Merge "Datarate test for temporal layers and frame dropper."
diff --git a/examples/vpx_temporal_scalable_patterns.c b/examples/vpx_temporal_scalable_patterns.c
index c40450e..73278a8 100644
--- a/examples/vpx_temporal_scalable_patterns.c
+++ b/examples/vpx_temporal_scalable_patterns.c
@@ -12,6 +12,7 @@
// encoding scheme based on temporal scalability for video applications
// that benefit from a scalable bitstream.
+#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
@@ -31,6 +32,86 @@
static int mode_to_num_layers[12] = {1, 2, 2, 3, 3, 3, 3, 5, 2, 3, 3, 3};
+// For rate control encoding stats.
+struct RateControlMetrics {
+ // Number of input frames per layer.
+ int layer_input_frames[VPX_TS_MAX_LAYERS];
+ // Total (cumulative) number of encoded frames per layer.
+ int layer_tot_enc_frames[VPX_TS_MAX_LAYERS];
+ // Number of encoded non-key frames per layer.
+ int layer_enc_frames[VPX_TS_MAX_LAYERS];
+ // Framerate per layer layer (cumulative).
+ float layer_framerate[VPX_TS_MAX_LAYERS];
+ // Target average frame size per layer (per-frame-bandwidth per layer).
+ float layer_pfb[VPX_TS_MAX_LAYERS];
+ // Actual average frame size per layer.
+ float layer_avg_frame_size[VPX_TS_MAX_LAYERS];
+ // Average rate mismatch per layer (|target - actual| / target).
+ float layer_avg_rate_mismatch[VPX_TS_MAX_LAYERS];
+ // Actual encoding bitrate per layer (cumulative).
+ float layer_encoding_bitrate[VPX_TS_MAX_LAYERS];
+};
+
+static void set_rate_control_metrics(struct RateControlMetrics *rc,
+ vpx_codec_enc_cfg_t *cfg) {
+ int i = 0;
+ // Set the layer (cumulative) framerate and the target layer (non-cumulative)
+ // per-frame-bandwidth, for the rate control encoding stats below.
+ float framerate = cfg->g_timebase.den / cfg->g_timebase.num;
+ rc->layer_framerate[0] = framerate / cfg->ts_rate_decimator[0];
+ rc->layer_pfb[0] = 1000.0 * cfg->ts_target_bitrate[0] /
+ rc->layer_framerate[0];
+ for (i = 0; i < cfg->ts_number_layers; ++i) {
+ if (i > 0) {
+ rc->layer_framerate[i] = framerate / cfg->ts_rate_decimator[i];
+ rc->layer_pfb[i] = 1000.0 *
+ (cfg->ts_target_bitrate[i] - cfg->ts_target_bitrate[i - 1]) /
+ (rc->layer_framerate[i] - rc->layer_framerate[i - 1]);
+ }
+ rc->layer_input_frames[i] = 0;
+ rc->layer_enc_frames[i] = 0;
+ rc->layer_tot_enc_frames[i] = 0;
+ rc->layer_encoding_bitrate[i] = 0.0;
+ rc->layer_avg_frame_size[i] = 0.0;
+ rc->layer_avg_rate_mismatch[i] = 0.0;
+ }
+}
+
+static void printout_rate_control_summary(struct RateControlMetrics *rc,
+ vpx_codec_enc_cfg_t *cfg,
+ int frame_cnt) {
+ int i = 0;
+ int check_num_frames = 0;
+ printf("Total number of processed frames: %d\n\n", frame_cnt -1);
+ printf("Rate control layer stats for %d layer(s):\n\n",
+ cfg->ts_number_layers);
+ for (i = 0; i < cfg->ts_number_layers; ++i) {
+ const int num_dropped = (i > 0) ?
+ (rc->layer_input_frames[i] - rc->layer_enc_frames[i]) :
+ (rc->layer_input_frames[i] - rc->layer_enc_frames[i] - 1);
+ rc->layer_encoding_bitrate[i] = 0.001 * rc->layer_framerate[i] *
+ rc->layer_encoding_bitrate[i] / rc->layer_tot_enc_frames[i];
+ rc->layer_avg_frame_size[i] = rc->layer_avg_frame_size[i] /
+ rc->layer_enc_frames[i];
+ rc->layer_avg_rate_mismatch[i] = 100.0 * rc->layer_avg_rate_mismatch[i] /
+ rc->layer_enc_frames[i];
+ printf("For layer#: %d \n", i);
+ printf("Bitrate (target vs actual): %d %f \n", cfg->ts_target_bitrate[i],
+ rc->layer_encoding_bitrate[i]);
+ printf("Average frame size (target vs actual): %f %f \n", rc->layer_pfb[i],
+ rc->layer_avg_frame_size[i]);
+ printf("Average rate_mismatch: %f \n", rc->layer_avg_rate_mismatch[i]);
+ printf("Number of input frames, encoded (non-key) frames, "
+ "and perc dropped frames: %d %d %f \n", rc->layer_input_frames[i],
+ rc->layer_enc_frames[i],
+ 100.0 * num_dropped / rc->layer_input_frames[i]);
+ check_num_frames += rc->layer_input_frames[i];
+ printf("\n");
+ }
+ if ((frame_cnt - 1) != check_num_frames)
+ die("Error: Number of input frames not equal to output! \n");
+}
+
// Temporal scaling parameters:
// NOTE: The 3 prediction frames cannot be used interchangeably due to
// differences in the way they are handled throughout the code. The
@@ -355,20 +436,20 @@
int pts = 0; // PTS starts at 0.
int frame_duration = 1; // 1 timebase tick per frame.
int layering_mode = 0;
- int frames_in_layer[VPX_TS_MAX_LAYERS] = {0};
int layer_flags[VPX_TS_MAX_PERIODICITY] = {0};
int flag_periodicity = 1;
int max_intra_size_pct;
vpx_svc_layer_id_t layer_id = {0, 0};
const VpxInterface *encoder = NULL;
struct VpxInputContext input_ctx = {0};
+ struct RateControlMetrics rc;
exec_name = argv[0];
// Check usage and arguments.
- if (argc < 10) {
+ if (argc < 11) {
die("Usage: %s <infile> <outfile> <codec_type(vp8/vp9)> <width> <height> "
- "<rate_num> <rate_den> <mode> <Rate_0> ... <Rate_nlayers-1> \n",
- argv[0]);
+ "<rate_num> <rate_den> <frame_drop_threshold> <mode> "
+ "<Rate_0> ... <Rate_nlayers-1> \n", argv[0]);
}
encoder = get_vpx_encoder_by_name(argv[3]);
@@ -383,12 +464,12 @@
die("Invalid resolution: %d x %d", width, height);
}
- layering_mode = strtol(argv[8], NULL, 0);
- if (layering_mode < 0 || layering_mode > 11) {
- die("Invalid mode (0..11) %s", argv[8]);
+ layering_mode = strtol(argv[9], NULL, 0);
+ if (layering_mode < 0 || layering_mode > 12) {
+ die("Invalid mode (0..12) %s", argv[9]);
}
- if (argc != 9 + mode_to_num_layers[layering_mode]) {
+ if (argc != 10 + mode_to_num_layers[layering_mode]) {
die("Invalid number of arguments");
}
@@ -411,12 +492,12 @@
cfg.g_timebase.num = strtol(argv[6], NULL, 0);
cfg.g_timebase.den = strtol(argv[7], NULL, 0);
- for (i = 9; i < 9 + mode_to_num_layers[layering_mode]; ++i) {
- cfg.ts_target_bitrate[i - 9] = strtol(argv[i], NULL, 0);
+ for (i = 10; i < 10 + mode_to_num_layers[layering_mode]; ++i) {
+ cfg.ts_target_bitrate[i - 10] = strtol(argv[i], NULL, 0);
}
// Real time parameters.
