summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--examples/vp9_spatial_scalable_encoder.c34
-rw-r--r--vp9/common/vp9_loopfilter.c77
-rw-r--r--vp9/common/vp9_rtcd_defs.pl4
-rw-r--r--vp9/decoder/vp9_decoder.h2
-rw-r--r--vp9/encoder/vp9_aq_cyclicrefresh.c4
-rw-r--r--vp9/encoder/vp9_onyx_int.h33
-rw-r--r--vp9/encoder/x86/vp9_sad4d_intrin_avx2.c167
-rw-r--r--vp9/vp9_dx_iface.c1
-rw-r--r--vp9/vp9cx.mk1
-rw-r--r--vpx/src/svc_encodeframe.c23
10 files changed, 275 insertions, 71 deletions
diff --git a/examples/vp9_spatial_scalable_encoder.c b/examples/vp9_spatial_scalable_encoder.c
index 5c80d34ea..64e62ef29 100644
--- a/examples/vp9_spatial_scalable_encoder.c
+++ b/examples/vp9_spatial_scalable_encoder.c
@@ -67,13 +67,22 @@ static const arg_def_t pass_arg =
ARG_DEF(NULL, "pass", 1, "Pass to execute (1/2)");
static const arg_def_t fpf_name_arg =
ARG_DEF(NULL, "fpf", 1, "First pass statistics file name");
+static const arg_def_t min_q_arg =
+ ARG_DEF(NULL, "min-q", 1, "Minimum quantizer");
+static const arg_def_t max_q_arg =
+ ARG_DEF(NULL, "max-q", 1, "Maximum quantizer");
+static const arg_def_t min_bitrate_arg =
+ ARG_DEF(NULL, "min-bitrate", 1, "Minimum bitrate");
+static const arg_def_t max_bitrate_arg =
+ ARG_DEF(NULL, "max-bitrate", 1, "Maximum bitrate");
static const arg_def_t *svc_args[] = {
&encoding_mode_arg, &frames_arg, &width_arg, &height_arg,
&timebase_arg, &bitrate_arg, &skip_frames_arg, &layers_arg,
&kf_dist_arg, &scale_factors_arg, &quantizers_arg,
- &quantizers_keyframe_arg, &passes_arg, &pass_arg,
- &fpf_name_arg, NULL
+ &quantizers_keyframe_arg, &passes_arg, &pass_arg,
+ &fpf_name_arg, &min_q_arg, &max_q_arg, &min_bitrate_arg,
+ &max_bitrate_arg, NULL
};
static const SVC_ENCODING_MODE default_encoding_mode =
@@ -120,6 +129,8 @@ static void parse_command_line(int argc, const char **argv_,
int passes = 0;
int pass = 0;
const char *fpf_file_name = NULL;
+ unsigned int min_bitrate = 0;
+ unsigned int max_bitrate = 0;
// initialize SvcContext with parameters that will be passed to vpx_svc_init
svc_ctx->log_level = SVC_LOG_DEBUG;
@@ -186,6 +197,14 @@ static void parse_command_line(int argc, const char **argv_,
}
} else if (arg_match(&arg, &fpf_name_arg, argi)) {
fpf_file_name = arg.val;
+ } else if (arg_match(&arg, &min_q_arg, argi)) {
+ enc_cfg->rc_min_quantizer = arg_parse_uint(&arg);
+ } else if (arg_match(&arg, &max_q_arg, argi)) {
+ enc_cfg->rc_max_quantizer = arg_parse_uint(&arg);
+ } else if (arg_match(&arg, &min_bitrate_arg, argi)) {
+ min_bitrate = arg_parse_uint(&arg);
+ } else if (arg_match(&arg, &max_bitrate_arg, argi)) {
+ max_bitrate = arg_parse_uint(&arg);
} else {
++argj;
}
@@ -221,6 +240,17 @@ static void parse_command_line(int argc, const char **argv_,
app_input->pass = pass;
}
+ if (enc_cfg->rc_target_bitrate > 0) {
+ if (min_bitrate > 0) {
+ enc_cfg->rc_2pass_vbr_minsection_pct =
+ min_bitrate * 100 / enc_cfg->rc_target_bitrate;
+ }
+ if (max_bitrate > 0) {
+ enc_cfg->rc_2pass_vbr_maxsection_pct =
