Merge "Combining integer and fractional parts of mvs for entropy coding." into experimental
diff --git a/test/dct16x16_test.cc b/test/dct16x16_test.cc
index f6d2d59..9fb45d6 100644
--- a/test/dct16x16_test.cc
+++ b/test/dct16x16_test.cc
@@ -17,6 +17,7 @@
extern "C" {
#include "vp9/common/vp9_entropy.h"
#include "vp9_rtcd.h"
+void vp9_short_idct16x16_add_c(short *input, uint8_t *output, int pitch);
}
#include "acm_random.h"
@@ -269,19 +270,23 @@
const int count_test_block = 1000;
for (int i = 0; i < count_test_block; ++i) {
int16_t in[256], coeff[256];
- int16_t out_c[256];
+ uint8_t dst[256], src[256];
double out_r[256];
+ for (int j = 0; j < 256; ++j) {
+ src[j] = rnd.Rand8();
+ dst[j] = rnd.Rand8();
+ }
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 256; ++j)
- in[j] = rnd.Rand8() - rnd.Rand8();
+ in[j] = src[j] - dst[j];
reference_16x16_dct_2d(in, out_r);
for (int j = 0; j < 256; j++)
coeff[j] = round(out_r[j]);
- vp9_short_idct16x16_c(coeff, out_c, 32);
+ vp9_short_idct16x16_add_c(coeff, dst, 16);
for (int j = 0; j < 256; ++j) {
- const int diff = out_c[j] - in[j];
+ const int diff = dst[j] - src[j];
const int error = diff * diff;
EXPECT_GE(1, error)
<< "Error: 16x16 IDCT has error " << error
@@ -289,7 +294,7 @@
}
}
}
-#if 1
+
// we need enable fdct test once we re-do the 16 point fdct.
TEST(VP9Fdct16x16Test, AccuracyCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
@@ -299,18 +304,22 @@
for (int i = 0; i < count_test_block; ++i) {
int16_t test_input_block[256];
int16_t test_temp_block[256];
- int16_t test_output_block[256];
+ uint8_t dst[256], src[256];
+ for (int j = 0; j < 256; ++j) {
+ src[j] = rnd.Rand8();
+ dst[j] = rnd.Rand8();
+ }
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 256; ++j)
- test_input_block[j] = rnd.Rand8() - rnd.Rand8();
+ test_input_block[j] = src[j] - dst[j];
const int pitch = 32;
vp9_short_fdct16x16_c(test_input_block, test_temp_block, pitch);
- vp9_short_idct16x16_c(test_temp_block, test_output_block, pitch);
+ vp9_short_idct16x16_add_c(test_temp_block, dst, 16);
for (int j = 0; j < 256; ++j) {
- const int diff = test_input_block[j] - test_output_block[j];
+ const int diff = dst[j] - src[j];
const int error = diff * diff;
if (max_error < error)
max_error = error;
@@ -354,6 +363,4 @@
}
}
}
-#endif
-
} // namespace
diff --git a/test/dct32x32_test.cc b/test/dct32x32_test.cc
index a565270..e05d482 100644
--- a/test/dct32x32_test.cc
+++ b/test/dct32x32_test.cc
@@ -18,7 +18,7 @@
#include "vp9/common/vp9_entropy.h"
#include "./vp9_rtcd.h"
void vp9_short_fdct32x32_c(int16_t *input, int16_t *out, int pitch);
- void vp9_short_idct32x32_c(short *input, short *output, int pitch);
+ void vp9_short_idct32x32_add_c(short *input, uint8_t *output, int pitch);
}
#include "test/acm_random.h"
@@ -91,28 +91,31 @@
}
}
-
TEST(VP9Idct32x32Test, AccuracyCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
const int count_test_block = 1000;
for (int i = 0; i < count_test_block; ++i) {
int16_t in[1024], coeff[1024];
- int16_t out_c[1024];
+ uint8_t dst[1024], src[1024];
double out_r[1024];
+ for (int j = 0; j < 1024; ++j) {
+ src[j] = rnd.Rand8();
+ dst[j] = rnd.Rand8();
+ }
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 1024; ++j)
- in[j] = rnd.Rand8() - rnd.Rand8();
+ in[j] = src[j] - dst[j];
reference_32x32_dct_2d(in, out_r);
for (int j = 0; j < 1024; j++)
coeff[j] = round(out_r[j]);
- vp9_short_idct32x32_c(coeff, out_c, 64);
+ vp9_short_idct32x32_add_c(coeff, dst, 32);
for (int j = 0; j < 1024; ++j) {
- const int diff = out_c[j] - in[j];
+ const int diff = dst[j] - src[j];
const int error = diff * diff;
EXPECT_GE(1, error)
- << "Error: 3x32 IDCT has error " << error
+ << "Error: 32x32 IDCT has error " << error
<< " at index " << j;
}
}
@@ -126,18 +129,22 @@
for (int i = 0; i < count_test_block; ++i) {
int16_t test_input_block[1024];
int16_t test_temp_block[1024];
- int16_t test_output_block[1024];
+ uint8_t dst[1024], src[1024];
+ for (int j = 0; j < 1024; ++j) {
+ src[j] = rnd.Rand8();
+ dst[j] = rnd.Rand8();
+ }
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 1024; ++j)
- test_input_block[j] = rnd.Rand8() - rnd.Rand8();
+ test_input_block[j] = src[j] - dst[j];
const int pitch = 64;
vp9_short_fdct32x32_c(test_input_block, test_temp_block, pitch);
- vp9_short_idct32x32_c(test_temp_block, test_output_block, pitch);
+ vp9_short_idct32x32_add_c(test_temp_block, dst, 32);
for (int j = 0; j < 1024; ++j) {
- const unsigned diff = test_input_block[j] - test_output_block[j];
+ const unsigned diff = dst[j] - src[j];
const unsigned error = diff * diff;
if (max_error < error)
max_error = error;
diff --git a/vp9/common/vp9_blockd.h b/vp9/common/vp9_blockd.h
index d14ed45..6a3fbb1 100644
--- a/vp9/common/vp9_blockd.h
+++ b/vp9/common/vp9_blockd.h
@@ -222,12 +222,21 @@
static INLINE int mi_width_log2(BLOCK_SIZE_TYPE sb_type) {
int a = b_width_log2(sb_type) - 1;
+#if CONFIG_AB4X4
+ // align 4x4 block to mode_info
+ if (a < 0)
+ a = 0;
+#endif
assert(a >= 0);
return a;
}
static INLINE int mi_height_log2(BLOCK_SIZE_TYPE sb_type) {
int a = b_height_log2(sb_type) - 1;
+#if CONFIG_AB4X4
+ if (a < 0)
+ a = 0;
+#endif
assert(a >= 0);
return a;
}
@@ -263,7 +272,7 @@
typedef struct {
MB_MODE_INFO mbmi;
- union b_mode_info bmi[4];
+ union b_mode_info bmi[16];
} MODE_INFO;
struct scale_factors {
@@ -442,9 +451,12 @@
int bhl = mi_height_log2(sb_type);
int boffset = mi_width_log2(BLOCK_SIZE_SB64X64) - bsl;
int i;
- // skip macroblock partition
+
+#if !CONFIG_AB4X4
+ // skip 8x8 block partition
if (bsl == 0)
return;
+#endif
// update the partition context at the end notes. set partition bits
// of block sizes larger than the current one to be one, and partition
@@ -492,7 +504,11 @@
above = (above > 0);
left = (left > 0);
+#if CONFIG_AB4X4
+ return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+#else
return (left * 2 + above) + (bsl - 1) * PARTITION_PLOFFSET;
+#endif
}
static BLOCK_SIZE_TYPE get_subsize(BLOCK_SIZE_TYPE bsize,
@@ -509,6 +525,10 @@
subsize = BLOCK_SIZE_SB32X16;
else if (bsize == BLOCK_SIZE_MB16X16)
subsize = BLOCK_SIZE_SB16X8;
+#if CONFIG_AB4X4
+ else if (bsize == BLOCK_SIZE_SB8X8)
+ subsize = BLOCK_SIZE_SB8X4;
+#endif
else
assert(0);
break;
@@ -519,6 +539,10 @@
subsize = BLOCK_SIZE_SB16X32;
else if (bsize == BLOCK_SIZE_MB16X16)
subsize = BLOCK_SIZE_SB8X16;
+#if CONFIG_AB4X4
+ else if (bsize == BLOCK_SIZE_SB8X8)
+ subsize = BLOCK_SIZE_SB4X8;
+#endif
else
assert(0);
break;
@@ -529,6 +553,10 @@
subsize = BLOCK_SIZE_MB16X16;
else if (bsize == BLOCK_SIZE_MB16X16)
subsize = BLOCK_SIZE_SB8X8;
+#if CONFIG_AB4X4
+ else if (bsize == BLOCK_SIZE_SB8X8)
+ subsize = BLOCK_SIZE_AB4X4;
+#endif
else
assert(0);
break;
@@ -538,12 +566,6 @@
return subsize;
}
-#define ACTIVE_HT 110 // quantization stepsize threshold
-
-#define ACTIVE_HT8 300
-
-#define ACTIVE_HT16 300
-
// convert MB_PREDICTION_MODE to B_PREDICTION_MODE
static B_PREDICTION_MODE pred_mode_conv(MB_PREDICTION_MODE mode) {
switch (mode) {
@@ -572,6 +594,7 @@
case B_V_PRED :
case B_D117_PRED :
+ case B_D63_PRED:
return ADST_DCT;
case B_H_PRED :
@@ -596,21 +619,17 @@
// is smaller than the prediction size
TX_TYPE tx_type = DCT_DCT;
const BLOCK_SIZE_TYPE sb_type = xd->mode_info_context->mbmi.sb_type;
- const int wb = b_width_log2(sb_type), hb = b_height_log2(sb_type);
+ const int wb = b_width_log2(sb_type);
#if !USE_ADST_FOR_SB
if (sb_type > BLOCK_SIZE_MB16X16)
return tx_type;
#endif
- if (ib >= (1 << (wb + hb))) // no chroma adst
- return tx_type;
if (xd->lossless)
return DCT_DCT;
- if (xd->mode_info_context->mbmi.mode == I4X4_PRED &&
- xd->q_index < ACTIVE_HT) {
+ if (xd->mode_info_context->mbmi.mode == I4X4_PRED) {
tx_type = txfm_map(
xd->mode_info_context->bmi[ib].as_mode.first);
- } else if (xd->mode_info_context->mbmi.mode <= TM_PRED &&
- xd->q_index < ACTIVE_HT) {
+ } else if (xd->mode_info_context->mbmi.mode <= TM_PRED) {
#if USE_ADST_FOR_I16X16_4X4
#if USE_ADST_PERIPHERY_ONLY
const int hmax = 1 << wb;
@@ -653,8 +672,7 @@
#endif
if (ib >= (1 << (wb + hb))) // no chroma adst
return tx_type;
- if (xd->mode_info_context->mbmi.mode <= TM_PRED &&
- xd->q_index < ACTIVE_HT8) {
+ if (xd->mode_info_context->mbmi.mode <= TM_PRED) {
#if USE_ADST_FOR_I16X16_8X8
#if USE_ADST_PERIPHERY_ONLY
const int hmax = 1 << wb;
@@ -695,8 +713,7 @@
#endif
if (ib >= (1 << (wb + hb)))
return tx_type;
- if (xd->mode_info_context->mbmi.mode <= TM_PRED &&
- xd->q_index < ACTIVE_HT16) {
+ if (xd->mode_info_context->mbmi.mode <= TM_PRED) {
tx_type = txfm_map(pred_mode_conv(xd->mode_info_context->mbmi.mode));
#if USE_ADST_PERIPHERY_ONLY
if (sb_type > BLOCK_SIZE_MB16X16) {
diff --git a/vp9/common/vp9_common.h b/vp9/common/vp9_common.h
index dbfb9ed..b6252d9 100644
--- a/vp9/common/vp9_common.h
+++ b/vp9/common/vp9_common.h
@@ -52,6 +52,10 @@
return value < low ? low : (value > high ? high : value);
}
+static INLINE double fclamp(double value, double low, double high) {
+ return value < low ? low : (value > high ? high : value);
+}
+
static INLINE int multiple16(int value) {
return (value + 15) & ~15;
}
diff --git a/vp9/common/vp9_entropymode.c b/vp9/common/vp9_entropymode.c
index dcee62f..577aab5 100644
--- a/vp9/common/vp9_entropymode.c
+++ b/vp9/common/vp9_entropymode.c
@@ -106,6 +106,12 @@
const vp9_prob vp9_partition_probs[NUM_PARTITION_CONTEXTS]
[PARTITION_TYPES - 1] = {
// FIXME(jingning,rbultje) put real probabilities here
+#if CONFIG_AB4X4
+ {202, 162, 107},
+ {16, 2, 169},
+ {3, 246, 19},
+ {104, 90, 134},
+#endif
{202, 162, 107},
{16, 2, 169},
{3, 246, 19},
@@ -513,6 +519,7 @@
vp9_sub_mv_ref_tree, fc->sub_mv_ref_counts[i],
fc->pre_sub_mv_ref_prob[i], fc->sub_mv_ref_prob[i],
LEFT4X4);
+
for (i = 0; i < NUM_PARTITION_CONTEXTS; i++)
update_mode_probs(PARTITION_TYPES, vp9_partition_tree,