- cfg.rc_dropframe_thresh = 0;
+ cfg.rc_dropframe_thresh = strtol(argv[8], NULL, 0);
cfg.rc_end_usage = VPX_CBR;
cfg.rc_resize_allowed = 0;
cfg.rc_min_quantizer = 2;
@@ -443,6 +524,8 @@
layer_flags,
&flag_periodicity);
+ set_rate_control_metrics(&rc, &cfg);
+
// Open input file.
input_ctx.filename = argv[1];
if (!(input_ctx.file = fopen(input_ctx.filename, "rb"))) {
@@ -494,9 +577,13 @@
layer_id.spatial_layer_id = 0;
layer_id.temporal_layer_id =
cfg.ts_layer_id[frame_cnt % cfg.ts_periodicity];
- vpx_codec_control(&codec, VP9E_SET_SVC_LAYER_ID, &layer_id);
+ if (strncmp(encoder->name, "vp9", 3) == 0) {
+ vpx_codec_control(&codec, VP9E_SET_SVC_LAYER_ID, &layer_id);
+ }
flags = layer_flags[frame_cnt % flag_periodicity];
frame_avail = !read_yuv_frame(&input_ctx, &raw);
+ if (frame_avail)
+ ++rc.layer_input_frames[layer_id.temporal_layer_id];
if (vpx_codec_encode(&codec, frame_avail? &raw : NULL, pts, 1, flags,
VPX_DL_REALTIME)) {
die_codec(&codec, "Failed to encode frame");
@@ -514,7 +601,17 @@
i < cfg.ts_number_layers; ++i) {
vpx_video_writer_write_frame(outfile[i], pkt->data.frame.buf,
pkt->data.frame.sz, pts);
- ++frames_in_layer[i];
+ ++rc.layer_tot_enc_frames[i];
+ rc.layer_encoding_bitrate[i] += 8.0 * pkt->data.frame.sz;
+ // Keep count of rate control stats per layer (for non-key frames).
+ if (i == cfg.ts_layer_id[frame_cnt % cfg.ts_periodicity] &&
+ !(pkt->data.frame.flags & VPX_FRAME_IS_KEY)) {
+ rc.layer_avg_frame_size[i] += 8.0 * pkt->data.frame.sz;
+ rc.layer_avg_rate_mismatch[i] +=
+ fabs(8.0 * pkt->data.frame.sz - rc.layer_pfb[i]) /
+ rc.layer_pfb[i];
+ ++rc.layer_enc_frames[i];
+ }
}
break;
default:
@@ -525,7 +622,8 @@
pts += frame_duration;
}
fclose(input_ctx.file);
- printf("Processed %d frames: \n", frame_cnt - 1);
+ printout_rate_control_summary(&rc, &cfg, frame_cnt);
+
if (vpx_codec_destroy(&codec))
die_codec(&codec, "Failed to destroy codec");
diff --git a/vp9/common/arm/neon/vp9_mb_lpf_neon.asm b/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
index 8cb913c..5fe2bba 100644
--- a/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
+++ b/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
@@ -439,6 +439,9 @@
tst r7, #1
bxne lr
+ orrs r5, r5, r6 ; Check for 0
+ orreq r7, r7, #2 ; Only do mbfilter branch
+
; mbfilter flat && mask branch
; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
; and using vibt on the q's?
diff --git a/vp9/common/vp9_blockd.h b/vp9/common/vp9_blockd.h
index c22c621..6086323 100644
--- a/vp9/common/vp9_blockd.h
+++ b/vp9/common/vp9_blockd.h
@@ -128,7 +128,7 @@
uint8_t mode_context[MAX_REF_FRAMES];
- unsigned char skip_coeff; // 0=need to decode coeffs, 1=no coefficients
+ unsigned char skip; // 0=need to decode coeffs, 1=no coefficients
unsigned char segment_id; // Segment id for this block.
// Flags used for prediction status of various bit-stream signals
diff --git a/vp9/common/vp9_debugmodes.c b/vp9/common/vp9_debugmodes.c
index 355ac1a..24c785f 100644
--- a/vp9/common/vp9_debugmodes.c
+++ b/vp9/common/vp9_debugmodes.c
@@ -58,7 +58,7 @@
print_mi_data(cm, mvs, "Partitions:", offsetof(MB_MODE_INFO, sb_type));
print_mi_data(cm, mvs, "Modes:", offsetof(MB_MODE_INFO, mode));
- print_mi_data(cm, mvs, "Skips:", offsetof(MB_MODE_INFO, skip_coeff));
+ print_mi_data(cm, mvs, "Skips:", offsetof(MB_MODE_INFO, skip));
print_mi_data(cm, mvs, "Ref frame:", offsetof(MB_MODE_INFO, ref_frame[0]));
print_mi_data(cm, mvs, "Transform:", offsetof(MB_MODE_INFO, tx_size));
print_mi_data(cm, mvs, "UV Modes:", offsetof(MB_MODE_INFO, uv_mode));
diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 07d7a92..04f8934 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -496,7 +496,7 @@
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_coeff;
+ 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]];
@@ -577,7 +577,7 @@
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_coeff;
+ 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]];
@@ -937,8 +937,7 @@
for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
const MODE_INFO *mi = mi_8x8[c];
const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
- const int skip_this = mi[0].mbmi.skip_coeff
- && is_inter_block(&mi[0].mbmi);
+ const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
// left edge of current unit is block/partition edge -> no skip
const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
!(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
diff --git a/vp9/common/vp9_postproc.c b/vp9/common/vp9_postproc.c
index a172ba6..7baa9ee 100644
--- a/vp9/common/vp9_postproc.c
+++ b/vp9/common/vp9_postproc.c
@@ -700,7 +700,7 @@
char zz[4];
int dc_diff = !(mi[mb_index].mbmi.mode != I4X4_PRED &&
mi[mb_index].mbmi.mode != SPLITMV &&
- mi[mb_index].mbmi.skip_coeff);
+ mi[mb_index].mbmi.skip);
if (cm->frame_type == KEY_FRAME)
snprintf(zz, sizeof(zz) - 1, "a");
diff --git a/vp9/common/vp9_pred_common.c b/vp9/common/vp9_pred_common.c
index 487f00c..197bcb6 100644
--- a/vp9/common/vp9_pred_common.c
+++ b/vp9/common/vp9_pred_common.c
@@ -353,10 +353,10 @@
const MB_MODE_INFO *const left_mbmi = get_mbmi(get_left_mi(xd));
const int has_above = above_mbmi != NULL;
const int has_left = left_mbmi != NULL;
- int above_ctx = (has_above && !above_mbmi->skip_coeff) ? above_mbmi->tx_size
- : max_tx_size;
- int left_ctx = (has_left && !left_mbmi->skip_coeff) ? left_mbmi->tx_size