+ max_bitrate * 100 / enc_cfg->rc_target_bitrate;
+ }
+ }
+
// Check for unrecognized options
for (argi = argv; *argi; ++argi)
if (argi[0][0] == '-' && strlen(argi[0]) > 1)
diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index af8afed84..e48d4178b 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -228,6 +228,12 @@ static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
}
}
+static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
+ const MB_MODE_INFO *mbmi) {
+ return lfi_n->lvl[mbmi->segment_id][mbmi->ref_frame[0]]
+ [mode_lf_lut[mbmi->mode]];
+}
+
void vp9_loop_filter_init(VP9_COMMON *cm) {
loop_filter_info_n *lfi = &cm->lf_info;
struct loopfilter *lf = &cm->lf;
@@ -493,27 +499,25 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
const MODE_INFO *mi, const int shift_y,
const int shift_uv,
LOOP_FILTER_MASK *lfm) {
- const BLOCK_SIZE block_size = mi->mbmi.sb_type;
- const TX_SIZE tx_size_y = mi->mbmi.tx_size;
- const TX_SIZE tx_size_uv = get_uv_tx_size(&mi->mbmi);
- const int skip = mi->mbmi.skip;
- const int seg = mi->mbmi.segment_id;
- const int ref = mi->mbmi.ref_frame[0];
- const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
- uint64_t *left_y = &lfm->left_y[tx_size_y];
- uint64_t *above_y = &lfm->above_y[tx_size_y];
- uint64_t *int_4x4_y = &lfm->int_4x4_y;
- uint16_t *left_uv = &lfm->left_uv[tx_size_uv];
- uint16_t *above_uv = &lfm->above_uv[tx_size_uv];
- uint16_t *int_4x4_uv = &lfm->int_4x4_uv;
+ const MB_MODE_INFO *mbmi = &mi->mbmi;
+ const BLOCK_SIZE block_size = mbmi->sb_type;
+ const TX_SIZE tx_size_y = mbmi->tx_size;
+ const TX_SIZE tx_size_uv = get_uv_tx_size(mbmi);
+ const int filter_level = get_filter_level(lfi_n, mbmi);
+ uint64_t *const left_y = &lfm->left_y[tx_size_y];
+ uint64_t *const above_y = &lfm->above_y[tx_size_y];
+ uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+ uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
+ uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
+ uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
int i;
- int w = num_8x8_blocks_wide_lookup[block_size];
- int h = num_8x8_blocks_high_lookup[block_size];
// If filter level is 0 we don't loop filter.
if (!filter_level) {
return;
} else {
+ const int w = num_8x8_blocks_wide_lookup[block_size];
+ const int h = num_8x8_blocks_high_lookup[block_size];
int index = shift_y;
for (i = 0; i < h; i++) {
vpx_memset(&lfm->lfl_y[index], filter_level, w);
@@ -540,7 +544,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
// If the block has no coefficients and is not intra we skip applying
// the loop filter on block edges.
- if (skip && ref > INTRA_FRAME)
+ if (mbmi->skip && is_inter_block(mbmi))
return;
// Here we are adding a mask for the transform size. The transform
@@ -561,12 +565,11 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
// boundaries. These differ from the 4x4 boundaries on the outside edge of
// an 8x8 in that the internal ones can be skipped and don't depend on
// the prediction block size.