fc->partition_counts[i], fc->pre_partition_prob[i],
diff --git a/vp9/common/vp9_enums.h b/vp9/common/vp9_enums.h
index 2f67074..626f0b9 100644
--- a/vp9/common/vp9_enums.h
+++ b/vp9/common/vp9_enums.h
@@ -48,6 +48,10 @@
} PARTITION_TYPE;
#define PARTITION_PLOFFSET 4 // number of probability models per block size
+#if CONFIG_AB4X4
+#define NUM_PARTITION_CONTEXTS (4 * PARTITION_PLOFFSET)
+#else
#define NUM_PARTITION_CONTEXTS (3 * PARTITION_PLOFFSET)
+#endif
#endif // VP9_COMMON_VP9_ENUMS_H_
diff --git a/vp9/common/vp9_idct.c b/vp9/common/vp9_idct.c
index 3ec093f..b166fcb 100644
--- a/vp9/common/vp9_idct.c
+++ b/vp9/common/vp9_idct.c
@@ -621,10 +621,9 @@
output[15] = step2[0] - step2[15];
}
-void vp9_short_idct16x16_c(int16_t *input, int16_t *output, int pitch) {
+void vp9_short_idct16x16_add_c(int16_t *input, uint8_t *dest, int dest_stride) {
int16_t out[16 * 16];
int16_t *outptr = out;
- const int half_pitch = pitch >> 1;
int i, j;
int16_t temp_in[16], temp_out[16];
@@ -641,7 +640,8 @@
temp_in[j] = out[j * 16 + i];
idct16_1d(temp_in, temp_out);
for (j = 0; j < 16; ++j)
- output[j * half_pitch + i] = ROUND_POWER_OF_TWO(temp_out[j], 6);
+ dest[j * dest_stride + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+ + dest[j * dest_stride + i]);
}
}
@@ -823,8 +823,8 @@
{ iadst16_1d, iadst16_1d } // ADST_ADST = 3
};
-void vp9_short_iht16x16_c(int16_t *input, int16_t *output,
- int pitch, int tx_type) {
+void vp9_short_iht16x16_add_c(int16_t *input, uint8_t *dest, int dest_stride,
+ int tx_type) {
int i, j;
int16_t out[16 * 16];
int16_t *outptr = out;
@@ -844,38 +844,38 @@
temp_in[j] = out[j * 16 + i];
ht.cols(temp_in, temp_out);
for (j = 0; j < 16; ++j)
- output[j * pitch + i] = ROUND_POWER_OF_TWO(temp_out[j], 6);
+ dest[j * dest_stride + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+ + dest[j * dest_stride + i]); }
+}
+
+void vp9_short_idct10_16x16_add_c(int16_t *input, uint8_t *dest,
+ int dest_stride) {
+ int16_t out[16 * 16];
+ int16_t *outptr = out;
+ int i, j;
+ int16_t temp_in[16], temp_out[16];
+
+ /* First transform rows. Since all non-zero dct coefficients are in
+ * upper-left 4x4 area, we only need to calculate first 4 rows here.
+ */
+ vpx_memset(out, 0, sizeof(out));
+ for (i = 0; i < 4; ++i) {
+ idct16_1d(input, outptr);
+ input += 16;
+ outptr += 16;
+ }
+
+ // Then transform columns
+ for (i = 0; i < 16; ++i) {
+ for (j = 0; j < 16; ++j)
+ temp_in[j] = out[j*16 + i];
+ idct16_1d(temp_in, temp_out);
+ for (j = 0; j < 16; ++j)
+ dest[j * dest_stride + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+ + dest[j * dest_stride + i]);
}
}
-void vp9_short_idct10_16x16_c(int16_t *input, int16_t *output, int pitch) {
- int16_t out[16 * 16];
- int16_t *outptr = out;
- const int half_pitch = pitch >> 1;
- int i, j;
- int16_t temp_in[16], temp_out[16];
-
- /* First transform rows. Since all non-zero dct coefficients are in
- * upper-left 4x4 area, we only need to calculate first 4 rows here.
- */
- vpx_memset(out, 0, sizeof(out));
- for (i = 0; i < 4; ++i) {
- idct16_1d(input, outptr);
- input += 16;
- outptr += 16;
- }
-
- // Then transform columns
- for (i = 0; i < 16; ++i) {
- for (j = 0; j < 16; ++j)
- temp_in[j] = out[j*16 + i];
- idct16_1d(temp_in, temp_out);
- for (j = 0; j < 16; ++j)
- output[j * half_pitch + i] = ROUND_POWER_OF_TWO(temp_out[j], 6);
- }
-}
-
-
void vp9_short_idct1_16x16_c(int16_t *input, int16_t *output) {
int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
out = dct_const_round_shift(out * cospi_16_64);
@@ -1249,10 +1249,9 @@
output[31] = step1[0] - step1[31];
}
-void vp9_short_idct32x32_c(int16_t *input, int16_t *output, int pitch) {
+void vp9_short_idct32x32_add_c(int16_t *input, uint8_t *dest, int dest_stride) {
int16_t out[32 * 32];
int16_t *outptr = out;
- const int half_pitch = pitch >> 1;
int i, j;
int16_t temp_in[32], temp_out[32];
@@ -1269,7 +1268,8 @@
temp_in[j] = out[j * 32 + i];
idct32_1d(temp_in, temp_out);
for (j = 0; j < 32; ++j)
- output[j * half_pitch + i] = ROUND_POWER_OF_TWO(temp_out[j], 6);
+ dest[j * dest_stride + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+ + dest[j * dest_stride + i]);
}
}
@@ -1279,10 +1279,10 @@
output[0] = ROUND_POWER_OF_TWO(out, 6);
}
-void vp9_short_idct10_32x32_c(int16_t *input, int16_t *output, int pitch) {
+void vp9_short_idct10_32x32_add_c(int16_t *input, uint8_t *dest,
+ int dest_stride) {
int16_t out[32 * 32];
int16_t *outptr = out;
- const int half_pitch = pitch >> 1;
int i, j;
int16_t temp_in[32], temp_out[32];
@@ -1302,6 +1302,7 @@
temp_in[j] = out[j * 32 + i];
idct32_1d(temp_in, temp_out);
for (j = 0; j < 32; ++j)
- output[j * half_pitch + i] = ROUND_POWER_OF_TWO(temp_out[j], 6);
+ dest[j * dest_stride + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+ + dest[j * dest_stride + i]);
}
}
diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h
index 80fccd5..589984f 100644
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -49,9 +49,6 @@
void sym(uint8_t *y, uint8_t *u, uint8_t *v, \
int ystride, int uv_stride, struct loop_filter_info *lfi)
-#define prototype_simple_loopfilter(sym) \
- void sym(uint8_t *y, int ystride, const unsigned char *blimit)
-
#if ARCH_X86 || ARCH_X86_64
#include "x86/vp9_loopfilter_x86.h"
#endif
diff --git a/vp9/common/vp9_loopfilter_filters.c b/vp9/common/vp9_loopfilter_filters.c
index bf97589..fc7fbc4 100644
--- a/vp9/common/vp9_loopfilter_filters.c
+++ b/vp9/common/vp9_loopfilter_filters.c
@@ -8,15 +8,13 @@
* be found in the AUTHORS file in the root of the source tree.
*/
-#include <stdlib.h>
#include "vpx_config.h"
+#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_loopfilter.h"
#include "vp9/common/vp9_onyxc_int.h"
static INLINE int8_t signed_char_clamp(int t) {
- t = (t < -128 ? -128 : t);
- t = (t > 127 ? 127 : t);
- return (int8_t) t;
+ return (int8_t)clamp(t, -128, 127);
}
// should we apply any filter at all: 11111111 yes, 00000000 no
@@ -36,7 +34,7 @@
return ~mask;
}
-// is there high variance internal edge: 11111111 yes, 00000000 no
+// is there high edge variance internal edge: 11111111 yes, 00000000 no
static INLINE int8_t hevmask(uint8_t thresh, uint8_t p1, uint8_t p0,
uint8_t q0, uint8_t q1) {
int8_t hev = 0;
@@ -68,12 +66,9 @@
*oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
*op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
- filter = filter1;
// outer tap adjustments
- filter += 1;
- filter >>= 1;
- filter &= ~hev;
+ filter = ((filter1 + 1) >> 1) & ~hev;
*oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
*op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
@@ -84,23 +79,19 @@
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
// loop filter designed to work using chars so that we can make maximum use
// of 8 bit simd instructions.
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4 * p], s[-3 * p], s[-2 * p], s[-1 * p],
- s[0 * p], s[1 * p], s[2 * p], s[3 * p]);
-
- // high edge variance
- const int8_t hev = hevmask(thresh[0],
- s[-2 * p], s[-1 * p], s[0 * p], s[1 * p]);
-
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(*thresh, p1, p0, q0, q1);
filter(mask, hev, s - 2 * p, s - 1 * p, s, s + 1 * p);
-
++s;
- } while (++i < count * 8);
+ }
}
void vp9_loop_filter_vertical_edge_c(uint8_t *s, int pitch,
@@ -108,21 +99,21 @@
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
// loop filter designed to work using chars so that we can make maximum use
// of 8 bit simd instructions.
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4], s[-3], s[-2], s[-1],
- s[0], s[1], s[2], s[3]);
-
- // high edge variance
- const int8_t hev = hevmask(thresh[0], s[-2], s[-1], s[0], s[1]);
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(*thresh, p1, p0, q0, q1);
filter(mask, hev, s - 2, s - 1, s, s + 1);
s += pitch;
- } while (++i < count * 8);
+ }
}
+
static INLINE int8_t flatmask4(uint8_t thresh,
uint8_t p3, uint8_t p2,
uint8_t p1, uint8_t p0,
@@ -157,14 +148,8 @@
uint8_t *oq2, uint8_t *oq3) {
// use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line
if (flat && mask) {
- const uint8_t p3 = *op3;
- const uint8_t p2 = *op2;
- const uint8_t p1 = *op1;
- const uint8_t p0 = *op0;
- const uint8_t q0 = *oq0;
- const uint8_t q1 = *oq1;
- const uint8_t q2 = *oq2;
- const uint8_t q3 = *oq3;
+ const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+ const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
*op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3);
*op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3);
@@ -173,33 +158,7 @@
*oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3);
*oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3);
} else {
- int8_t filter1, filter2;
-
- const int8_t ps1 = (int8_t) *op1 ^ 0x80;
- const int8_t ps0 = (int8_t) *op0 ^ 0x80;
- const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
- const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
-
- // add outer taps if we have high edge variance
- int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
-
- // inner taps
- filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
-
- filter1 = signed_char_clamp(filter + 4) >> 3;
- filter2 = signed_char_clamp(filter + 3) >> 3;
-
- *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
- *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
- filter = filter1;
-
- // outer tap adjustments
- filter += 1;
- filter >>= 1;
- filter &= ~hev;
-
- *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
- *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+ filter(mask, hev, op1, op0, oq0, oq1);
}
}
@@ -208,28 +167,23 @@
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
// loop filter designed to work using chars so that we can make maximum use
// of 8 bit simd instructions.
