it for
+ // quantization.
+ if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+ idx_arr[idx++] = i;
+ }
+
+ // 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 = scan[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] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+ if (abs_qcoeff)
+ eob = idx_arr[i];
+ }
+ }
+ *eob_ptr = eob + 1;
+}
+#endif
--- /dev/null
+++ b/vpx_dsp/quantize.h
@@ -1,0 +1,51 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_QUANTIZE_H_
+#define VPX_DSP_QUANTIZE_H_
+
+#include "./vpx_config.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {+#endif
+
+void vp9_quantize_dc(const tran_low_t *coeff_ptr,
+ int n_coeffs, int skip_block,
+ const int16_t *round_ptr, const int16_t quant_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+ const int16_t *round_ptr, const int16_t quant_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+ int n_coeffs, int skip_block,
+ const int16_t *round_ptr, const int16_t quant_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+ int skip_block,
+ const int16_t *round_ptr,
+ const int16_t quant_ptr,
+ tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr,
+ const int16_t dequant_ptr,
+ uint16_t *eob_ptr);
+#endif
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+#endif // VPX_DSP_QUANTIZE_H_
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -54,6 +54,21 @@
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_loopfilter_sse2.c
endif # CONFIG_VP9_HIGHBITDEPTH
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+DSP_SRCS-yes += quantize.c
+DSP_SRCS-yes += quantize.h
+
+DSP_SRCS-$(HAVE_SSE2) += x86/quantize_sse2.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_quantize_intrin_sse2.c
+endif
+ifeq ($(ARCH_X86_64),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSSE3) += x86/quantize_ssse3_x86_64.asm
+endif
+endif
+endif # CONFIG_VP9_ENCODER
+
ifeq ($(CONFIG_ENCODERS),yes)
DSP_SRCS-yes += sad.c
DSP_SRCS-yes += subtract.c
--- a/vpx_dsp/vpx_dsp_common.h
+++ b/vpx_dsp/vpx_dsp_common.h
@@ -24,6 +24,20 @@
#define MIN(x, y) (((x) < (y)) ? (x) : (y))
#define MAX(x, y) (((x) > (y)) ? (x) : (y))
+#if CONFIG_VP9_HIGHBITDEPTH
+// Note:
+// tran_low_t is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+#else
+// Note:
+// tran_low_t is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
static INLINE uint8_t clip_pixel(int val) {return (val > 255) ? 255 : (val < 0) ? 0 : val;
}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -5,12 +5,18 @@
*/
#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
EOF
}
forward_decls qw/vpx_dsp_forward_decls/;
-# Functions which use x86inc.asm instead of x86_abi_support.asm
+# x86inc.asm had specific constraints. break it out so it's easy to disable.
+# zero all the variables to avoid tricky else conditions.
+$mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc = $avx_x86inc =
+ $avx2_x86inc = '';
+$mmx_x86_64_x86inc = $sse_x86_64_x86inc = $sse2_x86_64_x86inc =
+ $ssse3_x86_64_x86inc = $avx_x86_64_x86inc = $avx2_x86_64_x86inc = '';
if (vpx_config("CONFIG_USE_X86INC") eq "yes") {$mmx_x86inc = 'mmx';
$sse_x86inc = 'sse';
@@ -18,12 +24,18 @@
$ssse3_x86inc = 'ssse3';
$avx_x86inc = 'avx';
$avx2_x86inc = 'avx2';
-} else {- $mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc =
- $avx_x86inc = $avx2_x86inc = '';
+ if ($opts{arch} eq "x86_64") {+ $mmx_x86_64_x86inc = 'mmx';
+ $sse_x86_64_x86inc = 'sse';
+ $sse2_x86_64_x86inc = 'sse2';
+ $ssse3_x86_64_x86inc = 'ssse3';
+ $avx_x86_64_x86inc = 'avx';
+ $avx2_x86_64_x86inc = 'avx2';
+ }
}
-# Functions which are 64 bit only.