- if (tx_size_y == TX_4X4) {
+ if (tx_size_y == TX_4X4)
*int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y;
- }
- if (tx_size_uv == TX_4X4) {
+
+ if (tx_size_uv == TX_4X4)
*int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
- }
}
// This function does the same thing as the one above with the exception that
@@ -575,22 +578,20 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
static void build_y_mask(const loop_filter_info_n *const lfi_n,
const MODE_INFO *mi, const int shift_y,
LOOP_FILTER_MASK *lfm) {
- const BLOCK_SIZE block_size = mi->mbmi.sb_type;
- const TX_SIZE tx_size_y = mi->mbmi.tx_size;
- const int skip = mi->mbmi.skip;
- const int seg = mi->mbmi.segment_id;
- const int ref = mi->mbmi.ref_frame[0];
- const int filter_level = lfi_n->lvl[seg][ref][mode_lf_lut[mi->mbmi.mode]];
- uint64_t *left_y = &lfm->left_y[tx_size_y];
- uint64_t *above_y = &lfm->above_y[tx_size_y];
- uint64_t *int_4x4_y = &lfm->int_4x4_y;
+ const MB_MODE_INFO *mbmi = &mi->mbmi;
+ const BLOCK_SIZE block_size = mbmi->sb_type;
+ const TX_SIZE tx_size_y = mbmi->tx_size;
+ const int filter_level = get_filter_level(lfi_n, mbmi);
+ uint64_t *const left_y = &lfm->left_y[tx_size_y];
+ uint64_t *const above_y = &lfm->above_y[tx_size_y];
+ uint64_t *const int_4x4_y = &lfm->int_4x4_y;
int i;
- int w = num_8x8_blocks_wide_lookup[block_size];
- int h = num_8x8_blocks_high_lookup[block_size];
if (!filter_level) {
return;
} else {
+ const int w = num_8x8_blocks_wide_lookup[block_size];
+ const int h = num_8x8_blocks_high_lookup[block_size];
int index = shift_y;
for (i = 0; i < h; i++) {
vpx_memset(&lfm->lfl_y[index], filter_level, w);
@@ -601,7 +602,7 @@ static void build_y_mask(const loop_filter_info_n *const lfi_n,
*above_y |= above_prediction_mask[block_size] << shift_y;
*left_y |= left_prediction_mask[block_size] << shift_y;
- if (skip && ref > INTRA_FRAME)
+ if (mbmi->skip && is_inter_block(mbmi))
return;
*above_y |= (size_mask[block_size] &
@@ -610,9 +611,8 @@ static void build_y_mask(const loop_filter_info_n *const lfi_n,
*left_y |= (size_mask[block_size] &
left_64x64_txform_mask[tx_size_y]) << shift_y;
- if (tx_size_y == TX_4X4) {
+ if (tx_size_y == TX_4X4)
*int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffff) << shift_y;
- }
}
// This function sets up the bit masks for the entire 64x64 region represented
@@ -868,13 +868,6 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
}
-static uint8_t build_lfi(const loop_filter_info_n *lfi_n,
- const MB_MODE_INFO *mbmi) {
- const int seg = mbmi->segment_id;
- const int ref = mbmi->ref_frame[0];
- return lfi_n->lvl[seg][ref][mode_lf_lut[mbmi->mode]];
-}
-
static void filter_selectively_vert(uint8_t *s, int pitch,
unsigned int mask_16x16,
unsigned int mask_8x8,
@@ -953,7 +946,7 @@ static void filter_block_plane_non420(VP9_COMMON *cm,
// Filter level can vary per MI
if (!(lfl[(r << 3) + (c >> ss_x)] =
- build_lfi(&cm->lf_info, &mi[0].mbmi)))
+ get_filter_level(&cm->lf_info, &mi[0].mbmi)))
continue;
// Build masks based on the transform size of each block
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index e4cd9d4a0..b874ef3ba 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -636,7 +636,7 @@ add_proto qw/void vp9_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const
specialize qw/vp9_sad4x4x8 sse4/;
add_proto qw/void vp9_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad64x64x4d sse2/;
+specialize qw/vp9_sad64x64x4d sse2 avx2/;
add_proto qw/void vp9_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad32x64x4d sse2/;
@@ -651,7 +651,7 @@ add_proto qw/void vp9_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, co
specialize qw/vp9_sad16x32x4d sse2/;
add_proto qw/void vp9_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad32x32x4d sse2/;
+specialize qw/vp9_sad32x32x4d sse2 avx2/;
add_proto qw/void vp9_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array";
specialize qw/vp9_sad16x16x4d sse2/;
diff --git a/vp9/decoder/vp9_decoder.h b/vp9/decoder/vp9_decoder.h
index a9458c4d4..e99b21af6 100644
--- a/vp9/decoder/vp9_decoder.h
+++ b/vp9/decoder/vp9_decoder.h
@@ -31,10 +31,8 @@ typedef struct {
int width;
int height;
int version;
- int postprocess;
int max_threads;
int inv_tile_order;
- int input_partition;
} VP9D_CONFIG;
typedef struct VP9Decompressor {
diff --git a/vp9/encoder/vp9_aq_cyclicrefresh.c b/vp9/encoder/vp9_aq_cyclicrefresh.c
index abf770327..5c80f8a8e 100644
--- a/vp9/encoder/vp9_aq_cyclicrefresh.c
+++ b/vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -150,7 +150,7 @@ void vp9_setup_cyclic_refresh_aq(VP9_COMP *const cpi) {
int qindex_delta = 0;
int mbs_in_frame = cm->mi_rows * cm->mi_cols;
int i, x, y, block_count, bl_index, bl_index2;
- int sum_map, new_value, mi_row, mi_col, xmis, ymis, qindex2;
+ int sum_map, mi_row, mi_col, xmis, ymis, qindex2;
// Rate target ratio to set q delta.