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4 * p], s[-3 * p], s[-2 * p], s[-1 * p],
- s[ 0 * p], s[ 1 * p], s[ 2 * p], s[ 3 * p]);
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
- const int8_t hev = hevmask(thresh[0],
- s[-2 * p], s[-1 * p], s[0 * p], s[1 * p]);
-
- const int8_t flat = flatmask4(1,
- s[-4 * p], s[-3 * p], s[-2 * p], s[-1 * p],
- s[ 0 * p], s[ 1 * p], s[ 2 * p], s[ 3 * p]);
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(*thresh, p1, p0, q0, q1);
+ const int8_t flat = flatmask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
mbfilter(mask, hev, flat,
s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
s, s + 1 * p, s + 2 * p, s + 3 * p);
-
++s;
- } while (++i < count * 8);
-
+ }
}
void vp9_mbloop_filter_vertical_edge_c(uint8_t *s, int pitch,
@@ -237,49 +191,19 @@
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4], s[-3], s[-2], s[-1],
- s[0], s[1], s[2], s[3]);
-
- const int8_t hev = hevmask(thresh[0], s[-2], s[-1], s[0], s[1]);
- const int8_t flat = flatmask4(1, s[-4], s[-3], s[-2], s[-1],
- s[ 0], s[ 1], s[ 2], s[ 3]);
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(thresh[0], p1, p0, q0, q1);
+ const int8_t flat = flatmask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
mbfilter(mask, hev, flat, s - 4, s - 3, s - 2, s - 1,
s, s + 1, s + 2, s + 3);
s += pitch;
- } while (++i < count * 8);
-
-}
-
-// should we apply any filter at all: 11111111 yes, 00000000 no
-static INLINE int8_t simple_filter_mask(uint8_t blimit,
- uint8_t p1, uint8_t p0,
- uint8_t q0, uint8_t q1) {
- return (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 <= blimit) * -1;
-}
-
-static INLINE void simple_filter(int8_t mask,
- uint8_t *op1, uint8_t *op0,
- uint8_t *oq0, uint8_t *oq1) {
- int8_t filter1, filter2;
- const int8_t p1 = (int8_t) *op1 ^ 0x80;
- const int8_t p0 = (int8_t) *op0 ^ 0x80;
- const int8_t q0 = (int8_t) *oq0 ^ 0x80;
- const int8_t q1 = (int8_t) *oq1 ^ 0x80;
-
- int8_t filter = signed_char_clamp(p1 - q1);
- filter = signed_char_clamp(filter + 3 * (q0 - p0));
- filter &= mask;
-
- // save bottom 3 bits so that we round one side +4 and the other +3
- filter1 = signed_char_clamp(filter + 4) >> 3;
- *oq0 = signed_char_clamp(q0 - filter1) ^ 0x80;
-
- filter2 = signed_char_clamp(filter + 3) >> 3;
- *op0 = signed_char_clamp(p0 + filter2) ^ 0x80;
+ }
}
/* Vertical MB Filtering */
@@ -395,22 +319,11 @@
uint8_t *oq7) {
// use a 15 tap filter [1,1,1,1,1,1,1,2,1,1,1,1,1,1,1] for flat line
if (flat2 && flat && mask) {
- const uint8_t p7 = *op7;
- const uint8_t p6 = *op6;
- const uint8_t p5 = *op5;
- const uint8_t p4 = *op4;
- const uint8_t p3 = *op3;
- const uint8_t p2 = *op2;
- const uint8_t p1 = *op1;
- const uint8_t p0 = *op0;
- const uint8_t q0 = *oq0;
- const uint8_t q1 = *oq1;
- const uint8_t q2 = *oq2;
- const uint8_t q3 = *oq3;
- const uint8_t q4 = *oq4;
- const uint8_t q5 = *oq5;
- const uint8_t q6 = *oq6;
- const uint8_t q7 = *oq7;
+ const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
+ p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+
+ const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
+ q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
*op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
q0, 4);
@@ -440,49 +353,8 @@
q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
*oq6 = ROUND_POWER_OF_TWO(p0 +
q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
- } else if (flat && mask) {
- const uint8_t p3 = *op3;
- const uint8_t p2 = *op2;
- const uint8_t p1 = *op1;
- const uint8_t p0 = *op0;
- const uint8_t q0 = *oq0;
- const uint8_t q1 = *oq1;
- const uint8_t q2 = *oq2;
- const uint8_t q3 = *oq3;
-
- *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3);
- *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3);
- *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3);
- *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3);
- *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3);
- *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3);
} else {
- int8_t filter1, filter2;
-
- const int8_t ps1 = (int8_t) * op1 ^ 0x80;
- const int8_t ps0 = (int8_t) * op0 ^ 0x80;
- const int8_t qs0 = (int8_t) * oq0 ^ 0x80;
- const int8_t qs1 = (int8_t) * oq1 ^ 0x80;
-
- // add outer taps if we have high edge variance
- int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
-
- // inner taps
- filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
- filter1 = signed_char_clamp(filter + 4) >> 3;
- filter2 = signed_char_clamp(filter + 3) >> 3;
-
- *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
- *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
- filter = filter1;
-
- // outer tap adjustments
- filter += 1;
- filter >>= 1;
- filter &= ~hev;
-
- *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
- *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+ mbfilter(mask, hev, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
}
}
@@ -491,25 +363,20 @@
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
// loop filter designed to work using chars so that we can make maximum use
// of 8 bit simd instructions.
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4 * p], s[-3 * p], s[-2 * p], s[-1 * p],
- s[ 0 * p], s[ 1 * p], s[ 2 * p], s[ 3 * p]);
-
- const int8_t hev = hevmask(thresh[0],
- s[-2 * p], s[-1 * p], s[0 * p], s[1 * p]);
-
- const int8_t flat = flatmask4(1,
- s[-4 * p], s[-3 * p], s[-2 * p], s[-1 * p],
- s[ 0 * p], s[ 1 * p], s[ 2 * p], s[ 3 * p]);
-
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(*thresh, p1, p0, q0, q1);
+ const int8_t flat = flatmask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
const int8_t flat2 = flatmask5(1,
- s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], s[-1 * p],
- s[ 0 * p], s[ 4 * p], s[ 5 * p], s[ 6 * p], s[ 7 * p]);
+ s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+ q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
wide_mbfilter(mask, hev, flat, flat2,
s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
@@ -518,33 +385,31 @@
s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
++s;
- } while (++i < count * 8);
+ }
}
+
void vp9_mb_lpf_vertical_edge_w(uint8_t *s, int p,
const uint8_t *blimit,
const uint8_t *limit,
const uint8_t *thresh,
int count) {
- int i = 0;
+ int i;
- do {
- const int8_t mask = filter_mask(limit[0], blimit[0],
- s[-4], s[-3], s[-2], s[-1],
- s[0], s[1], s[2], s[3]);
-
- const int8_t hev = hevmask(thresh[0], s[-2], s[-1], s[0], s[1]);
- const int8_t flat = flatmask4(1, s[-4], s[-3], s[-2], s[-1],
- s[ 0], s[ 1], s[ 2], s[ 3]);
- const int8_t flat2 = flatmask5(1, s[-8], s[-7], s[-6], s[-5], s[-1],
- s[ 0], s[ 4], s[ 5], s[ 6], s[ 7]);
+ for (i = 0; i < 8 * count; ++i) {
+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t hev = hevmask(*thresh, p1, p0, q0, q1);
+ const int8_t flat = flatmask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat2 = flatmask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+ q0, s[4], s[5], s[6], s[7]);
wide_mbfilter(mask, hev, flat, flat2,
- s - 8, s - 7, s - 6, s - 5,
- s - 4, s - 3, s - 2, s - 1,
- s, s + 1, s + 2, s + 3,
- s + 4, s + 5, s + 6, s + 7);
+ s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+ s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
s += p;
- } while (++i < count * 8);
+ }
}
void vp9_lpf_mbv_w_c(uint8_t *y, uint8_t *u, uint8_t *v,
diff --git a/vp9/common/vp9_postproc.c b/vp9/common/vp9_postproc.c
index f81690a..8001adb 100644
--- a/vp9/common/vp9_postproc.c
+++ b/vp9/common/vp9_postproc.c
@@ -132,14 +132,15 @@
/****************************************************************************
*/
-void vp9_post_proc_down_and_across_c(uint8_t *src_ptr,
+void vp9_post_proc_down_and_across_c(const uint8_t *src_ptr,
uint8_t *dst_ptr,
int src_pixels_per_line,
int dst_pixels_per_line,
int rows,
int cols,
int flimit) {
- uint8_t *p_src, *p_dst;
+ uint8_t const *p_src;
+ uint8_t *p_dst;
int row;
int col;
int i;
@@ -313,51 +314,52 @@
source->uv_height, source->uv_width, ppl);
}
-void vp9_deblock(YV12_BUFFER_CONFIG *source,
- YV12_BUFFER_CONFIG *post,
- int q,
- int low_var_thresh,
- int flag) {
- double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
- int ppl = (int)(level + .5);
- (void) low_var_thresh;
- (void) flag;
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+ int q) {
+ const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+ + 0.0065 + 0.5);
+ int i;
- vp9_post_proc_down_and_across(source->y_buffer, post->y_buffer,
- source->y_stride, post->y_stride,
- source->y_height, source->y_width, ppl);
+ const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+ const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+ const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+ const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
- vp9_post_proc_down_and_across(source->u_buffer, post->u_buffer,
- source->uv_stride, post->uv_stride,
- source->uv_height, source->uv_width, ppl);
+ uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+ const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
- vp9_post_proc_down_and_across(source->v_buffer, post->v_buffer,
- source->uv_stride, post->uv_stride,
- source->uv_height, source->uv_width, ppl);
+ for (i = 0; i < MAX_MB_PLANE; ++i)
+ vp9_post_proc_down_and_across(srcs[i], dsts[i],
+ src_strides[i], dst_strides[i],
+ src_heights[i], src_widths[i], ppl);
}
-void vp9_denoise(YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *post,
- int q, int low_var_thresh, int flag) {
- double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
- int ppl = (int)(level + .5);
- (void) post;
- (void) low_var_thresh;
- (void) flag;
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+ int q) {
+ const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+ + 0.0065 + 0.5);
+ int i;
- vp9_post_proc_down_and_across(src->y_buffer + 2 * src->y_stride + 2,
- src->y_buffer + 2 * src->y_stride + 2,
- src->y_stride, src->y_stride, src->y_height - 4,
- src->y_width - 4, ppl);
+ const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+ const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+ const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+ const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
- vp9_post_proc_down_and_across(src->u_buffer + 2 * src->uv_stride + 2,
- src->u_buffer + 2 * src->uv_stride + 2,
- src->uv_stride, src->uv_stride,
- src->uv_height - 4, src->uv_width - 4, ppl);
+ uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+ const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
- vp9_post_proc_down_and_across(src->v_buffer + 2 * src->uv_stride + 2,
- src->v_buffer + 2 * src->uv_stride + 2,
- src->uv_stride, src->uv_stride,
- src->uv_height - 4, src->uv_width - 4, ppl);
+ for (i = 0; i < MAX_MB_PLANE; ++i) {
+ const int src_stride = src_strides[i];
+ const uint8_t *const src = srcs[i] + 2 * src_stride + 2;
+ const int src_width = src_widths[i] - 4;
+ const int src_height = src_heights[i] - 4;
+
+ const int dst_stride = dst_strides[i];
+ uint8_t *const dst = dsts[i] + 2 * dst_stride + 2;
+
+ vp9_post_proc_down_and_across(src, dst, src_stride, dst_stride,
+ src_height, src_width, ppl);
+ }
}
double vp9_gaussian(double sigma, double mu, double x) {
@@ -642,7 +644,7 @@
deblock_and_de_macro_block(oci->frame_to_show, &oci->post_proc_buffer,
q + (deblock_level - 5) * 10, 1, 0);
} else if (flags & VP9D_DEBLOCK) {
- vp9_deblock(oci->frame_to_show, &oci->post_proc_buffer, q, 1, 0);
+ vp9_deblock(oci->frame_to_show, &oci->post_proc_buffer, q);
} else {
vp8_yv12_copy_frame(oci->frame_to_show, &oci->post_proc_buffer);
}
diff --git a/vp9/common/vp9_postproc.h b/vp9/common/vp9_postproc.h
index c2f556e..2c0d333 100644
--- a/vp9/common/vp9_postproc.h
+++ b/vp9/common/vp9_postproc.h
@@ -29,10 +29,8 @@
int vp9_post_proc_frame(struct VP9Common *oci, YV12_BUFFER_CONFIG *dest,
vp9_ppflags_t *flags);
-void vp9_denoise(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *post,
- int q, int low_var_thresh, int flag);
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
-void vp9_deblock(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *post,
- int q, int low_var_thresh, int flag);
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
#endif // VP9_COMMON_VP9_POSTPROC_H_
diff --git a/vp9/common/vp9_reconinter.c b/vp9/common/vp9_reconinter.c
index 3668fcd..e7303f1 100644
--- a/vp9/common/vp9_reconinter.c
+++ b/vp9/common/vp9_reconinter.c
@@ -17,6 +17,78 @@
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
+static int scale_value_x_with_scaling(int val,
+ const struct scale_factors *scale) {
+ return val * scale->x_num / scale->x_den;
+}
+
+static int scale_value_y_with_scaling(int val,
+ const struct scale_factors *scale) {
+ return val * scale->y_num / scale->y_den;
+}
+
+static int unscaled_value(int val, const struct scale_factors *scale) {
+ (void) scale;
+ return val;
+}
+
+static int_mv32 mv_q3_to_q4_with_scaling(const int_mv *src_mv,
+ const struct scale_factors *scale) {
+ // returns mv * scale + offset
+ int_mv32 result;
+ const int32_t mv_row_q4 = src_mv->as_mv.row << 1;
+ const int32_t mv_col_q4 = src_mv->as_mv.col << 1;
+
+ /* TODO(jkoleszar): make fixed point, or as a second multiply? */
+ result.as_mv.row = mv_row_q4 * scale->y_num / scale->y_den
+ + scale->y_offset_q4;
+ result.as_mv.col = mv_col_q4 * scale->x_num / scale->x_den
+ + scale->x_offset_q4;
+ return result;
+}
+
+static int_mv32 mv_q3_to_q4_without_scaling(const int_mv *src_mv,
+ const struct scale_factors *scale) {
+ // returns mv * scale + offset
+ int_mv32 result;
+
+ result.as_mv.row = src_mv->as_mv.row << 1;
+ result.as_mv.col = src_mv->as_mv.col << 1;
+ return result;
+}
+
+static int32_t mv_component_q4_with_scaling(int mv_q4, int num, int den,
+ int offset_q4) {
+ // returns the scaled and offset value of the mv component.