+# functions that are 64 bit only.
+$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
if ($opts{arch} eq "x86_64") {$mmx_x86_64 = 'mmx';
$sse2_x86_64 = 'sse2';
@@ -30,9 +42,6 @@
$ssse3_x86_64 = 'ssse3';
$avx_x86_64 = 'avx';
$avx2_x86_64 = 'avx2';
-} else {- $mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 =
- $avx_x86_64 = $avx2_x86_64 = '';
}
#
@@ -109,6 +118,31 @@
add_proto qw/void vpx_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2/;
+} # CONFIG_VP9_HIGHBITDEPTH
+
+#
+# Encoder functions.
+#
+if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {+ 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/;
+
+ 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/;
+
+ 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/;
+
+ 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/;
+} else {+ 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_x86inc";
+
+ 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_x86inc";
+} # CONFIG_VP9_ENCODER
} # CONFIG_VP9_HIGHBITDEPTH
if (vpx_config("CONFIG_ENCODERS") eq "yes") {--- /dev/null
+++ b/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -1,0 +1,180 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// from vp9_idct.h: typedef int32_t tran_low_t;
+void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
+ intptr_t count,
+ 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) {+ int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+ __m128i zbins[2];
+ __m128i nzbins[2];
+
+ zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
+ (int)zbin_ptr[1],
+ (int)zbin_ptr[1],
+ (int)zbin_ptr[0]);
+ zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+ nzbins[0] = _mm_setzero_si128();
+ nzbins[1] = _mm_setzero_si128();
+ nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+ nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+ (void)scan;
+
+ 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;
+ }
+
+ // 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 tmp1 = abs_coeff[j] + round_ptr[k != 0];
+ const int64_t tmp2 = ((tmp1 * quant_ptr[k != 0]) >> 16) + tmp1;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp2 * 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;
+ }
+ }
+ }
+ }
+ *eob_ptr = eob_i + 1;
+}
+
+
+void vp9_highbd_quantize_b_32x32_sse2(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) {+ __m128i zbins[2];
+ __m128i nzbins[2];
+ int idx = 0;
+ int idx_arr[1024];
+ int i, eob = -1;
+ 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;
+ zbins[0] = _mm_set_epi32(zbin1_tmp,
+ zbin1_tmp,
+ zbin1_tmp,
+ zbin0_tmp);
+ zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+ nzbins[0] = _mm_setzero_si128();
+ nzbins[1] = _mm_setzero_si128();
+ nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+ nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+ 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;
+ }
+
+ // 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;
+}
+#endif
--- /dev/null
+++ b/vpx_dsp/x86/quantize_sse2.c
@@ -1,0 +1,223 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vp9_quantize_b_sse2(const int16_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, int16_t* qcoeff_ptr,
+ int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+ uint16_t* eob_ptr,
+ const int16_t* scan_ptr,
+ const int16_t* iscan_ptr) {+ __m128i zero;
+ (void)scan_ptr;
+
+ coeff_ptr += n_coeffs;
+ iscan_ptr += n_coeffs;
+ qcoeff_ptr += 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_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+ __m128i cmp_mask0, cmp_mask1;
+ // Do DC and first 15 AC
+ coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+ coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+ // 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);
+
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+ }
+
+ {+ // 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;
+ }
+
+ // 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 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+ coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+ // 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);
+
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+ }
+
+ {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob0, eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob0 = _mm_max_epi16(eob0, eob1);
+ eob = _mm_max_epi16(eob, eob0);
+ }
+ n_coeffs += 8 * 2;
+ }
+
+ // Accumulate EOB
+ {+ __m128i eob_shuffled;
+ 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);
+ }
+ } else {+ do {+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+ _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+ _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+ n_coeffs += 8 * 2;
+ } while (n_coeffs < 0);
+ *eob_ptr = 0;
+ }
+}
--- /dev/null
+++ b/vpx_dsp/x86/quantize_ssse3_x86_64.asm
@@ -1,0 +1,216 @@
+;
+; Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+; TODO(yunqingwang)fix quantize_b code for skip=1 case.