- : max_tx_size;
+ int above_ctx = (has_above && !above_mbmi->skip) ? above_mbmi->tx_size
+ : max_tx_size;
+ int left_ctx = (has_left && !left_mbmi->skip) ? left_mbmi->tx_size
+ : max_tx_size;
if (!has_left)
left_ctx = above_ctx;
diff --git a/vp9/common/vp9_pred_common.h b/vp9/common/vp9_pred_common.h
index 33ae5a8..7f9e563 100644
--- a/vp9/common/vp9_pred_common.h
+++ b/vp9/common/vp9_pred_common.h
@@ -47,8 +47,8 @@
static INLINE int vp9_get_skip_context(const MACROBLOCKD *xd) {
const MODE_INFO *const above_mi = get_above_mi(xd);
const MODE_INFO *const left_mi = get_left_mi(xd);
- const int above_skip = (above_mi != NULL) ? above_mi->mbmi.skip_coeff : 0;
- const int left_skip = (left_mi != NULL) ? left_mi->mbmi.skip_coeff : 0;
+ const int above_skip = (above_mi != NULL) ? above_mi->mbmi.skip : 0;
+ const int left_skip = (left_mi != NULL) ? left_mi->mbmi.skip : 0;
return above_skip + left_skip;
}
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 878c751..6317103 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -264,13 +264,13 @@
specialize vp9_convolve_avg $sse2_x86inc neon dspr2
prototype void vp9_convolve8 "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
-specialize vp9_convolve8 sse2 ssse3 neon dspr2
+specialize vp9_convolve8 sse2 ssse3 avx2 neon dspr2
prototype void vp9_convolve8_horiz "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
-specialize vp9_convolve8_horiz sse2 ssse3 neon dspr2
+specialize vp9_convolve8_horiz sse2 ssse3 avx2 neon dspr2
prototype void vp9_convolve8_vert "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
-specialize vp9_convolve8_vert sse2 ssse3 neon dspr2
+specialize vp9_convolve8_vert sse2 ssse3 avx2 neon dspr2
prototype void vp9_convolve8_avg "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"
specialize vp9_convolve8_avg sse2 ssse3 neon dspr2
diff --git a/vp9/common/x86/vp9_asm_stubs.c b/vp9/common/x86/vp9_asm_stubs.c
index 60018ea..a2cf910 100644
--- a/vp9/common/x86/vp9_asm_stubs.c
+++ b/vp9/common/x86/vp9_asm_stubs.c
@@ -139,7 +139,49 @@
filter_x, x_step_q4, filter_y, y_step_q4, w, h); \
} \
}
+#if HAVE_AVX2
+filter8_1dfunction vp9_filter_block1d16_v8_avx2;
+filter8_1dfunction vp9_filter_block1d16_h8_avx2;
+filter8_1dfunction vp9_filter_block1d8_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_ssse3;
+filter8_1dfunction vp9_filter_block1d16_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h2_ssse3;
+#define vp9_filter_block1d8_v8_avx2 vp9_filter_block1d8_v8_ssse3
+#define vp9_filter_block1d8_h8_avx2 vp9_filter_block1d8_h8_ssse3
+#define vp9_filter_block1d4_v8_avx2 vp9_filter_block1d4_v8_ssse3
+#define vp9_filter_block1d4_h8_avx2 vp9_filter_block1d4_h8_ssse3
+#define vp9_filter_block1d16_v2_avx2 vp9_filter_block1d16_v2_ssse3
+#define vp9_filter_block1d16_h2_avx2 vp9_filter_block1d16_h2_ssse3
+#define vp9_filter_block1d8_v2_avx2 vp9_filter_block1d8_v2_ssse3
+#define vp9_filter_block1d8_h2_avx2 vp9_filter_block1d8_h2_ssse3
+#define vp9_filter_block1d4_v2_avx2 vp9_filter_block1d4_v2_ssse3
+#define vp9_filter_block1d4_h2_avx2 vp9_filter_block1d4_h2_ssse3
+// void vp9_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h);
+// void vp9_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
+// void vp9_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h);
+FUN_CONV_2D(, avx2);
+#endif
#if HAVE_SSSE3
filter8_1dfunction vp9_filter_block1d16_v8_ssse3;
filter8_1dfunction vp9_filter_block1d16_h8_ssse3;
diff --git a/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c b/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
new file mode 100644
index 0000000..0ffb1bc
--- /dev/null
+++ b/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
@@ -0,0 +1,542 @@
+/*
+ * Copyright (c) 2010 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>
+#include "vpx_ports/mem.h"
+
+// filters for 16_h8 and 16_v8
+DECLARE_ALIGNED(32, const unsigned char, filt1_global_avx2[32])= {
+ 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+ 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8};
+
+DECLARE_ALIGNED(32, const unsigned char, filt2_global_avx2[32])= {
+ 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
+ 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10};
+
+DECLARE_ALIGNED(32, const unsigned char, filt3_global_avx2[32])= {
+ 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
+ 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12};
+
+DECLARE_ALIGNED(32, const unsigned char, filt4_global_avx2[32])= {
+ 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
+ 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14};
+
+
+void vp9_filter_block1d16_h8_avx2(unsigned char *src_ptr,
+ unsigned int src_pixels_per_line,
+ unsigned char *output_ptr,
+ unsigned int output_pitch,
+ unsigned int output_height,
+ int16_t *filter) {
+ __m128i filtersReg;
+ __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+ __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+ __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+ __m256i srcReg32b1, srcReg32b2, filtersReg32;
+ unsigned int i;
+ unsigned int src_stride, dst_stride;
+
+ // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+ addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+ filtersReg = _mm_loadu_si128((__m128i *)filter);
+ // converting the 16 bit (short) to 8 bit (byte) and have the same data
+ // in both lanes of 128 bit register.
+ filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+ // have the same data in both lanes of a 256 bit register
+#if defined (__GNUC__)
+#if ( __GNUC__ < 4 || (__GNUC__ == 4 && \
+(__GNUC_MINOR__ < 6 || (__GNUC_MINOR__ == 6 && __GNUC_PATCHLEVEL__ > 0))))
+ filtersReg32 = _mm_broadcastsi128_si256((__m128i const *)&filtersReg);
+#elif(__GNUC__ == 4 && (__GNUC_MINOR__ == 7 && __GNUC_PATCHLEVEL__ > 0))
+ filtersReg32 = _mm_broadcastsi128_si256(filtersReg);
+#else
+ filtersReg32 = _mm256_broadcastsi128_si256(filtersReg);
+#endif
+#else
+ filtersReg32 = _mm256_broadcastsi128_si256(filtersReg);
+#endif
+
+ // duplicate only the first 16 bits (first and second byte)
+ // across 256 bit register
+ firstFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x100u));
+ // duplicate only the second 16 bits (third and forth byte)
+ // across 256 bit register
+ secondFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x302u));
+ // duplicate only the third 16 bits (fifth and sixth byte)
+ // across 256 bit register
+ thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x504u));
+ // duplicate only the forth 16 bits (seventh and eighth byte)
+ // across 256 bit register
+ forthFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x706u));
+
+ filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+ filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+ filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+ filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+
+ // multiple the size of the source and destination stride by two
+ src_stride = src_pixels_per_line << 1;
+ dst_stride = output_pitch << 1;
+ for (i = output_height; i > 1; i-=2) {
+ // load the 2 strides of source
+ srcReg32b1 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr-3)));
+ srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+ _mm_loadu_si128((__m128i *)
+ (src_ptr+src_pixels_per_line-3)), 1);
+
+ // filter the source buffer
+ srcRegFilt32b1_1= _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+ srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+ srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters);
+
+ // add and saturate the results together
+ srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+ // filter the source buffer
+ srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+ srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, forthFilters);
+ srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+ // add and saturate the results together
+ srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+ _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+ // reading 2 strides of the next 16 bytes
+ // (part of it was being read by earlier read)
+ srcReg32b2 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+5)));
+ srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+ _mm_loadu_si128((__m128i *)
+ (src_ptr+src_pixels_per_line+5)), 1);
+
+ // add and saturate the results together
+ srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+ _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+ // filter the source buffer
+ srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
+ srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
+ srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters);
+
+ // add and saturate the results together
+ srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
+
+ // filter the source buffer
+ srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+ srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, forthFilters);
+ srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+ // add and saturate the results together
+ srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+ _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+ srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+ _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+
+ srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
+
+ srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
+
+ // shift by 7 bit each 16 bit
+ srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
+ srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
+
+ // shrink to 8 bit each 16 bits, the first lane contain the first
+ // convolve result and the second lane contain the second convolve
+ // result
+ srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1,
+ srcRegFilt32b2_1);
+
+ src_ptr+=src_stride;
+
+ // save 16 bytes
+ _mm_store_si128((__m128i*)output_ptr,
+ _mm256_castsi256_si128(srcRegFilt32b1_1));
+
+ // save the next 16 bits
+ _mm_store_si128((__m128i*)(output_ptr+output_pitch),
+ _mm256_extractf128_si256(srcRegFilt32b1_1, 1));
+ output_ptr+=dst_stride;
+ }
+
+ // if the number of strides is odd.
+ // process only 16 bytes
+ if (i > 0) {
+ __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
+ __m128i srcRegFilt2, srcRegFilt3;
+
+ srcReg1 = _mm_loadu_si128((__m128i *)(src_ptr-3));
+
+ // filter the source buffer
+ srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1,
+ _mm256_castsi256_si128(filt1Reg));
+ srcRegFilt2 = _mm_shuffle_epi8(srcReg1,
+ _mm256_castsi256_si128(filt2Reg));
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1,
+ _mm256_castsi256_si128(firstFilters));
+ srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+ _mm256_castsi256_si128(secondFilters));
+
+ // add and saturate the results together
+ srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+ // filter the source buffer
+ srcRegFilt3= _mm_shuffle_epi8(srcReg1,
+ _mm256_castsi256_si128(filt4Reg));
+ srcRegFilt2= _mm_shuffle_epi8(srcReg1,
+ _mm256_castsi256_si128(filt3Reg));
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+ _mm256_castsi256_si128(forthFilters));
+ srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+ _mm256_castsi256_si128(thirdFilters));
+
+ // add and saturate the results together
+ srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+ _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+
+ // reading the next 16 bytes
+ // (part of it was being read by earlier read)
+ srcReg2 = _mm_loadu_si128((__m128i *)(src_ptr+5));
+
+ // add and saturate the results together
+ srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+ _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+ // filter the source buffer
+ srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2,
+ _mm256_castsi256_si128(filt1Reg));
+ srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+ _mm256_castsi256_si128(filt2Reg));
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1,
+ _mm256_castsi256_si128(firstFilters));
+ srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+ _mm256_castsi256_si128(secondFilters));
+
+ // add and saturate the results together
+ srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+ // filter the source buffer
+ srcRegFilt3 = _mm_shuffle_epi8(srcReg2,
+ _mm256_castsi256_si128(filt4Reg));
+ srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+ _mm256_castsi256_si128(filt3Reg));
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+ _mm256_castsi256_si128(forthFilters));
+ srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+ _mm256_castsi256_si128(thirdFilters));
+
+ // add and saturate the results together
+ srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+ _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+ srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+ _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+
+ srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+ _mm256_castsi256_si128(addFilterReg64));
+
+ srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+ _mm256_castsi256_si128(addFilterReg64));
+
+ // shift by 7 bit each 16 bit
+ srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
+ srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+
+ // shrink to 8 bit each 16 bits, the first lane contain the first
+ // convolve result and the second lane contain the second convolve
+ // result
+ srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+ // save 16 bytes
+ _mm_store_si128((__m128i*)output_ptr, srcRegFilt1_1);
+ }
+}
+
+void vp9_filter_block1d16_v8_avx2(unsigned char *src_ptr,
+ unsigned int src_pitch,
+ unsigned char *output_ptr,
+ unsigned int out_pitch,
+ unsigned int output_height,
+ int16_t *filter) {
+ __m128i filtersReg;
+ __m256i addFilterReg64;
+ __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
+ __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
+ __m256i srcReg32b11, srcReg32b12, srcReg32b13, filtersReg32;
+ __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+ unsigned int i;
+ unsigned int src_stride, dst_stride;
+
+ // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+ addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+ filtersReg = _mm_loadu_si128((__m128i *)filter);
+ // converting the 16 bit (short) to 8 bit (byte) and have the
+ // same data in both lanes of 128 bit register.
+ filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+ // have the same data in both lanes of a 256 bit register
+#if defined (__GNUC__)
+#if ( __GNUC__ < 4 || (__GNUC__ == 4 && \
+(__GNUC_MINOR__ < 6 || (__GNUC_MINOR__ == 6 && __GNUC_PATCHLEVEL__ > 0))))
+ filtersReg32 = _mm_broadcastsi128_si256((__m128i const *)&filtersReg);
+#elif(__GNUC__ == 4 && (__GNUC_MINOR__ == 7 && __GNUC_PATCHLEVEL__ > 0))
+ filtersReg32 = _mm_broadcastsi128_si256(filtersReg);
+#else
+ filtersReg32 = _mm256_broadcastsi128_si256(filtersReg);
+#endif
+#else
+ filtersReg32 = _mm256_broadcastsi128_si256(filtersReg);
+#endif
+
+ // duplicate only the first 16 bits (first and second byte)
+ // across 256 bit register
+ firstFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x100u));
+ // duplicate only the second 16 bits (third and forth byte)
+ // across 256 bit register
+ secondFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x302u));
+ // duplicate only the third 16 bits (fifth and sixth byte)
+ // across 256 bit register
+ thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x504u));
+ // duplicate only the forth 16 bits (seventh and eighth byte)
+ // across 256 bit register
+ forthFilters = _mm256_shuffle_epi8(filtersReg32,
+ _mm256_set1_epi16(0x706u));
+
+ // multiple the size of the source and destination stride by two
+ src_stride = src_pitch << 1;
+ dst_stride = out_pitch << 1;
+
+ // load 16 bytes 7 times in stride of src_pitch
+ srcReg32b1 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr)));
+ srcReg32b2 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch)));
+ srcReg32b3 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*2)));
+ srcReg32b4 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*3)));
+ srcReg32b5 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*4)));
+ srcReg32b6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*5)));
+ srcReg32b7 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*6)));
+
+ // have each consecutive loads on the same 256 register
+ srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+ _mm256_castsi256_si128(srcReg32b2), 1);
+ srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+ _mm256_castsi256_si128(srcReg32b3), 1);
+ srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
+ _mm256_castsi256_si128(srcReg32b4), 1);
+ srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
+ _mm256_castsi256_si128(srcReg32b5), 1);
+ srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
+ _mm256_castsi256_si128(srcReg32b6), 1);
+ srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
+ _mm256_castsi256_si128(srcReg32b7), 1);
+
+ // merge every two consecutive registers except the last one
+ srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
+ srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
+
+ // save
+ srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
+
+ // save
+ srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
+
+ // save
+ srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+
+ // save
+ srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+
+
+ for (i = output_height; i > 1; i-=2) {
+ // load the last 2 loads of 16 bytes and have every two
+ // consecutive loads in the same 256 bit register
+ srcReg32b8 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*7)));
+ srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+ _mm256_castsi256_si128(srcReg32b8), 1);
+ srcReg32b9 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*8)));
+ srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
+ _mm256_castsi256_si128(srcReg32b9), 1);
+
+ // merge every two consecutive registers
+ // save
+ srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
+ srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
+ srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
+ srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
+ srcReg32b8 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
+
+ // add and saturate the results together
+ srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
+ srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b8);
+
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
+ srcReg32b6 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
+ srcReg32b13 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
+
+
+ // add and saturate the results together
+ srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+ _mm256_min_epi16(srcReg32b8, srcReg32b12));
+ srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+ _mm256_min_epi16(srcReg32b6, srcReg32b13));
+
+ // add and saturate the results together
+ srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+ _mm256_max_epi16(srcReg32b8, srcReg32b12));
+ srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+ _mm256_max_epi16(srcReg32b6, srcReg32b13));
+
+
+ srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
+ srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
+
+ // shift by 7 bit each 16 bit
+ srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
+ srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
+
+ // shrink to 8 bit each 16 bits, the first lane contain the first
+ // convolve result and the second lane contain the second convolve
+ // result
+ srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+
+ src_ptr+=src_stride;
+
+ // save 16 bytes
+ _mm_store_si128((__m128i*)output_ptr,
+ _mm256_castsi256_si128(srcReg32b1));
+
+ // save the next 16 bits
+ _mm_store_si128((__m128i*)(output_ptr+out_pitch),
+ _mm256_extractf128_si256(srcReg32b1, 1));
+
+ output_ptr+=dst_stride;
+
+ // save part of the registers for next strides
+ srcReg32b10 = srcReg32b11;
+ srcReg32b1 = srcReg32b3;
+ srcReg32b11 = srcReg32b2;
+ srcReg32b3 = srcReg32b5;
+ srcReg32b2 = srcReg32b4;
+ srcReg32b5 = srcReg32b7;
+ srcReg32b7 = srcReg32b9;
+ }
+ if (i > 0) {
+ __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
+ __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
+ // load the last 16 bytes
+ srcRegFilt8 = _mm_loadu_si128((__m128i *)(src_ptr+src_pitch*7));
+
+ // merge the last 2 results together
+ srcRegFilt4 = _mm_unpacklo_epi8(
+ _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+ srcRegFilt7 = _mm_unpackhi_epi8(
+ _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
+ _mm256_castsi256_si128(firstFilters));
+ srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4,
+ _mm256_castsi256_si128(forthFilters));
+ srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
+ _mm256_castsi256_si128(firstFilters));
+ srcRegFilt7 = _mm_maddubs_epi16(srcRegFilt7,
+ _mm256_castsi256_si128(forthFilters));
+
+ // add and saturate the results together
+ srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+ srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