+
+ /* TODO(jkoleszar): make fixed point, or as a second multiply? */
+ return mv_q4 * num / den + offset_q4;
+}
+
+static int32_t mv_component_q4_without_scaling(int mv_q4, int num, int den,
+ int offset_q4) {
+ // returns the scaled and offset value of the mv component.
+ (void)num;
+ (void)den;
+ (void)offset_q4;
+ return mv_q4;
+}
+
+static void set_offsets_with_scaling(struct scale_factors *scale,
+ int row, int col) {
+ const int x_q4 = 16 * col;
+ const int y_q4 = 16 * row;
+
+ scale->x_offset_q4 = (x_q4 * scale->x_num / scale->x_den) & 0xf;
+ scale->y_offset_q4 = (y_q4 * scale->y_num / scale->y_den) & 0xf;
+}
+
+static void set_offsets_without_scaling(struct scale_factors *scale,
+ int row, int col) {
+ scale->x_offset_q4 = 0;
+ scale->y_offset_q4 = 0;
+}
+
void vp9_setup_scale_factors_for_frame(struct scale_factors *scale,
int other_w, int other_h,
int this_w, int this_h) {
@@ -34,18 +106,14 @@
scale->scale_value_x = unscaled_value;
scale->scale_value_y = unscaled_value;
scale->set_scaled_offsets = set_offsets_without_scaling;
- scale->scale_motion_vector_q3_to_q4 =
- motion_vector_q3_to_q4_without_scaling;
- scale->scale_motion_vector_component_q4 =
- motion_vector_component_q4_without_scaling;
+ scale->scale_motion_vector_q3_to_q4 = mv_q3_to_q4_without_scaling;
+ scale->scale_motion_vector_component_q4 = mv_component_q4_without_scaling;
} else {
scale->scale_value_x = scale_value_x_with_scaling;
scale->scale_value_y = scale_value_y_with_scaling;
scale->set_scaled_offsets = set_offsets_with_scaling;
- scale->scale_motion_vector_q3_to_q4 =
- motion_vector_q3_to_q4_with_scaling;
- scale->scale_motion_vector_component_q4 =
- motion_vector_component_q4_with_scaling;
+ scale->scale_motion_vector_q3_to_q4 = mv_q3_to_q4_with_scaling;
+ scale->scale_motion_vector_component_q4 = mv_component_q4_with_scaling;
}
// TODO(agrange): Investigate the best choice of functions to use here
@@ -424,3 +492,18 @@
vp9_build_inter_predictors_sbuv(xd, mb_row, mb_col,
BLOCK_SIZE_MB16X16);
}
+
+// TODO(dkovalev: find better place for this function)
+void vp9_setup_scale_factors(VP9_COMMON *cm, int i) {
+ const int ref = cm->active_ref_idx[i];
+ struct scale_factors *const sf = &cm->active_ref_scale[i];
+ if (ref >= NUM_YV12_BUFFERS) {
+ memset(sf, 0, sizeof(*sf));
+ } else {
+ YV12_BUFFER_CONFIG *const fb = &cm->yv12_fb[ref];
+ vp9_setup_scale_factors_for_frame(sf,
+ fb->y_crop_width, fb->y_crop_height,
+ cm->width, cm->height);
+ }
+}
+
diff --git a/vp9/common/vp9_reconinter.h b/vp9/common/vp9_reconinter.h
index faf018c..8f76195 100644
--- a/vp9/common/vp9_reconinter.h
+++ b/vp9/common/vp9_reconinter.h
@@ -52,21 +52,6 @@
int w, int h, int do_avg,
const struct subpix_fn_table *subpix);
-static int scale_value_x_with_scaling(int val,
- const struct scale_factors *scale) {
- return val * scale->x_num / scale->x_den;
-}
-
-static int scale_value_y_with_scaling(int val,
- const struct scale_factors *scale) {
- return val * scale->y_num / scale->y_den;
-}
-
-static int unscaled_value(int val, const struct scale_factors *scale) {
- (void) scale;
- return val;
-}
-
static int scaled_buffer_offset(int x_offset, int y_offset, int stride,
const struct scale_factors *scale) {
const int x = scale ? scale->scale_value_x(x_offset, scale) : x_offset;
@@ -137,66 +122,6 @@
xd->scale_factor_uv[1] = xd->scale_factor[1];
}
-static void set_offsets_with_scaling(struct scale_factors *scale,
- int row, int col) {
- const int x_q4 = 16 * col;
- const int y_q4 = 16 * row;
+void vp9_setup_scale_factors(VP9_COMMON *cm, int i);
- scale->x_offset_q4 = (x_q4 * scale->x_num / scale->x_den) & 0xf;
- scale->y_offset_q4 = (y_q4 * scale->y_num / scale->y_den) & 0xf;
-}
-
-static void set_offsets_without_scaling(struct scale_factors *scale,
- int row, int col) {
- scale->x_offset_q4 = 0;
- scale->y_offset_q4 = 0;
-}
-
-static int_mv32 motion_vector_q3_to_q4_with_scaling(
- const int_mv *src_mv,
- const struct scale_factors *scale) {
- // returns mv * scale + offset
- int_mv32 result;
- const int32_t mv_row_q4 = src_mv->as_mv.row << 1;
- const int32_t mv_col_q4 = src_mv->as_mv.col << 1;
-
- /* TODO(jkoleszar): make fixed point, or as a second multiply? */
- result.as_mv.row = mv_row_q4 * scale->y_num / scale->y_den
- + scale->y_offset_q4;
- result.as_mv.col = mv_col_q4 * scale->x_num / scale->x_den
- + scale->x_offset_q4;
- return result;
-}
-
-static int_mv32 motion_vector_q3_to_q4_without_scaling(
- const int_mv *src_mv,
- const struct scale_factors *scale) {
- // returns mv * scale + offset
- int_mv32 result;
-
- result.as_mv.row = src_mv->as_mv.row << 1;
- result.as_mv.col = src_mv->as_mv.col << 1;
- return result;
-}
-
-static int32_t motion_vector_component_q4_with_scaling(int mv_q4,
- int num,
- int den,
- int offset_q4) {
- // returns the scaled and offset value of the mv component.
-
- /* TODO(jkoleszar): make fixed point, or as a second multiply? */
- return mv_q4 * num / den + offset_q4;
-}
-
-static int32_t motion_vector_component_q4_without_scaling(int mv_q4,
- int num,
- int den,
- int offset_q4) {
- // returns the scaled and offset value of the mv component.