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+ shift, qcoeff, dqcoeff, dequant, \
+ eob, scan, iscan
+ cmp dword skipm, 0
+ jne .blank
+
+ ; actual quantize loop - setup pointers, rounders, etc.
+ movifnidn coeffq, coeffmp
+ movifnidn ncoeffq, ncoeffmp
+ mov r2, dequantmp
+ movifnidn zbinq, zbinmp
+ movifnidn roundq, roundmp
+ movifnidn quantq, quantmp
+ mova m0, [zbinq] ; m0 = zbin
+ mova m1, [roundq] ; m1 = round
+ mova m2, [quantq] ; m2 = quant
+%ifidn %1, b_32x32
+ pcmpeqw m5, m5
+ psrlw m5, 15
+ paddw m0, m5
+ paddw m1, m5
+ psrlw m0, 1 ; m0 = (m0 + 1) / 2
+ psrlw m1, 1 ; m1 = (m1 + 1) / 2
+%endif
+ mova m3, [r2q] ; m3 = dequant
+ psubw m0, [pw_1]
+ mov r2, shiftmp
+ mov r3, qcoeffmp
+ mova m4, [r2] ; m4 = shift
+ mov r4, dqcoeffmp
+ mov r5, iscanmp
+%ifidn %1, b_32x32
+ psllw m4, 1
+%endif
+ pxor m5, m5 ; m5 = dedicated zero
+ DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+ lea coeffq, [ coeffq+ncoeffq*2]
+ lea iscanq, [ iscanq+ncoeffq*2]
+ lea qcoeffq, [ qcoeffq+ncoeffq*2]
+ lea dqcoeffq, [dqcoeffq+ncoeffq*2]
+ neg ncoeffq
+
+ ; get DC and first 15 AC coeffs
+ mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+ mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
+ pabsw m6, m9 ; m6 = abs(m9)
+ pabsw m11, m10 ; m11 = abs(m10)
+ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
+ punpckhqdq m0, m0
+ pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
+ paddsw m6, m1 ; m6 += round
+ punpckhqdq m1, m1
+ paddsw m11, m1 ; m11 += round
+ pmulhw m8, m6, m2 ; m8 = m6*q>>16
+ punpckhqdq m2, m2
+ pmulhw m13, m11, m2 ; m13 = m11*q>>16
+ paddw m8, m6 ; m8 += m6
+ paddw m13, m11 ; m13 += m11
+ pmulhw m8, m4 ; m8 = m8*qsh>>16
+ punpckhqdq m4, m4
+ pmulhw m13, m4 ; m13 = m13*qsh>>16
+ psignw m8, m9 ; m8 = reinsert sign
+ psignw m13, m10 ; m13 = reinsert sign
+ pand m8, m7
+ pand m13, m12
+ mova [qcoeffq+ncoeffq*2+ 0], m8
+ mova [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, b_32x32
+ pabsw m8, m8
+ pabsw m13, m13
+%endif
+ pmullw m8, m3 ; dqc[i] = qc[i] * q
+ punpckhqdq m3, m3
+ pmullw m13, m3 ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+ psrlw m8, 1
+ psrlw m13, 1
+ psignw m8, m9
+ psignw m13, m10
+%endif
+ mova [dqcoeffq+ncoeffq*2+ 0], m8
+ mova [dqcoeffq+ncoeffq*2+16], m13
+ pcmpeqw m8, m5 ; m8 = c[i] == 0
+ pcmpeqw m13, m5 ; m13 = c[i] == 0
+ mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+ mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
+ psubw m6, m7 ; m6 = scan[i] + 1
+ psubw m11, m12 ; m11 = scan[i] + 1
+ pandn m8, m6 ; m8 = max(eob)
+ pandn m13, m11 ; m13 = max(eob)
+ pmaxsw m8, m13
+ add ncoeffq, mmsize
+ jz .accumulate_eob
+
+.ac_only_loop:
+ mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+ mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
+ pabsw m6, m9 ; m6 = abs(m9)
+ pabsw m11, m10 ; m11 = abs(m10)
+ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
+ pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
+%ifidn %1, b_32x32
+ pmovmskb r6d, m7
+ pmovmskb r2d, m12
+ or r6, r2
+ jz .skip_iter
+%endif
+ paddsw m6, m1 ; m6 += round
+ paddsw m11, m1 ; m11 += round
+ pmulhw m14, m6, m2 ; m14 = m6*q>>16
+ pmulhw m13, m11, m2 ; m13 = m11*q>>16
+ paddw m14, m6 ; m14 += m6
+ paddw m13, m11 ; m13 += m11
+ pmulhw m14, m4 ; m14 = m14*qsh>>16
+ pmulhw m13, m4 ; m13 = m13*qsh>>16
+ psignw m14, m9 ; m14 = reinsert sign
+ psignw m13, m10 ; m13 = reinsert sign
+ pand m14, m7
+ pand m13, m12
+ mova [qcoeffq+ncoeffq*2+ 0], m14
+ mova [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, b_32x32
+ pabsw m14, m14
+ pabsw m13, m13
+%endif
+ pmullw m14, m3 ; dqc[i] = qc[i] * q
+ pmullw m13, m3 ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+ psrlw m14, 1
+ psrlw m13, 1
+ psignw m14, m9
+ psignw m13, m10
+%endif
+ mova [dqcoeffq+ncoeffq*2+ 0], m14
+ mova [dqcoeffq+ncoeffq*2+16], m13
+ pcmpeqw m14, m5 ; m14 = c[i] == 0
+ pcmpeqw m13, m5 ; m13 = c[i] == 0
+ mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+ mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
+ psubw m6, m7 ; m6 = scan[i] + 1
+ psubw m11, m12 ; m11 = scan[i] + 1
+ pandn m14, m6 ; m14 = max(eob)
+ pandn m13, m11 ; m13 = max(eob)
+ pmaxsw m8, m14
+ pmaxsw m8, m13
+ add ncoeffq, mmsize
+ jl .ac_only_loop
+
+%ifidn %1, b_32x32
+ jmp .accumulate_eob
+.skip_iter:
+ mova [qcoeffq+ncoeffq*2+ 0], m5
+ mova [qcoeffq+ncoeffq*2+16], m5
+ mova [dqcoeffq+ncoeffq*2+ 0], m5
+ mova [dqcoeffq+ncoeffq*2+16], m5
+ add ncoeffq, mmsize
+ jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+ ; horizontally accumulate/max eobs and write into [eob] memory pointer
+ mov r2, eobmp
+ pshufd m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0xe
+ pmaxsw m8, m7
+ pshuflw m7, m8, 0x1
+ pmaxsw m8, m7
+ pextrw r6, m8, 0
+ mov [r2], r6
+ RET
+
+ ; skip-block, i.e. just write all zeroes
+.blank:
+ mov r0, dqcoeffmp
+ movifnidn ncoeffq, ncoeffmp
+ mov r2, qcoeffmp
+ mov r3, eobmp
+ DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+ lea dqcoeffq, [dqcoeffq+ncoeffq*2]
+ lea qcoeffq, [ qcoeffq+ncoeffq*2]
+ neg ncoeffq
+ pxor m7, m7
+.blank_loop:
+ mova [dqcoeffq+ncoeffq*2+ 0], m7
+ mova [dqcoeffq+ncoeffq*2+16], m7
+ mova [qcoeffq+ncoeffq*2+ 0], m7
+ mova [qcoeffq+ncoeffq*2+16], m7
+ add ncoeffq, mmsize
+ jl .blank_loop
+ mov word [eobq], 0
+ RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7