float rate_ratio_qdelta = 2.0;
@@ -249,9 +249,9 @@ void vp9_setup_cyclic_refresh_aq(VP9_COMP *const cpi) {
bl_index2 = bl_index + y * cm->mi_cols + x;
sum_map += seg_map[bl_index2];
}
- new_value = 0;
// If segment is partial over superblock, reset.
if (sum_map > 0 && sum_map < xmis * ymis) {
+ int new_value;
if (sum_map < xmis * ymis / 2)
new_value = 0;
else
diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h
index 4bfaacdff..a2147d0b4 100644
--- a/vp9/encoder/vp9_onyx_int.h
+++ b/vp9/encoder/vp9_onyx_int.h
@@ -874,8 +874,8 @@ void vp9_set_svc(VP9_COMP *cpi, int use_svc);
int vp9_get_quantizer(struct VP9_COMP *cpi);
-static int get_ref_frame_idx(const VP9_COMP *cpi,
- MV_REFERENCE_FRAME ref_frame) {
+static INLINE int get_ref_frame_idx(const VP9_COMP *cpi,
+ MV_REFERENCE_FRAME ref_frame) {
if (ref_frame == LAST_FRAME) {
return cpi->lst_fb_idx;
} else if (ref_frame == GOLDEN_FRAME) {
@@ -885,15 +885,25 @@ static int get_ref_frame_idx(const VP9_COMP *cpi,
}
}
-static YV12_BUFFER_CONFIG *get_ref_frame_buffer(VP9_COMP *cpi,
- MV_REFERENCE_FRAME ref_frame) {
- VP9_COMMON *const cm = &cpi->common;
- return &cm->frame_bufs[cm->ref_frame_map[get_ref_frame_idx(cpi,
- ref_frame)]].buf;
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
+ VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+ VP9_COMMON * const cm = &cpi->common;
+ return &cm->frame_bufs[cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)]]
+ .buf;
}
void vp9_set_speed_features(VP9_COMP *cpi);
+static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
+ // TODO(JBB): make this work for alpha channel and double check we can't
+ // exceed this token count if we have a 32x32 transform crossing a boundary
+ // at a multiple of 16.
+ // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
+ // resolution. We assume up to 1 token per pixel, and then allow
+ // a head room of 4.