- (void)num;
- (void)den;
- (void)offset_q4;
- return mv_q4;
-}
#endif // VP9_COMMON_VP9_RECONINTER_H_
diff --git a/vp9/common/vp9_reconintra.c b/vp9/common/vp9_reconintra.c
index 9e580c7..ea4805f 100644
--- a/vp9/common/vp9_reconintra.c
+++ b/vp9/common/vp9_reconintra.c
@@ -378,19 +378,21 @@
void vp9_build_intra_predictors_sbuv_s(MACROBLOCKD *xd,
BLOCK_SIZE_TYPE bsize) {
- const int bwl = b_width_log2(bsize), bw = 2 << bwl;
- const int bhl = b_height_log2(bsize), bh = 2 << bhl;
+ int p;
- vp9_build_intra_predictors(xd->plane[1].dst.buf, xd->plane[1].dst.stride,
- xd->plane[1].dst.buf, xd->plane[1].dst.stride,
- xd->mode_info_context->mbmi.uv_mode,
- bw, bh, xd->up_available,
- xd->left_available, 0 /*xd->right_available*/);
- vp9_build_intra_predictors(xd->plane[2].dst.buf, xd->plane[1].dst.stride,
- xd->plane[2].dst.buf, xd->plane[1].dst.stride,
- xd->mode_info_context->mbmi.uv_mode,
- bw, bh, xd->up_available,
- xd->left_available, 0 /*xd->right_available*/);
+ for (p = 1; p < MAX_MB_PLANE; p++) {
+ const struct macroblockd_plane* const pd = &xd->plane[p];
+ const int bwl = b_width_log2(bsize) - pd->subsampling_x;
+ const int bw = 4 << bwl;
+ const int bhl = b_height_log2(bsize) - pd->subsampling_y;
+ const int bh = 4 << bhl;
+
+ vp9_build_intra_predictors(pd->dst.buf, pd->dst.stride,
+ pd->dst.buf, pd->dst.stride,
+ xd->mode_info_context->mbmi.uv_mode,
+ bw, bh, xd->up_available,
+ xd->left_available, 0 /*xd->right_available*/);
+ }
}
void vp9_intra4x4_predict(MACROBLOCKD *xd,
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 02d3253..b55cc74 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -91,12 +91,6 @@
prototype void vp9_add_residual_8x8 "const int16_t *diff, uint8_t *dest, int stride"
specialize vp9_add_residual_8x8 sse2
-prototype void vp9_add_residual_16x16 "const int16_t *diff, uint8_t *dest, int stride"
-specialize vp9_add_residual_16x16 sse2
-
-prototype void vp9_add_residual_32x32 "const int16_t *diff, uint8_t *dest, int stride"
-specialize vp9_add_residual_32x32 sse2
-
prototype void vp9_add_constant_residual_8x8 "const int16_t diff, uint8_t *dest, int stride"
specialize vp9_add_constant_residual_8x8 sse2
@@ -146,7 +140,7 @@
specialize vp9_mbpost_proc_across_ip sse2
vp9_mbpost_proc_across_ip_sse2=vp9_mbpost_proc_across_ip_xmm
-prototype void vp9_post_proc_down_and_across "uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit"
+prototype void vp9_post_proc_down_and_across "const uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit"
specialize vp9_post_proc_down_and_across mmx sse2
vp9_post_proc_down_and_across_sse2=vp9_post_proc_down_and_across_xmm
@@ -203,24 +197,23 @@
prototype void vp9_short_idct1_8x8 "int16_t *input, int16_t *output"
specialize vp9_short_idct1_8x8
-prototype void vp9_short_idct16x16 "int16_t *input, int16_t *output, int pitch"
-specialize vp9_short_idct16x16 sse2
+prototype void vp9_short_idct16x16_add "int16_t *input, uint8_t *dest, int dest_stride"
+specialize vp9_short_idct16x16_add sse2
-prototype void vp9_short_idct10_16x16 "int16_t *input, int16_t *output, int pitch"
-specialize vp9_short_idct10_16x16 sse2
+prototype void vp9_short_idct10_16x16_add "int16_t *input, uint8_t *dest, int dest_stride"
+specialize vp9_short_idct10_16x16_add sse2
prototype void vp9_short_idct1_16x16 "int16_t *input, int16_t *output"
specialize vp9_short_idct1_16x16
-
-prototype void vp9_short_idct32x32 "int16_t *input, int16_t *output, int pitch"
-specialize vp9_short_idct32x32 sse2
+prototype void vp9_short_idct32x32_add "int16_t *input, uint8_t *dest, int dest_stride"
+specialize vp9_short_idct32x32_add sse2
prototype void vp9_short_idct1_32x32 "int16_t *input, int16_t *output"
specialize vp9_short_idct1_32x32
-prototype void vp9_short_idct10_32x32 "int16_t *input, int16_t *output, int pitch"
-specialize vp9_short_idct10_32x32
+prototype void vp9_short_idct10_32x32_add "int16_t *input, uint8_t *dest, int dest_stride"
+specialize vp9_short_idct10_32x32_add
prototype void vp9_short_iht8x8 "int16_t *input, int16_t *output, int pitch, int tx_type"
specialize vp9_short_iht8x8
@@ -228,8 +221,8 @@
prototype void vp9_short_iht4x4 "int16_t *input, int16_t *output, int pitch, int tx_type"
specialize vp9_short_iht4x4
-prototype void vp9_short_iht16x16 "int16_t *input, int16_t *output, int pitch, int tx_type"
-specialize vp9_short_iht16x16
+prototype void vp9_short_iht16x16_add "int16_t *input, uint8_t *output, int pitch, int tx_type"
+specialize vp9_short_iht16x16_add
prototype void vp9_idct4_1d "int16_t *input, int16_t *output"
specialize vp9_idct4_1d sse2
@@ -374,6 +367,19 @@
prototype unsigned int vp9_sub_pixel_avg_variance8x8 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred"
specialize vp9_sub_pixel_avg_variance8x8
+# TODO(jingning): need to convert 8x4/4x8 functions into mmx/sse form
+prototype unsigned int vp9_sub_pixel_variance8x4 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
+specialize vp9_sub_pixel_variance8x4
+
+prototype unsigned int vp9_sub_pixel_avg_variance8x4 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred"
+specialize vp9_sub_pixel_avg_variance8x4
+
+prototype unsigned int vp9_sub_pixel_variance4x8 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
+specialize vp9_sub_pixel_variance4x8
+
+prototype unsigned int vp9_sub_pixel_avg_variance4x8 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred"
+specialize vp9_sub_pixel_avg_variance4x8
+
prototype unsigned int vp9_sub_pixel_variance4x4 "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_sub_pixel_variance4x4 sse2 mmx
vp9_sub_pixel_variance4x4_sse2=vp9_sub_pixel_variance4x4_wmt
@@ -411,6 +417,13 @@
prototype unsigned int vp9_sad8x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
specialize vp9_sad8x8 mmx sse2
+# TODO(jingning): need to covert these functions into mmx/sse2 form
+prototype unsigned int vp9_sad8x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
+specialize vp9_sad8x4
+
+prototype unsigned int vp9_sad4x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
+specialize vp9_sad4x8
+
prototype unsigned int vp9_sad4x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad"
specialize vp9_sad4x4 mmx sse
@@ -516,6 +529,13 @@
prototype void vp9_sad8x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad8x8x4d sse2
+# TODO(jingning): need to convert these 4x8/8x4 functions into sse2 form
+prototype void vp9_sad8x4x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
+specialize vp9_sad8x4x4d
+
+prototype void vp9_sad4x8x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
+specialize vp9_sad4x8x4d
+
prototype void vp9_sad4x4x4d "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array"
specialize vp9_sad4x4x4d sse
prototype unsigned int vp9_sub_pixel_mse16x16 "const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse"
diff --git a/vp9/common/vp9_seg_common.c b/vp9/common/vp9_seg_common.c
index 02b7053..e5511a1 100644
--- a/vp9/common/vp9_seg_common.c
+++ b/vp9/common/vp9_seg_common.c
@@ -89,7 +89,7 @@
#if CONFIG_IMPLICIT_SEGMENTATION
// This function defines an implicit segmentation for the next frame based
// on predcition and transform decisions in the current frame.
-// For test purposes at the moment only TX size is used.
+// For test purposes at the moment it uses ref frame and prediction size
void vp9_implicit_segment_map_update(VP9_COMMON * cm) {
int row, col;
MODE_INFO *mi, *mi_ptr = cm->mi;
@@ -97,9 +97,66 @@
for (row = 0; row < cm->mb_rows; row++) {
mi = mi_ptr;
- // Experimental use of tx size to define implicit segmentation
+
for (col = 0; col < cm->mb_cols; ++col, ++mi) {
- map_ptr[col] = 1 + mi->mbmi.txfm_size;
+ // Inter prediction
+ if (mi->mbmi.ref_frame != INTRA_FRAME) {
+ // Zero motion and prediction block size >= 16
+ if ((mi->mbmi.sb_type >= BLOCK_SIZE_MB16X16) &&
+ (mi->mbmi.mv[0].as_int == 0))
+ map_ptr[col] = 1;
+ else if (mi->mbmi.sb_type >= BLOCK_SIZE_SB32X32)
+ map_ptr[col] = 2;
+ else if (mi->mbmi.sb_type >= BLOCK_SIZE_MB16X16)
+ map_ptr[col] = 3;
+ else
+ map_ptr[col] = 6;
+
+ // Intra prediction
+ } else {
+ if (mi->mbmi.sb_type >= BLOCK_SIZE_SB32X32)
+ map_ptr[col] = 4;
+ else if (mi->mbmi.sb_type >= BLOCK_SIZE_MB16X16)
+ map_ptr[col] = 5;
+ else
+ map_ptr[col] = 7;
+ }
+ }
+ mi_ptr += cm->mode_info_stride;
+ map_ptr += cm->mb_cols;
+ }
+}
+
+// This function defines an implicit segmentation for the next frame based
+// on predcition and transform decisions in the current frame.
+// For test purposes at the moment only TX size is used.
+void vp9_implicit_segment_map_update_tx(VP9_COMMON * cm) {
+ int row, col;
+ MODE_INFO *mi, *mi_ptr = cm->mi;
+ unsigned char * map_ptr = cm->last_frame_seg_map;
+
+ for (row = 0; row < cm->mb_rows; row++) {
+ mi = mi_ptr;
+ for (col = 0; col < cm->mb_cols; ++col, ++mi) {
+ // Intra modes
+ if (mi->mbmi.ref_frame == INTRA_FRAME) {
+ if (mi->mbmi.txfm_size == TX_4X4)
+ map_ptr[col] = 7;
+ else if (mi->mbmi.txfm_size <= TX_16X16)
+ map_ptr[col] = 5;
+ else
+ map_ptr[col] = 4;
+ } else {
+ // Inter Modes
+ if (mi->mbmi.txfm_size == TX_4X4)
+ map_ptr[col] = 6;
+ else if (mi->mbmi.txfm_size == TX_8X8)
+ map_ptr[col] = 3;
+ else if (mi->mbmi.txfm_size == TX_16X16)
+ map_ptr[col] = 2;
+ else
+ map_ptr[col] = 1;
+ }
}
mi_ptr += cm->mode_info_stride;
map_ptr += cm->mb_cols;
@@ -107,6 +164,7 @@
}
#endif
+
const vp9_tree_index vp9_segment_tree[14] = {
2, 4, 6, 8, 10, 12,
0, -1, -2, -3, -4, -5, -6, -7
diff --git a/vp9/common/x86/vp9_idct_intrin_sse2.c b/vp9/common/x86/vp9_idct_intrin_sse2.c
index dd7e68a..667da33 100644
--- a/vp9/common/x86/vp9_idct_intrin_sse2.c
+++ b/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -752,8 +752,17 @@
stp2_10, stp2_13, stp2_11, stp2_12) \
}
-void vp9_short_idct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
- const int half_pitch = pitch >> 1;
+#define RECON_AND_STORE(dest, in_x) \
+ { \
+ __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+ d0 = _mm_unpacklo_epi8(d0, zero); \
+ in_x = _mm_add_epi16(in_x, d0); \
+ in_x = _mm_packus_epi16(in_x, in_x); \
+ _mm_storel_epi64((__m128i *)(dest), in_x); \
+ dest += stride; \
+ }
+
+void vp9_short_idct16x16_add_sse2(int16_t *input, uint8_t *dest, int stride) {
const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
const __m128i final_rounding = _mm_set1_epi16(1<<5);
const __m128i zero = _mm_setzero_si128();
@@ -938,31 +947,30 @@
in14 = _mm_srai_epi16(in14, 6);
in15 = _mm_srai_epi16(in15, 6);
- // Store results
- _mm_store_si128((__m128i *)output, in0);
- _mm_store_si128((__m128i *)(output + half_pitch * 1), in1);
- _mm_store_si128((__m128i *)(output + half_pitch * 2), in2);
- _mm_store_si128((__m128i *)(output + half_pitch * 3), in3);
- _mm_store_si128((__m128i *)(output + half_pitch * 4), in4);
- _mm_store_si128((__m128i *)(output + half_pitch * 5), in5);
- _mm_store_si128((__m128i *)(output + half_pitch * 6), in6);
- _mm_store_si128((__m128i *)(output + half_pitch * 7), in7);
- _mm_store_si128((__m128i *)(output + half_pitch * 8), in8);
- _mm_store_si128((__m128i *)(output + half_pitch * 9), in9);
- _mm_store_si128((__m128i *)(output + half_pitch * 10), in10);
- _mm_store_si128((__m128i *)(output + half_pitch * 11), in11);
- _mm_store_si128((__m128i *)(output + half_pitch * 12), in12);
- _mm_store_si128((__m128i *)(output + half_pitch * 13), in13);
- _mm_store_si128((__m128i *)(output + half_pitch * 14), in14);
- _mm_store_si128((__m128i *)(output + half_pitch * 15), in15);
+ RECON_AND_STORE(dest, in0);
+ RECON_AND_STORE(dest, in1);
+ RECON_AND_STORE(dest, in2);
+ RECON_AND_STORE(dest, in3);
+ RECON_AND_STORE(dest, in4);
+ RECON_AND_STORE(dest, in5);
+ RECON_AND_STORE(dest, in6);
+ RECON_AND_STORE(dest, in7);
+ RECON_AND_STORE(dest, in8);
+ RECON_AND_STORE(dest, in9);
+ RECON_AND_STORE(dest, in10);
+ RECON_AND_STORE(dest, in11);
+ RECON_AND_STORE(dest, in12);
+ RECON_AND_STORE(dest, in13);
+ RECON_AND_STORE(dest, in14);
+ RECON_AND_STORE(dest, in15);
- output += 8;
+ dest += 8 - (stride * 16);
}
}
}
-void vp9_short_idct10_16x16_sse2(int16_t *input, int16_t *output, int pitch) {
- const int half_pitch = pitch >> 1;
+void vp9_short_idct10_16x16_add_sse2(int16_t *input, uint8_t *dest,
+ int stride) {
const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
const __m128i final_rounding = _mm_set1_epi16(1<<5);
const __m128i zero = _mm_setzero_si128();
@@ -1007,7 +1015,6 @@
stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
__m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
int i;
-
// 1-D idct. Load input data.