+
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+ _mm256_castsi256_si128(secondFilters));
+ srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
+ _mm256_castsi256_si128(secondFilters));
+
+ // multiply 2 adjacent elements with the filter and add the result
+ srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
+ _mm256_castsi256_si128(thirdFilters));
+ srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
+ _mm256_castsi256_si128(thirdFilters));
+
+ // add and saturate the results together
+ srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+ _mm_min_epi16(srcRegFilt4, srcRegFilt6));
+ srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+ _mm_min_epi16(srcRegFilt5, srcRegFilt7));
+
+ // add and saturate the results together
+ srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+ _mm_max_epi16(srcRegFilt4, srcRegFilt6));
+ srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+ _mm_max_epi16(srcRegFilt5, srcRegFilt7));
+
+
+ srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+ _mm256_castsi256_si128(addFilterReg64));
+ srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+ _mm256_castsi256_si128(addFilterReg64));
+
+ // shift by 7 bit each 16 bit
+ srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+ srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
+
+ // shrink to 8 bit each 16 bits, the first lane contain the first
+ // convolve result and the second lane contain the second convolve
+ // result
+ srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+
+ // save 16 bytes
+ _mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
+ }
+}
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index d37afa5..b1900e6 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -303,7 +303,7 @@
dst, pd->dst.stride, dst, pd->dst.stride,
x, y, plane);
- if (!mi->mbmi.skip_coeff) {
+ if (!mi->mbmi.skip) {
const int eob = vp9_decode_block_tokens(cm, xd, plane, block,
plane_bsize, x, y, tx_size,
args->r);
@@ -397,7 +397,7 @@
// Has to be called after set_offsets
mbmi = &xd->mi_8x8[0]->mbmi;
- if (mbmi->skip_coeff) {
+ if (mbmi->skip) {
reset_skip_context(xd, bsize);
} else {
if (cm->seg.enabled)
@@ -421,12 +421,12 @@
vp9_dec_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
// Reconstruction
- if (!mbmi->skip_coeff) {
+ if (!mbmi->skip) {
int eobtotal = 0;
struct inter_args arg = { cm, xd, r, &eobtotal };
vp9_foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
if (!less8x8 && eobtotal == 0)
- mbmi->skip_coeff = 1; // skip loopfilter
+ mbmi->skip = 1; // skip loopfilter
}
}
diff --git a/vp9/decoder/vp9_decodemv.c b/vp9/decoder/vp9_decodemv.c
index c7fb71d..856c8b5 100644
--- a/vp9/decoder/vp9_decodemv.c
+++ b/vp9/decoder/vp9_decodemv.c
@@ -146,8 +146,8 @@
return segment_id;
}
-static int read_skip_coeff(VP9_COMMON *cm, const MACROBLOCKD *xd,
- int segment_id, vp9_reader *r) {
+static int read_skip(VP9_COMMON *cm, const MACROBLOCKD *xd,
+ int segment_id, vp9_reader *r) {
if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
return 1;
} else {
@@ -169,7 +169,7 @@
const BLOCK_SIZE bsize = mbmi->sb_type;
mbmi->segment_id = read_intra_segment_id(cm, xd, mi_row, mi_col, r);
- mbmi->skip_coeff = read_skip_coeff(cm, xd, mbmi->segment_id, r);
+ mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
mbmi->tx_size = read_tx_size(cm, xd, cm->tx_mode, bsize, 1, r);
mbmi->ref_frame[0] = INTRA_FRAME;
mbmi->ref_frame[1] = NONE;
@@ -520,10 +520,10 @@
mbmi->mv[0].as_int = 0;
mbmi->mv[1].as_int = 0;
mbmi->segment_id = read_inter_segment_id(cm, xd, mi_row, mi_col, r);
- mbmi->skip_coeff = read_skip_coeff(cm, xd, mbmi->segment_id, r);
+ mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
inter_block = read_is_inter_block(cm, xd, mbmi->segment_id, r);
mbmi->tx_size = read_tx_size(cm, xd, cm->tx_mode, mbmi->sb_type,
- !mbmi->skip_coeff || !inter_block, r);
+ !mbmi->skip || !inter_block, r);
if (inter_block)
read_inter_block_mode_info(cm, xd, tile, mi, mi_row, mi_col, r);
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index c8f334f..01ee92f 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -104,13 +104,13 @@
}
}
-static int write_skip_coeff(const VP9_COMP *cpi, int segment_id, MODE_INFO *m,
- vp9_writer *w) {
+static int write_skip(const VP9_COMP *cpi, int segment_id, MODE_INFO *m,
+ vp9_writer *w) {
const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
if (vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
return 1;
} else {
- const int skip = m->mbmi.skip_coeff;
+ const int skip = m->mbmi.skip;
vp9_write(w, skip, vp9_get_skip_prob(&cpi->common, xd));
return skip;
}
@@ -253,7 +253,7 @@
const MV_REFERENCE_FRAME sec_rf = mi->ref_frame[1];
const MB_PREDICTION_MODE mode = mi->mode;
const int segment_id = mi->segment_id;
- int skip_coeff;
+ int skip;
const BLOCK_SIZE bsize = mi->sb_type;
const int allow_hp = cm->allow_high_precision_mv;
@@ -273,14 +273,14 @@
}
}
- skip_coeff = write_skip_coeff(cpi, segment_id, m, bc);
+ skip = write_skip(cpi, segment_id, m, bc);
if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
vp9_write(bc, rf != INTRA_FRAME, vp9_get_intra_inter_prob(cm, xd));
if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
!(rf != INTRA_FRAME &&
- (skip_coeff || vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)))) {
+ (skip || vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)))) {
write_selected_tx_size(cpi, m, mi->tx_size, bsize, bc);
}
@@ -382,7 +382,7 @@
if (seg->update_map)
write_segment_id(bc, seg, m->mbmi.segment_id);
- write_skip_coeff(cpi, segment_id, m, bc);
+ write_skip(cpi, segment_id, m, bc);
if (m->mbmi.sb_type >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
write_selected_tx_size(cpi, m, m->mbmi.tx_size, m->mbmi.sb_type, bc);
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 5432c1f..73c1992 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -365,8 +365,6 @@
(bw * bh);
if (projected_rate < (target_rate / 4)) {
- segment = 2;
- } else if (projected_rate < (target_rate / 2)) {
segment = 1;
} else {
segment = 0;
@@ -660,14 +658,25 @@
x->skip_recode = 0;
// Set to zero to make sure we do not use the previous encoded frame stats
- xd->mi_8x8[0]->mbmi.skip_coeff = 0;
+ xd->mi_8x8[0]->mbmi.skip = 0;
x->source_variance = get_sby_perpixel_variance(cpi, x, bsize);
if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
: vp9_block_energy(cpi, x, bsize);
- xd->mi_8x8[0]->mbmi.segment_id = vp9_vaq_segment_id(energy);
+
+ if (cm->frame_type == KEY_FRAME ||
+ cpi->refresh_alt_ref_frame ||
+ (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+ xd->mi_8x8[0]->mbmi.segment_id = vp9_vaq_segment_id(energy);
+ } else {
+ const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
+ : cm->last_frame_seg_map;
+ xd->mi_8x8[0]->mbmi.segment_id =
+ vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
+ }
+
rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
vp9_mb_init_quantizer(cpi, x);
}
@@ -681,11 +690,12 @@
} else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
const int mi_offset = mi_row * cm->mi_cols + mi_col;
unsigned char complexity = cpi->complexity_map[mi_offset];
- const int is_edge = (mi_row == 0) || (mi_row == (cm->mi_rows - 1)) ||
- (mi_col == 0) || (mi_col == (cm->mi_cols - 1));
+ const int is_edge = (mi_row <= 1) || (mi_row >= (cm->mi_rows - 2)) ||
+ (mi_col <= 1) || (mi_col >= (cm->mi_cols - 2));
- if (!is_edge && (complexity > 128))
+ if (!is_edge && (complexity > 128)) {
x->rdmult = x->rdmult + ((x->rdmult * (complexity - 128)) / 256);
+ }
}
// Find best coding mode & reconstruct the MB so it is available
@@ -709,6 +719,9 @@
*totalrate = round(*totalrate * rdmult_ratio);
}
}
+ else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+ x->rdmult = orig_rdmult;
+ }
}
static void update_stats(VP9_COMP *cpi) {
@@ -2087,7 +2100,7 @@
for (y = 0; y < ymbs; y++) {
for (x = 0; x < xmbs; x++) {
- if (!mi_8x8[y * mis + x]->mbmi.skip_coeff)
+ if (!mi_8x8[y * mis + x]->mbmi.skip)
return 0;
}
}
@@ -2246,7 +2259,7 @@
mbmi->ref_frame[1] = INTRA_FRAME;
mbmi->tx_size = max_txsize_lookup[bsize];
mbmi->uv_mode = mode;
- mbmi->skip_coeff = 0;
+ mbmi->skip = 0;
mbmi->sb_type = bsize;
mbmi->segment_id = 0;
}
@@ -2723,7 +2736,7 @@
if (!is_inter_block(mbmi)) {
int plane;
- mbmi->skip_coeff = 1;
+ mbmi->skip = 1;
for (plane = 0; plane < MAX_MB_PLANE; ++plane)
vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
if (output_enabled)
@@ -2742,11 +2755,11 @@
if (!is_inter_block(mbmi)) {
vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
} else if (!x->skip) {
- mbmi->skip_coeff = 1;
+ mbmi->skip = 1;
vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
} else {
- mbmi->skip_coeff = 1;
+ mbmi->skip = 1;
if (output_enabled)
cm->counts.skip[vp9_get_skip_context(xd)][1]++;
reset_skip_context(xd, MAX(bsize, BLOCK_8X8));
@@ -2756,7 +2769,7 @@
if (cm->tx_mode == TX_MODE_SELECT &&
mbmi->sb_type >= BLOCK_8X8 &&
!(is_inter_block(mbmi) &&
- (mbmi->skip_coeff ||
+ (mbmi->skip ||
vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)))) {
++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
&cm->counts.tx)[mbmi->tx_size];
diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index ad28f31..8770107 100644
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -32,7 +32,7 @@
struct encode_b_args {
MACROBLOCK *x;
struct optimize_ctx *ctx;
- unsigned char *skip_coeff;
+ unsigned char *skip;
};
void vp9_subtract_block_c(int rows, int cols,
@@ -439,7 +439,7 @@
}
if (p->eobs[block])
- *(args->skip_coeff) = 0;
+ *(args->skip) = 0;
if (x->skip_encode || p->eobs[block] == 0)
return;
@@ -489,7 +489,7 @@
MACROBLOCKD *const xd = &x->e_mbd;
struct optimize_ctx ctx;
MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
- struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
+ struct encode_b_args arg = {x, &ctx, &mbmi->skip};
vp9_subtract_sby(x, bsize);
if (x->optimize)
@@ -503,7 +503,7 @@
MACROBLOCKD *const xd = &x->e_mbd;
struct optimize_ctx ctx;
MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
- struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
+ struct encode_b_args arg = {x, &ctx, &mbmi->skip};
if (!x->skip_recode)
vp9_subtract_sb(x, bsize);
@@ -648,20 +648,20 @@
assert(0);
}
if (*eob)
- *(args->skip_coeff) = 0;
+ *(args->skip) = 0;
}
void vp9_encode_block_intra(MACROBLOCK *x, int plane, int block,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
- unsigned char *skip_coeff) {
- struct encode_b_args arg = {x, NULL, skip_coeff};
+ unsigned char *skip) {
+ struct encode_b_args arg = {x, NULL, skip};
encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
}
void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
const MACROBLOCKD *const xd = &x->e_mbd;
- struct encode_b_args arg = {x, NULL, &xd->mi_8x8[0]->mbmi.skip_coeff};
+ struct encode_b_args arg = {x, NULL, &xd->mi_8x8[0]->mbmi.skip};
vp9_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block_intra,
&arg);
diff --git a/vp9/encoder/vp9_encodemb.h b/vp9/encoder/vp9_encodemb.h
index cf7097d..515935f 100644
--- a/vp9/encoder/vp9_encodemb.h
+++ b/vp9/encoder/vp9_encodemb.h
@@ -32,7 +32,7 @@
void vp9_encode_block_intra(MACROBLOCK *x, int plane, int block,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
- unsigned char *skip_coeff);
+ unsigned char *skip);
void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 5efb00a..bf9dd3e 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -1566,13 +1566,7 @@
(i >= MIN_GF_INTERVAL) &&
// for real scene cuts (not forced kfs) dont allow arf very near kf.
(rc->next_key_frame_forced ||
- (i <= (rc->frames_to_key - MIN_GF_INTERVAL))) &&
- ((next_frame.pcnt_inter > 0.75) ||
- (next_frame.pcnt_second_ref > 0.5)) &&
- ((mv_in_out_accumulator / (double)i > -0.2) ||
- (mv_in_out_accumulator > -2.0)) &&
- (boost_score > 100)) {
-
+ (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
// Alternative boost calculation for alt ref
rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
&b_boost);
diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c
index e972355..cef7e04 100644
--- a/vp9/encoder/vp9_onyx_if.c
+++ b/vp9/encoder/vp9_onyx_if.c
@@ -96,7 +96,7 @@
void vp9_init_quantizer(VP9_COMP *cpi);
static const double in_frame_q_adj_ratio[MAX_SEGMENTS] =
- {1.0, 1.5, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0};
+ {1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
static INLINE void Scale2Ratio(int mode, int *hr, int *hs) {
switch (mode) {
@@ -267,7 +267,6 @@
// Clear down the complexity map used for rd
vpx_memset(cpi->complexity_map, 0, cm->mi_rows * cm->mi_cols);
- // Enable segmentation
vp9_enable_segmentation((VP9_PTR)cpi);
vp9_clearall_segfeatures(seg);
@@ -278,7 +277,7 @@
vp9_disable_segfeature(seg, 0, SEG_LVL_ALT_Q);
// Use some of the segments for in frame Q adjustment
- for (segment = 1; segment < 3; segment++) {
+ for (segment = 1; segment < 2; segment++) {
qindex_delta =
vp9_compute_qdelta_by_rate(cpi, cm->base_qindex,
in_frame_q_adj_ratio[segment]);
@@ -3707,8 +3706,9 @@
xd->interp_kernel = vp9_get_interp_kernel(
DEFAULT_INTERP_FILTER == SWITCHABLE ? EIGHTTAP : DEFAULT_INTERP_FILTER);
- if (cpi->oxcf.aq_mode == VARIANCE_AQ)
+ if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
vp9_vaq_init();
+ }
if (cpi->use_svc) {
SvcEncode(cpi, size, dest, frame_flags);
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 9129685..6fed3d5 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -170,7 +170,7 @@
tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
mbmi->interp_filter = cpi->common.interp_filter == SWITCHABLE ?