+ return mb_rows * mb_cols * (16 * 16 * 3 + 4);
+}
+
int vp9_calc_ss_err(const YV12_BUFFER_CONFIG *source,
const YV12_BUFFER_CONFIG *reference);
@@ -908,16 +918,13 @@ void vp9_scale_references(VP9_COMP *cpi);
void vp9_update_reference_frames(VP9_COMP *cpi);
-static int get_token_alloc(int mb_rows, int mb_cols) {
- return mb_rows * mb_cols * (48 * 16 + 4);
-}
-
extern const int q_trans[];
int64_t vp9_rescale(int64_t val, int64_t num, int denom);
-static void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
- MV_REFERENCE_FRAME ref0, MV_REFERENCE_FRAME ref1) {
+static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
+ MV_REFERENCE_FRAME ref0,
+ MV_REFERENCE_FRAME ref1) {
xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
: 0];
xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
diff --git a/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c b/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
new file mode 100644
index 000000000..f31b176e5
--- /dev/null
+++ b/vp9/encoder/x86/vp9_sad4d_intrin_avx2.c
@@ -0,0 +1,167 @@
+/*
+ * 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 <immintrin.h> // AVX2
+#include "vpx/vpx_integer.h"
+
+void vp9_sad32x32x4d_avx2(uint8_t *src,
+ int src_stride,
+ uint8_t *ref[4],
+ int ref_stride,
+ unsigned int res[4]) {
+ __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
+ __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+ __m256i sum_mlow, sum_mhigh;
+ int i;
+ uint8_t *ref0, *ref1, *ref2, *ref3;
+
+ ref0 = ref[0];
+ ref1 = ref[1];
+ ref2 = ref[2];
+ ref3 = ref[3];
+ sum_ref0 = _mm256_set1_epi16(0);
+ sum_ref1 = _mm256_set1_epi16(0);
+ sum_ref2 = _mm256_set1_epi16(0);
+ sum_ref3 = _mm256_set1_epi16(0);
+ for (i = 0; i < 32 ; i++) {
+ // load src and all refs
+ src_reg = _mm256_load_si256((__m256i *)(src));
+ ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
+ ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
+ ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
+ ref3_reg = _mm256_loadu_si256((__m256i *) (ref3));
+ // sum of the absolute differences between every ref-i to src
+ ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+ ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+ ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+ ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+ // sum every ref-i
+ sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+ sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+ sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+ sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+
+ src+= src_stride;
+ ref0+= ref_stride;
+ ref1+= ref_stride;
+ ref2+= ref_stride;
+ ref3+= ref_stride;
+ }
+ {
+ __m128i sum;
+ // in sum_ref-i the result is saved in the first 4 bytes
+ // the other 4 bytes are zeroed.
+ // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+ sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+ sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+ // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+ sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+ sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+ // merge every 64 bit from each sum_ref-i
+ sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+ sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+ // add the low 64 bit to the high 64 bit
+ sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+ // add the low 128 bit to the high 128 bit
+ sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+ _mm256_extractf128_si256(sum_mlow, 1));
+
+ _mm_storeu_si128((__m128i *)(res), sum);
+ }
+}
+
+void vp9_sad64x64x4d_avx2(uint8_t *src,
+ int src_stride,
+ uint8_t *ref[4],
+ int ref_stride,
+ unsigned int res[4]) {
+ __m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg;
+ __m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg;
+ __m256i ref3_reg, ref3next_reg;
+ __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+ __m256i sum_mlow, sum_mhigh;
+ int i;
+ uint8_t *ref0, *ref1, *ref2, *ref3;
+
+ ref0 = ref[0];
+ ref1 = ref[1];
+ ref2 = ref[2];
+ ref3 = ref[3];
+ sum_ref0 = _mm256_set1_epi16(0);
+ sum_ref1 = _mm256_set1_epi16(0);
+ sum_ref2 = _mm256_set1_epi16(0);
+ sum_ref3 = _mm256_set1_epi16(0);
+ for (i = 0; i < 64 ; i++) {
+ // load 64 bytes from src and all refs
+ src_reg = _mm256_load_si256((__m256i *)(src));
+ srcnext_reg = _mm256_load_si256((__m256i *)(src + 32));
+ ref0_reg = _mm256_loadu_si256((__m256i *) (ref0));
+ ref0next_reg = _mm256_loadu_si256((__m256i *) (ref0 + 32));
+ ref1_reg = _mm256_loadu_si256((__m256i *) (ref1));
+ ref1next_reg = _mm256_loadu_si256((__m256i *) (ref1 + 32));
+ ref2_reg = _mm256_loadu_si256((__m256i *) (ref2));
+ ref2next_reg = _mm256_loadu_si256((__m256i *) (ref2 + 32));
+ ref3_reg = _mm256_loadu_si256((__m256i *) (ref3));
+ ref3next_reg = _mm256_loadu_si256((__m256i *) (ref3 + 32));
+ // sum of the absolute differences between every ref-i to src
+ ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+ ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+ ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+ ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+ ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg);
+ ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg);
+ ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg);
+ ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg);
+
+ // sum every ref-i
+ sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+ sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+ sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+ sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+ sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg);
+ sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg);
+ sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg);
+ sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg);
+ src+= src_stride;
+ ref0+= ref_stride;
+ ref1+= ref_stride;
+ ref2+= ref_stride;
+ ref3+= ref_stride;
+ }
+ {
+ __m128i sum;
+
+ // in sum_ref-i the result is saved in the first 4 bytes
+ // the other 4 bytes are zeroed.
+ // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+ sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+ sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+ // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+ sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+ sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+ // merge every 64 bit from each sum_ref-i
+ sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+ sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+ // add the low 64 bit to the high 64 bit
+ sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+ // add the low 128 bit to the high 128 bit
+ sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+ _mm256_extractf128_si256(sum_mlow, 1));
+
+ _mm_storeu_si128((__m128i *)(res), sum);
+ }
+}
diff --git a/vp9/vp9_dx_iface.c b/vp9/vp9_dx_iface.c
index 72701d9d1..4c9350edb 100644
--- a/vp9/vp9_dx_iface.c
+++ b/vp9/vp9_dx_iface.c
@@ -280,7 +280,6 @@ static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
oxcf.width = ctx->si.w;
oxcf.height = ctx->si.h;
oxcf.version = 9;
- oxcf.postprocess = 0;
oxcf.max_threads = ctx->cfg.threads;
oxcf.inv_tile_order = ctx->invert_tile_order;
optr = vp9_create_decompressor(&oxcf);
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index a61f737aa..d89571935 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -89,6 +89,7 @@ VP9_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp9_sad_mmx.asm
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_impl_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_impl_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad4d_sse2.asm
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_sad4d_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subpel_variance_impl_sse2.asm
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
diff --git a/vpx/src/svc_encodeframe.c b/vpx/src/svc_encodeframe.c
index d48a761ff..d4f4e9f38 100644
--- a/vpx/src/svc_encodeframe.c
+++ b/vpx/src/svc_encodeframe.c
@@ -583,8 +583,12 @@ vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
enc_cfg->rc_dropframe_thresh = 0;
enc_cfg->rc_end_usage = VPX_CBR;
enc_cfg->rc_resize_allowed = 0;
- enc_cfg->rc_min_quantizer = 33;
- enc_cfg->rc_max_quantizer = 33;
+
+ if (enc_cfg->g_pass == VPX_RC_ONE_PASS) {
+ enc_cfg->rc_min_quantizer = 33;
+ enc_cfg->rc_max_quantizer = 33;
+ }
+
enc_cfg->rc_undershoot_pct = 100;
enc_cfg->rc_overshoot_pct = 15;
enc_cfg->rc_buf_initial_sz = 500;
@@ -784,12 +788,17 @@ static void set_svc_parameters(SvcContext *svc_ctx,
}
layer_index = layer + VPX_SS_MAX_LAYERS - si->layers;
- if (vpx_svc_is_keyframe(svc_ctx)) {
- svc_params.min_quantizer = si->quantizer_keyframe[layer_index];
- svc_params.max_quantizer = si->quantizer_keyframe[layer_index];
+ if (codec_ctx->config.enc->g_pass == VPX_RC_ONE_PASS) {
+ if (vpx_svc_is_keyframe(svc_ctx)) {
+ svc_params.min_quantizer = si->quantizer_keyframe[layer_index];
+ svc_params.max_quantizer = si->quantizer_keyframe[layer_index];
+ } else {
+ svc_params.min_quantizer = si->quantizer[layer_index];
+ svc_params.max_quantizer = si->quantizer[layer_index];
+ }
} else {
- svc_params.min_quantizer = si->quantizer[layer_index];
- svc_params.max_quantizer = si->quantizer[layer_index];
+ svc_params.min_quantizer = codec_ctx->config.enc->rc_min_quantizer;
+ svc_params.max_quantizer = codec_ctx->config.enc->rc_max_quantizer;
}
svc_params.distance_from_i_frame = si->frame_within_gop;
generated by cgit v1.2.3 (git 2.25.1) at 2024-10-12 05:29:53 +0000