in0 = _mm_load_si128((__m128i *)input);
in8 = _mm_load_si128((__m128i *)(input + 8 * 1));
@@ -1298,29 +1305,28 @@
in14 = _mm_srai_epi16(in14, 6);
in15 = _mm_srai_epi16(in15, 6);
- // Store results
- _mm_store_si128((__m128i *)output, in0);
- _mm_store_si128((__m128i *)(output + half_pitch * 1), in1);
- _mm_store_si128((__m128i *)(output + half_pitch * 2), in2);
- _mm_store_si128((__m128i *)(output + half_pitch * 3), in3);
- _mm_store_si128((__m128i *)(output + half_pitch * 4), in4);
- _mm_store_si128((__m128i *)(output + half_pitch * 5), in5);
- _mm_store_si128((__m128i *)(output + half_pitch * 6), in6);
- _mm_store_si128((__m128i *)(output + half_pitch * 7), in7);
- _mm_store_si128((__m128i *)(output + half_pitch * 8), in8);
- _mm_store_si128((__m128i *)(output + half_pitch * 9), in9);
- _mm_store_si128((__m128i *)(output + half_pitch * 10), in10);
- _mm_store_si128((__m128i *)(output + half_pitch * 11), in11);
- _mm_store_si128((__m128i *)(output + half_pitch * 12), in12);
- _mm_store_si128((__m128i *)(output + half_pitch * 13), in13);
- _mm_store_si128((__m128i *)(output + half_pitch * 14), in14);
- _mm_store_si128((__m128i *)(output + half_pitch * 15), in15);
- output += 8;
+ RECON_AND_STORE(dest, in0);
+ RECON_AND_STORE(dest, in1);
+ RECON_AND_STORE(dest, in2);
+ RECON_AND_STORE(dest, in3);
+ RECON_AND_STORE(dest, in4);
+ RECON_AND_STORE(dest, in5);
+ RECON_AND_STORE(dest, in6);
+ RECON_AND_STORE(dest, in7);
+ RECON_AND_STORE(dest, in8);
+ RECON_AND_STORE(dest, in9);
+ RECON_AND_STORE(dest, in10);
+ RECON_AND_STORE(dest, in11);
+ RECON_AND_STORE(dest, in12);
+ RECON_AND_STORE(dest, in13);
+ RECON_AND_STORE(dest, in14);
+ RECON_AND_STORE(dest, in15);
+
+ dest += 8 - (stride * 16);
}
}
-void vp9_short_idct32x32_sse2(int16_t *input, int16_t *output, int pitch) {
- const int half_pitch = pitch >> 1;
+void vp9_short_idct32x32_add_sse2(int16_t *input, uint8_t *dest, int stride) {
const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
const __m128i final_rounding = _mm_set1_epi16(1<<5);
@@ -1832,6 +1838,8 @@
col[i * 32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
col[i * 32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
} else {
+ const __m128i zero = _mm_setzero_si128();
+
// 2_D: Calculate the results and store them to destination.
in0 = _mm_add_epi16(stp1_0, stp1_31);
in1 = _mm_add_epi16(stp1_1, stp1_30);
@@ -1933,41 +1941,40 @@
in30 = _mm_srai_epi16(in30, 6);
in31 = _mm_srai_epi16(in31, 6);
- // Store results
- _mm_store_si128((__m128i *)output, in0);
- _mm_store_si128((__m128i *)(output + half_pitch * 1), in1);
- _mm_store_si128((__m128i *)(output + half_pitch * 2), in2);
- _mm_store_si128((__m128i *)(output + half_pitch * 3), in3);
- _mm_store_si128((__m128i *)(output + half_pitch * 4), in4);
- _mm_store_si128((__m128i *)(output + half_pitch * 5), in5);
- _mm_store_si128((__m128i *)(output + half_pitch * 6), in6);
- _mm_store_si128((__m128i *)(output + half_pitch * 7), in7);
- _mm_store_si128((__m128i *)(output + half_pitch * 8), in8);
- _mm_store_si128((__m128i *)(output + half_pitch * 9), in9);
- _mm_store_si128((__m128i *)(output + half_pitch * 10), in10);
- _mm_store_si128((__m128i *)(output + half_pitch * 11), in11);
- _mm_store_si128((__m128i *)(output + half_pitch * 12), in12);
- _mm_store_si128((__m128i *)(output + half_pitch * 13), in13);
- _mm_store_si128((__m128i *)(output + half_pitch * 14), in14);
- _mm_store_si128((__m128i *)(output + half_pitch * 15), in15);
- _mm_store_si128((__m128i *)(output + half_pitch * 16), in16);
- _mm_store_si128((__m128i *)(output + half_pitch * 17), in17);
- _mm_store_si128((__m128i *)(output + half_pitch * 18), in18);
- _mm_store_si128((__m128i *)(output + half_pitch * 19), in19);
- _mm_store_si128((__m128i *)(output + half_pitch * 20), in20);
- _mm_store_si128((__m128i *)(output + half_pitch * 21), in21);
- _mm_store_si128((__m128i *)(output + half_pitch * 22), in22);
- _mm_store_si128((__m128i *)(output + half_pitch * 23), in23);
- _mm_store_si128((__m128i *)(output + half_pitch * 24), in24);
- _mm_store_si128((__m128i *)(output + half_pitch * 25), in25);
- _mm_store_si128((__m128i *)(output + half_pitch * 26), in26);
- _mm_store_si128((__m128i *)(output + half_pitch * 27), in27);
- _mm_store_si128((__m128i *)(output + half_pitch * 28), in28);
- _mm_store_si128((__m128i *)(output + half_pitch * 29), in29);
- _mm_store_si128((__m128i *)(output + half_pitch * 30), in30);
- _mm_store_si128((__m128i *)(output + half_pitch * 31), in31);
+ RECON_AND_STORE(dest, in0);
+ RECON_AND_STORE(dest, in1);
+ RECON_AND_STORE(dest, in2);
+ RECON_AND_STORE(dest, in3);
+ RECON_AND_STORE(dest, in4);
+ RECON_AND_STORE(dest, in5);
+ RECON_AND_STORE(dest, in6);
+ RECON_AND_STORE(dest, in7);
+ RECON_AND_STORE(dest, in8);
+ RECON_AND_STORE(dest, in9);
+ RECON_AND_STORE(dest, in10);
+ RECON_AND_STORE(dest, in11);
+ RECON_AND_STORE(dest, in12);
+ RECON_AND_STORE(dest, in13);
+ RECON_AND_STORE(dest, in14);
+ RECON_AND_STORE(dest, in15);
+ RECON_AND_STORE(dest, in16);
+ RECON_AND_STORE(dest, in17);
+ RECON_AND_STORE(dest, in18);
+ RECON_AND_STORE(dest, in19);
+ RECON_AND_STORE(dest, in20);
+ RECON_AND_STORE(dest, in21);
+ RECON_AND_STORE(dest, in22);
+ RECON_AND_STORE(dest, in23);
+ RECON_AND_STORE(dest, in24);
+ RECON_AND_STORE(dest, in25);
+ RECON_AND_STORE(dest, in26);
+ RECON_AND_STORE(dest, in27);
+ RECON_AND_STORE(dest, in28);
+ RECON_AND_STORE(dest, in29);
+ RECON_AND_STORE(dest, in30);
+ RECON_AND_STORE(dest, in31);
- output += 8;
+ dest += 8 - (stride * 32);
}
}
}
diff --git a/vp9/decoder/vp9_decodemv.c b/vp9/decoder/vp9_decodemv.c
index 018f92b..3864d3c 100644
--- a/vp9/decoder/vp9_decodemv.c
+++ b/vp9/decoder/vp9_decodemv.c
@@ -119,13 +119,25 @@
m->mbmi.mb_skip_coeff = vp9_read(r, vp9_get_pred_prob(cm, xd, PRED_MBSKIP));
// luma mode
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8)
+ m->mbmi.mode = read_kf_sb_ymode(r,
+ cm->sb_kf_ymode_prob[cm->kf_ymode_probs_index]);
+ else
+ m->mbmi.mode = I4X4_PRED;
+#else
m->mbmi.mode = m->mbmi.sb_type > BLOCK_SIZE_SB8X8 ?
read_kf_sb_ymode(r, cm->sb_kf_ymode_prob[cm->kf_ymode_probs_index]):
read_kf_mb_ymode(r, cm->kf_ymode_prob[cm->kf_ymode_probs_index]);
+#endif
m->mbmi.ref_frame = INTRA_FRAME;
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type < BLOCK_SIZE_SB8X8) {
+#else
if (m->mbmi.mode == I4X4_PRED) {
+#endif
int i;
for (i = 0; i < 4; ++i) {
const B_PREDICTION_MODE a = above_block_mode(m, i, mis);
@@ -139,7 +151,13 @@
m->mbmi.uv_mode = read_uv_mode(r, cm->kf_uv_mode_prob[m->mbmi.mode]);
if (cm->txfm_mode == TX_MODE_SELECT &&
- !m->mbmi.mb_skip_coeff && m->mbmi.mode != I4X4_PRED) {
+ !m->mbmi.mb_skip_coeff &&
+#if CONFIG_AB4X4
+ m->mbmi.sb_type >= BLOCK_SIZE_SB8X8
+#else
+ m->mbmi.mode != I4X4_PRED
+#endif
+ ) {
const int allow_16x16 = m->mbmi.sb_type >= BLOCK_SIZE_MB16X16;
const int allow_32x32 = m->mbmi.sb_type >= BLOCK_SIZE_SB32X32;
m->mbmi.txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
@@ -150,7 +168,13 @@
m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 &&
m->mbmi.mode <= TM_PRED) {
m->mbmi.txfm_size = TX_16X16;
- } else if (cm->txfm_mode >= ALLOW_8X8 && m->mbmi.mode != I4X4_PRED) {
+ } else if (cm->txfm_mode >= ALLOW_8X8 &&
+#if CONFIG_AB4X4
+ m->mbmi.sb_type >= BLOCK_SIZE_SB8X8
+#else
+ m->mbmi.mode != I4X4_PRED
+#endif
+ ) {
m->mbmi.txfm_size = TX_8X8;
} else {
m->mbmi.txfm_size = TX_4X4;
@@ -588,9 +612,16 @@
if (vp9_segfeature_active(xd, mbmi->segment_id, SEG_LVL_SKIP)) {
mbmi->mode = ZEROMV;
} else {
+#if CONFIG_AB4X4
+ if (mbmi->sb_type >= BLOCK_SIZE_SB8X8)
+ mbmi->mode = read_sb_mv_ref(r, mv_ref_p);
+ else
+ mbmi->mode = SPLITMV;
+#else
mbmi->mode = mbmi->sb_type > BLOCK_SIZE_SB8X8 ?