EIGHTTAP : cpi->common.interp_filter;
- mbmi->skip_coeff = 0;
+ mbmi->skip = 0;
mbmi->segment_id = 0;
for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) {
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 03def20..aed4de5 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -650,8 +650,7 @@
return;
if (!is_inter_block(mbmi))
- vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size,
- &mbmi->skip_coeff);
+ vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
else
vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 0f68d83..510ef78 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -314,7 +314,7 @@
const int skip_inc = !vp9_segfeature_active(&cm->seg, mbmi->segment_id,
SEG_LVL_SKIP);
struct tokenize_b_args arg = {cpi, xd, t, cpi->mb.token_cache};
- if (mbmi->skip_coeff) {
+ if (mbmi->skip) {
if (!dry_run)
cm->counts.skip[ctx][1] += skip_inc;
reset_skip_context(xd, bsize);
diff --git a/vp9/encoder/vp9_vaq.c b/vp9/encoder/vp9_vaq.c
index 1f9cb87..acd7c41 100644
--- a/vp9/encoder/vp9_vaq.c
+++ b/vp9/encoder/vp9_vaq.c
@@ -19,8 +19,8 @@
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/common/vp9_systemdependent.h"
-#define ENERGY_MIN (-3)
-#define ENERGY_MAX (3)
+#define ENERGY_MIN (-1)
+#define ENERGY_MAX (1)
#define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN + 1)
#define ENERGY_IN_BOUNDS(energy)\
assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
@@ -65,7 +65,7 @@
vp9_clear_system_state(); // __asm emms;
- base_ratio = 1.8;
+ base_ratio = 1.5;
for (i = ENERGY_MIN; i <= ENERGY_MAX; i++) {
Q_RATIO(i) = pow(base_ratio, i/3.0);
@@ -80,30 +80,34 @@
cm->y_dc_delta_q);
int i;
- vp9_enable_segmentation((VP9_PTR)cpi);
- vp9_clearall_segfeatures(seg);
+ if (cm->frame_type == KEY_FRAME ||
+ cpi->refresh_alt_ref_frame ||
+ (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+ vp9_enable_segmentation((VP9_PTR)cpi);
+ vp9_clearall_segfeatures(seg);
- seg->abs_delta = SEGMENT_DELTADATA;
+ seg->abs_delta = SEGMENT_DELTADATA;
- vp9_clear_system_state(); // __asm emms;
+ vp9_clear_system_state(); // __asm emms;
- for (i = ENERGY_MIN; i <= ENERGY_MAX; i++) {
- int qindex_delta, segment_rdmult;
+ for (i = ENERGY_MIN; i <= ENERGY_MAX; i++) {
+ int qindex_delta, segment_rdmult;
- if (Q_RATIO(i) == 1) {
- // No need to enable SEG_LVL_ALT_Q for this segment
- RDMULT_RATIO(i) = 1;
- continue;
+ if (Q_RATIO(i) == 1) {
+ // No need to enable SEG_LVL_ALT_Q for this segment
+ RDMULT_RATIO(i) = 1;
+ continue;
+ }
+
+ qindex_delta = vp9_compute_qdelta(cpi, base_q, base_q * Q_RATIO(i));
+ vp9_set_segdata(seg, SEGMENT_ID(i), SEG_LVL_ALT_Q, qindex_delta);
+ vp9_enable_segfeature(seg, SEGMENT_ID(i), SEG_LVL_ALT_Q);
+
+ segment_rdmult = vp9_compute_rd_mult(cpi, cm->base_qindex + qindex_delta +
+ cm->y_dc_delta_q);
+
+ RDMULT_RATIO(i) = (double) segment_rdmult / base_rdmult;
}
-
- qindex_delta = vp9_compute_qdelta(cpi, base_q, base_q * Q_RATIO(i));
- vp9_set_segdata(seg, SEGMENT_ID(i), SEG_LVL_ALT_Q, qindex_delta);
- vp9_enable_segfeature(seg, SEGMENT_ID(i), SEG_LVL_ALT_Q);
-
- segment_rdmult = vp9_compute_rd_mult(cpi, cm->base_qindex + qindex_delta +
- cm->y_dc_delta_q);
-
- RDMULT_RATIO(i) = (double) segment_rdmult / base_rdmult;
}
}
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index 85e83b8..48d6a7c 100644
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -79,6 +79,7 @@
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_subpixel_bilinear_sse2.asm
VP9_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/vp9_subpixel_8t_ssse3.asm
VP9_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/vp9_subpixel_bilinear_ssse3.asm
+VP9_COMMON_SRCS-$(HAVE_AVX2) += common/x86/vp9_subpixel_8t_intrin_avx2.c
ifeq ($(CONFIG_VP9_POSTPROC),yes)
VP9_COMMON_SRCS-$(HAVE_MMX) += common/x86/vp9_postproc_mmx.asm
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_postproc_sse2.asm
diff --git a/vpx_ports/vpx_ports.mk b/vpx_ports/vpx_ports.mk
index e6cb52f..869a204 100644
--- a/vpx_ports/vpx_ports.mk
+++ b/vpx_ports/vpx_ports.mk
@@ -19,7 +19,6 @@
PORTS_SRCS-$(BUILD_LIBVPX) += emms.asm
PORTS_SRCS-$(BUILD_LIBVPX) += x86.h
PORTS_SRCS-$(BUILD_LIBVPX) += x86_abi_support.asm
-PORTS_SRCS-$(BUILD_LIBVPX) += x86_cpuid.c
endif
PORTS_SRCS-$(ARCH_ARM) += arm_cpudetect.c
diff --git a/vpx_ports/x86.h b/vpx_ports/x86.h
index 603e2b6..bc99f89 100644
--- a/vpx_ports/x86.h
+++ b/vpx_ports/x86.h
@@ -168,8 +168,6 @@
return flags & mask;
}
-vpx_cpu_t vpx_x86_vendor(void);
-
#if ARCH_X86_64 && defined(_MSC_VER)
unsigned __int64 __rdtsc(void);
#pragma intrinsic(__rdtsc)
diff --git a/vpx_ports/x86_cpuid.c b/vpx_ports/x86_cpuid.c
deleted file mode 100644
index 02d382c..0000000
--- a/vpx_ports/x86_cpuid.c
+++ /dev/null
@@ -1,49 +0,0 @@
-/*
- * Copyright (c) 2010 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 <string.h>
-#include "x86.h"
-
-struct cpuid_vendors {
- char vendor_string[12];
- vpx_cpu_t vendor_id;
-};
-
-static struct cpuid_vendors cpuid_vendor_list[VPX_CPU_LAST] = {
- { "AuthenticAMD", VPX_CPU_AMD },
- { "AMDisbetter!", VPX_CPU_AMD_OLD },
- { "CentaurHauls", VPX_CPU_CENTAUR },
- { "CyrixInstead", VPX_CPU_CYRIX },
- { "GenuineIntel", VPX_CPU_INTEL },
- { "NexGenDriven", VPX_CPU_NEXGEN },
- { "Geode by NSC", VPX_CPU_NSC },
- { "RiseRiseRise", VPX_CPU_RISE },
- { "SiS SiS SiS ", VPX_CPU_SIS },
- { "GenuineTMx86", VPX_CPU_TRANSMETA },
- { "TransmetaCPU", VPX_CPU_TRANSMETA_OLD },
- { "UMC UMC UMC ", VPX_CPU_UMC },
- { "VIA VIA VIA ", VPX_CPU_VIA },
-};
-
-vpx_cpu_t vpx_x86_vendor(void) {
- unsigned int reg_eax;
- unsigned int vs[3];
- int i;
-
- /* Get the Vendor String from the CPU */
- cpuid(0, 0, reg_eax, vs[0], vs[2], vs[1]);
-
- for (i = 0; i < VPX_CPU_LAST; i++) {
- if (strncmp((const char *)vs, cpuid_vendor_list[i].vendor_string, 12) == 0)
- return (cpuid_vendor_list[i].vendor_id);
- }
-
- return VPX_CPU_UNKNOWN;
-}