read_sb_mv_ref(r, mv_ref_p)
: read_mv_ref(r, mv_ref_p);
+#endif
vp9_accum_mv_refs(cm, mbmi->mode, mbmi->mb_mode_context[ref_frame]);
}
@@ -790,6 +821,14 @@
// required for left and above block mv
mv0->as_int = 0;
+#if CONFIG_AB4X4
+ if (mbmi->sb_type >= BLOCK_SIZE_SB8X8) {
+ mbmi->mode = read_sb_ymode(r, cm->fc.sb_ymode_prob);
+ cm->fc.sb_ymode_counts[mbmi->mode]++;
+ } else {
+ mbmi->mode = I4X4_PRED;
+ }
+#else
if (mbmi->sb_type > BLOCK_SIZE_SB8X8) {
mbmi->mode = read_sb_ymode(r, cm->fc.sb_ymode_prob);
cm->fc.sb_ymode_counts[mbmi->mode]++;
@@ -797,9 +836,14 @@
mbmi->mode = read_ymode(r, cm->fc.ymode_prob);
cm->fc.ymode_counts[mbmi->mode]++;
}
+#endif
// If MB mode is I4X4_PRED read the block modes
+#if CONFIG_AB4X4
+ if (mbmi->sb_type < BLOCK_SIZE_SB8X8) {
+#else
if (mbmi->mode == I4X4_PRED) {
+#endif
int j = 0;
do {
int m = read_bmode(r, cm->fc.bmode_prob);
@@ -812,9 +856,14 @@
cm->fc.uv_mode_counts[mbmi->mode][mbmi->uv_mode]++;
}
+#if CONFIG_AB4X4
+ if (cm->txfm_mode == TX_MODE_SELECT && mbmi->mb_skip_coeff == 0 &&
+ mbmi->sb_type >= BLOCK_SIZE_SB8X8) {
+#else
if (cm->txfm_mode == TX_MODE_SELECT && mbmi->mb_skip_coeff == 0 &&
((mbmi->ref_frame == INTRA_FRAME && mbmi->mode != I4X4_PRED) ||
(mbmi->ref_frame != INTRA_FRAME && mbmi->mode != SPLITMV))) {
+#endif
const int allow_16x16 = mbmi->sb_type >= BLOCK_SIZE_MB16X16;
const int allow_32x32 = mbmi->sb_type >= BLOCK_SIZE_SB32X32;
mbmi->txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
@@ -822,13 +871,21 @@
cm->txfm_mode >= ALLOW_32X32) {
mbmi->txfm_size = TX_32X32;
} else if (cm->txfm_mode >= ALLOW_16X16 &&
- mbmi->sb_type >= BLOCK_SIZE_MB16X16 &&
- ((mbmi->ref_frame == INTRA_FRAME && mbmi->mode <= TM_PRED) ||
- (mbmi->ref_frame != INTRA_FRAME && mbmi->mode != SPLITMV))) {
+ mbmi->sb_type >= BLOCK_SIZE_MB16X16
+#if !CONFIG_AB4X4
+ && ((mbmi->ref_frame == INTRA_FRAME && mbmi->mode <= TM_PRED) ||
+ (mbmi->ref_frame != INTRA_FRAME && mbmi->mode != SPLITMV))
+#endif
+ ) {
mbmi->txfm_size = TX_16X16;
} else if (cm->txfm_mode >= ALLOW_8X8 &&
+#if CONFIG_AB4X4
+ (mbmi->sb_type >= BLOCK_SIZE_SB8X8))
+#else
(!(mbmi->ref_frame == INTRA_FRAME && mbmi->mode == I4X4_PRED) &&
- !(mbmi->ref_frame != INTRA_FRAME && mbmi->mode == SPLITMV))) {
+ !(mbmi->ref_frame != INTRA_FRAME && mbmi->mode == SPLITMV)))
+#endif
+ {
mbmi->txfm_size = TX_8X8;
} else {
mbmi->txfm_size = TX_4X4;
diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c
index ea5905b..b200e6c 100644
--- a/vp9/decoder/vp9_decodframe.c
+++ b/vp9/decoder/vp9_decodframe.c
@@ -417,10 +417,14 @@
vp9_decode_mb_mode_mv(pbi, xd, mi_row, mi_col, r);
set_refs(pbi, mi_row, mi_col);
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8)
+#else
if (bsize == BLOCK_SIZE_SB8X8 &&
(xd->mode_info_context->mbmi.mode == SPLITMV ||
xd->mode_info_context->mbmi.mode == I4X4_PRED))
- decode_atom(pbi, xd, mi_row, mi_col, r, bsize);
+#endif
+ decode_atom(pbi, xd, mi_row, mi_col, r, BLOCK_SIZE_SB8X8);
else
decode_sb(pbi, xd, mi_row, mi_col, r, bsize);
@@ -439,7 +443,17 @@
if (mi_row >= pc->mi_rows || mi_col >= pc->mi_cols)
return;
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8)
+ if (xd->ab_index != 0)
+ return;
+#endif
+
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8) {
+#else
if (bsize > BLOCK_SIZE_SB8X8) {
+#endif
int pl;
// read the partition information
xd->left_seg_context = pc->left_seg_context + (mi_row & MI_MASK);
@@ -451,6 +465,7 @@
}
subsize = get_subsize(bsize, partition);
+
switch (partition) {
case PARTITION_NONE:
decode_modes_b(pbi, mi_row, mi_col, r, subsize);
@@ -476,8 +491,13 @@
assert(0);
}
// update partition context
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8 &&
+ (bsize == BLOCK_SIZE_SB8X8 || partition != PARTITION_SPLIT)) {
+#else
if (bsize > BLOCK_SIZE_SB8X8 &&
(bsize == BLOCK_SIZE_MB16X16 || partition != PARTITION_SPLIT)) {
+#endif
set_partition_seg_context(pc, xd, mi_row, mi_col);
update_partition_context(xd, subsize, bsize);
}
@@ -969,7 +989,14 @@
pc->clamp_type = (CLAMP_TYPE)vp9_read_bit(&header_bc);
pc->error_resilient_mode = vp9_read_bit(&header_bc);
- xd->lossless = vp9_read_bit(&header_bc);
+ setup_loopfilter(pc, xd, &header_bc);
+
+ setup_quantization(pbi, &header_bc);
+
+ xd->lossless = pc->base_qindex == 0 &&
+ pc->y_dc_delta_q == 0 &&
+ pc->uv_dc_delta_q == 0 &&
+ pc->uv_ac_delta_q == 0;
if (xd->lossless) {
xd->inv_txm4x4_1 = vp9_short_iwalsh4x4_1;
xd->inv_txm4x4 = vp9_short_iwalsh4x4;
@@ -984,10 +1011,6 @@
xd->itxm_add_uv_block = vp9_idct_add_uv_block;
}
- setup_loopfilter(pc, xd, &header_bc);
-
- setup_quantization(pbi, &header_bc);
-
// Determine if the golden frame or ARF buffer should be updated and how.
// For all non key frames the GF and ARF refresh flags and sign bias
// flags must be set explicitly.
@@ -1001,23 +1024,13 @@
// Select active reference frames and calculate scaling factors
for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i) {
const int ref = vp9_read_literal(&header_bc, NUM_REF_FRAMES_LG2);
- const int mapped_ref = pc->ref_frame_map[ref];
- YV12_BUFFER_CONFIG *const fb = &pc->yv12_fb[mapped_ref];
- struct scale_factors *const sf = &pc->active_ref_scale[i];
-
- pc->active_ref_idx[i] = mapped_ref;
- if (mapped_ref >= NUM_YV12_BUFFERS)
- memset(sf, 0, sizeof(*sf));
- else
- vp9_setup_scale_factors_for_frame(sf,
- fb->y_crop_width, fb->y_crop_height,
- pc->width, pc->height);
+ pc->active_ref_idx[i] = pc->ref_frame_map[ref];
+ vp9_setup_scale_factors(pc, i);
}
// Read the sign bias for each reference frame buffer.
- for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i) {
+ for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i)
pc->ref_frame_sign_bias[i + 1] = vp9_read_bit(&header_bc);
- }
xd->allow_high_precision_mv = vp9_read_bit(&header_bc);
pc->mcomp_filter_type = read_mcomp_filter_type(&header_bc);
diff --git a/vp9/decoder/vp9_decodframe.h b/vp9/decoder/vp9_decodframe.h
index 3aaae65..00b6d67 100644
--- a/vp9/decoder/vp9_decodframe.h
+++ b/vp9/decoder/vp9_decodframe.h
@@ -13,7 +13,9 @@
#define VP9_DECODER_VP9_DECODFRAME_H_
struct VP9Common;
+struct VP9Decompressor;
void vp9_init_dequantizer(struct VP9Common *pc);
+int vp9_decode_frame(struct VP9Decompressor *cpi, const uint8_t **p_data_end);
#endif // VP9_DECODER_VP9_DECODFRAME_H_
diff --git a/vp9/decoder/vp9_idct_blk.c b/vp9/decoder/vp9_idct_blk.c
index 3480df2..bc943fa 100644
--- a/vp9/decoder/vp9_idct_blk.c
+++ b/vp9/decoder/vp9_idct_blk.c
@@ -105,14 +105,6 @@
add_residual(diff, dest, stride, 8, 8);
}
-void vp9_add_residual_16x16_c(const int16_t *diff, uint8_t *dest, int stride) {
- add_residual(diff, dest, stride, 16, 16);
-}
-
-void vp9_add_residual_32x32_c(const int16_t *diff, uint8_t *dest, int stride) {
- add_residual(diff, dest, stride, 32, 32);
-}
-
static void add_constant_residual(const int16_t diff, uint8_t *dest, int stride,
int width, int height) {
int r, c;
@@ -264,19 +256,14 @@
if (tx_type == DCT_DCT) {
vp9_idct_add_16x16(input, dest, stride, eob);
} else {
- DECLARE_ALIGNED_ARRAY(16, int16_t, output, 256);
-
if (eob > 0) {
- vp9_short_iht16x16(input, output, 16, tx_type);
+ vp9_short_iht16x16_add(input, dest, stride, tx_type);
vpx_memset(input, 0, 512);
- vp9_add_residual_16x16(output, dest, stride);
}
}
}
void vp9_idct_add_16x16_c(int16_t *input, uint8_t *dest, int stride, int eob) {
- DECLARE_ALIGNED_ARRAY(16, int16_t, output, 256);
-
/* The calculation can be simplified if there are not many non-zero dct
* coefficients. Use eobs to separate different cases. */
if (eob) {
@@ -292,21 +279,15 @@
vp9_add_constant_residual_16x16(out, dest, stride);
#if !CONFIG_SCATTERSCAN
} else if (eob <= 10) {
- // the idct halves ( >> 1) the pitch
- vp9_short_idct10_16x16(input, output, 32);
-
+ vp9_short_idct10_16x16_add(input, dest, stride);
input[0] = input[1] = input[2] = input[3] = 0;
input[16] = input[17] = input[18] = 0;
input[32] = input[33] = 0;
input[48] = 0;
-
- vp9_add_residual_16x16(output, dest, stride);
#endif
} else {
- // the idct halves ( >> 1) the pitch
- vp9_short_idct16x16(input, output, 16 << 1);
+ vp9_short_idct16x16_add(input, dest, stride);
vpx_memset(input, 0, 512);
- vp9_add_residual_16x16(output, dest, stride);
}
}
}
@@ -321,20 +302,16 @@
input[0] = 0;
#if !CONFIG_SCATTERSCAN
} else if (eob <= 10) {
- // the idct halves ( >> 1) the pitch
- vp9_short_idct10_32x32(input, output, 64);
-
+ vp9_short_idct10_32x32_add_c(input, dest, stride);
input[0] = input[1] = input[2] = input[3] = 0;
input[32] = input[33] = input[34] = 0;
input[64] = input[65] = 0;
input[96] = 0;
- vp9_add_residual_32x32(output, dest, stride);
#endif
} else {
- vp9_short_idct32x32(input, output, 64);
+ vp9_short_idct32x32_add(input, dest, stride);
vpx_memset(input, 0, 2048);
- vp9_add_residual_32x32(output, dest, stride);
}
}
}
diff --git a/vp9/decoder/vp9_onyxd_if.c b/vp9/decoder/vp9_onyxd_if.c
index c0ce311..6a55e8f 100644
--- a/vp9/decoder/vp9_onyxd_if.c
+++ b/vp9/decoder/vp9_onyxd_if.c
@@ -352,7 +352,8 @@
vp9_loop_filter_frame(cm, &pbi->mb, cm->filter_level, 0,
cm->dering_enabled);
}
- vp8_yv12_extend_frame_borders(cm->frame_to_show);
+ vp9_extend_frame_borders(cm->frame_to_show,
+ cm->subsampling_x, cm->subsampling_y);
}
#if WRITE_RECON_BUFFER == 1
diff --git a/vp9/decoder/vp9_onyxd_int.h b/vp9/decoder/vp9_onyxd_int.h
index a7d444e..8698570 100644
--- a/vp9/decoder/vp9_onyxd_int.h
+++ b/vp9/decoder/vp9_onyxd_int.h
@@ -41,8 +41,6 @@
int initial_height;
} VP9D_COMP;
-int vp9_decode_frame(VP9D_COMP *cpi, const uint8_t **p_data_end);
-
#if CONFIG_DEBUG
#define CHECK_MEM_ERROR(lval,expr) do {\
diff --git a/vp9/decoder/x86/vp9_dequantize_sse2.c b/vp9/decoder/x86/vp9_dequantize_sse2.c
index 1296b70..796fc12 100644
--- a/vp9/decoder/x86/vp9_dequantize_sse2.c
+++ b/vp9/decoder/x86/vp9_dequantize_sse2.c
@@ -122,124 +122,6 @@
_mm_storel_epi64((__m128i *)(dest + 7 * stride), p6);
}
-void vp9_add_residual_16x16_sse2(const int16_t *diff, uint8_t *dest,
- int stride) {
- const int width = 16;
- int i = 4;
- const __m128i zero = _mm_setzero_si128();
-
- // Diff data
- __m128i d0, d1, d2, d3, d4, d5, d6, d7;
- __m128i p0, p1, p2, p3, p4, p5, p6, p7;
-
- do {
- d0 = _mm_load_si128((const __m128i *)(diff + 0 * width));
- d1 = _mm_load_si128((const __m128i *)(diff + 0 * width + 8));
- d2 = _mm_load_si128((const __m128i *)(diff + 1 * width));
- d3 = _mm_load_si128((const __m128i *)(diff + 1 * width + 8));
- d4 = _mm_load_si128((const __m128i *)(diff + 2 * width));
- d5 = _mm_load_si128((const __m128i *)(diff + 2 * width + 8));
- d6 = _mm_load_si128((const __m128i *)(diff + 3 * width));
- d7 = _mm_load_si128((const __m128i *)(diff + 3 * width + 8));
-
- // Prediction data.
- p1 = _mm_load_si128((const __m128i *)(dest + 0 * stride));
- p3 = _mm_load_si128((const __m128i *)(dest + 1 * stride));
- p5 = _mm_load_si128((const __m128i *)(dest + 2 * stride));
- p7 = _mm_load_si128((const __m128i *)(dest + 3 * stride));
-
- p0 = _mm_unpacklo_epi8(p1, zero);
- p1 = _mm_unpackhi_epi8(p1, zero);
- p2 = _mm_unpacklo_epi8(p3, zero);
- p3 = _mm_unpackhi_epi8(p3, zero);
- p4 = _mm_unpacklo_epi8(p5, zero);
- p5 = _mm_unpackhi_epi8(p5, zero);
- p6 = _mm_unpacklo_epi8(p7, zero);
- p7 = _mm_unpackhi_epi8(p7, zero);
-
- p0 = _mm_add_epi16(p0, d0);
- p1 = _mm_add_epi16(p1, d1);
- p2 = _mm_add_epi16(p2, d2);
- p3 = _mm_add_epi16(p3, d3);
- p4 = _mm_add_epi16(p4, d4);
- p5 = _mm_add_epi16(p5, d5);
- p6 = _mm_add_epi16(p6, d6);
- p7 = _mm_add_epi16(p7, d7);
-
- p0 = _mm_packus_epi16(p0, p1);
- p1 = _mm_packus_epi16(p2, p3);
- p2 = _mm_packus_epi16(p4, p5);
- p3 = _mm_packus_epi16(p6, p7);
-
- _mm_store_si128((__m128i *)(dest + 0 * stride), p0);
- _mm_store_si128((__m128i *)(dest + 1 * stride), p1);
- _mm_store_si128((__m128i *)(dest + 2 * stride), p2);
- _mm_store_si128((__m128i *)(dest + 3 * stride), p3);
-
- diff += 4 * width;
- dest += 4 * stride;
- } while (--i);
-}
-
-void vp9_add_residual_32x32_sse2(const int16_t *diff, uint8_t *dest,
- int stride) {
- const int width = 32;
- int i = 16;
- const __m128i zero = _mm_setzero_si128();
-
- // Diff data
- __m128i d0, d1, d2, d3, d4, d5, d6, d7;
- __m128i p0, p1, p2, p3, p4, p5, p6, p7;
-
- do {
- d0 = _mm_load_si128((const __m128i *)(diff + 0 * width));
- d1 = _mm_load_si128((const __m128i *)(diff + 0 * width + 8));
- d2 = _mm_load_si128((const __m128i *)(diff + 0 * width + 16));
- d3 = _mm_load_si128((const __m128i *)(diff + 0 * width + 24));
- d4 = _mm_load_si128((const __m128i *)(diff + 1 * width));
- d5 = _mm_load_si128((const __m128i *)(diff + 1 * width + 8));
- d6 = _mm_load_si128((const __m128i *)(diff + 1 * width + 16));
- d7 = _mm_load_si128((const __m128i *)(diff + 1 * width + 24));
-
- // Prediction data.
- p1 = _mm_load_si128((const __m128i *)(dest + 0 * stride));
- p3 = _mm_load_si128((const __m128i *)(dest + 0 * stride + 16));
- p5 = _mm_load_si128((const __m128i *)(dest + 1 * stride));
- p7 = _mm_load_si128((const __m128i *)(dest + 1 * stride + 16));
-
- p0 = _mm_unpacklo_epi8(p1, zero);
- p1 = _mm_unpackhi_epi8(p1, zero);
- p2 = _mm_unpacklo_epi8(p3, zero);
- p3 = _mm_unpackhi_epi8(p3, zero);
- p4 = _mm_unpacklo_epi8(p5, zero);
- p5 = _mm_unpackhi_epi8(p5, zero);
- p6 = _mm_unpacklo_epi8(p7, zero);
- p7 = _mm_unpackhi_epi8(p7, zero);
-
- p0 = _mm_add_epi16(p0, d0);
- p1 = _mm_add_epi16(p1, d1);
- p2 = _mm_add_epi16(p2, d2);
- p3 = _mm_add_epi16(p3, d3);
- p4 = _mm_add_epi16(p4, d4);
- p5 = _mm_add_epi16(p5, d5);
- p6 = _mm_add_epi16(p6, d6);
- p7 = _mm_add_epi16(p7, d7);
-
- p0 = _mm_packus_epi16(p0, p1);
- p1 = _mm_packus_epi16(p2, p3);
- p2 = _mm_packus_epi16(p4, p5);
- p3 = _mm_packus_epi16(p6, p7);
-
- _mm_store_si128((__m128i *)(dest + 0 * stride), p0);
- _mm_store_si128((__m128i *)(dest + 0 * stride + 16), p1);
- _mm_store_si128((__m128i *)(dest + 1 * stride), p2);
- _mm_store_si128((__m128i *)(dest + 1 * stride + 16), p3);
-
- diff += 2 * width;
- dest += 2 * stride;
- } while (--i);
-}
-
void vp9_add_constant_residual_8x8_sse2(const int16_t diff, uint8_t *dest,
int stride) {
uint8_t abs_diff;
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 32cef5e..9d72a88 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -619,12 +619,21 @@
active_section = 6;
#endif
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8)
+ write_sb_ymode(bc, mode, pc->fc.sb_ymode_prob);
+#else
if (m->mbmi.sb_type > BLOCK_SIZE_SB8X8)
write_sb_ymode(bc, mode, pc->fc.sb_ymode_prob);
else
write_ymode(bc, mode, pc->fc.ymode_prob);
+#endif
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type < BLOCK_SIZE_SB8X8) {
+#else
if (mode == I4X4_PRED) {
+#endif
int j = 0;
do {
write_bmode(bc, m->bmi[j].as_mode.first,
@@ -644,11 +653,16 @@
// If segment skip is not enabled code the mode.
if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
+#if CONFIG_AB4X4
+ if (mi->sb_type >= BLOCK_SIZE_SB8X8)
+ write_sb_mv_ref(bc, mode, mv_ref_p);
+#else
if (mi->sb_type > BLOCK_SIZE_SB8X8) {
write_sb_mv_ref(bc, mode, mv_ref_p);
} else {
write_mv_ref(bc, mode, mv_ref_p);
}
+#endif
vp9_accum_mv_refs(&cpi->common, mode, mi->mb_mode_context[rf]);
}
@@ -730,11 +744,20 @@
}
}
+#if CONFIG_AB4X4
+ if (((rf == INTRA_FRAME && mi->sb_type >= BLOCK_SIZE_SB8X8) ||
+ (rf != INTRA_FRAME && mi->sb_type >= BLOCK_SIZE_SB8X8)) &&
+ pc->txfm_mode == TX_MODE_SELECT &&
+ !(skip_coeff || vp9_segfeature_active(xd, segment_id,
+ SEG_LVL_SKIP)))
+#else
if (((rf == INTRA_FRAME && mode != I4X4_PRED) ||
(rf != INTRA_FRAME && mode != SPLITMV)) &&
pc->txfm_mode == TX_MODE_SELECT &&
!(skip_coeff || vp9_segfeature_active(xd, segment_id,
- SEG_LVL_SKIP))) {
+ SEG_LVL_SKIP)))
+#endif
+ {
TX_SIZE sz = mi->txfm_size;
// FIXME(rbultje) code ternary symbol once all experiments are merged
vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]);
@@ -766,12 +789,21 @@
vp9_write(bc, skip_coeff, vp9_get_pred_prob(c, xd, PRED_MBSKIP));
}
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8)
+ sb_kfwrite_ymode(bc, ym, c->sb_kf_ymode_prob[c->kf_ymode_probs_index]);
+#else
if (m->mbmi.sb_type > BLOCK_SIZE_SB8X8)
sb_kfwrite_ymode(bc, ym, c->sb_kf_ymode_prob[c->kf_ymode_probs_index]);
else
kfwrite_ymode(bc, ym, c->kf_ymode_prob[c->kf_ymode_probs_index]);
+#endif
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type < BLOCK_SIZE_SB8X8) {
+#else
if (ym == I4X4_PRED) {
+#endif
int i = 0;
do {
const B_PREDICTION_MODE a = above_block_mode(m, i, mis);
@@ -789,8 +821,13 @@
write_uv_mode(bc, m->mbmi.uv_mode, c->kf_uv_mode_prob[ym]);
+#if CONFIG_AB4X4
+ if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 && c->txfm_mode == TX_MODE_SELECT &&
+ !(skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) {
+#else
if (ym != I4X4_PRED && c->txfm_mode == TX_MODE_SELECT &&
!(skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) {
+#endif
TX_SIZE sz = m->mbmi.txfm_size;
// FIXME(rbultje) code ternary symbol once all experiments are merged
vp9_write(bc, sz != TX_4X4, c->prob_tx[0]);
@@ -862,7 +899,19 @@
else
assert(0);
+#if CONFIG_AB4X4
+ if (bsize == BLOCK_SIZE_SB8X8 && m->mbmi.sb_type < BLOCK_SIZE_SB8X8)
+ partition = PARTITION_SPLIT;
+ if (bsize < BLOCK_SIZE_SB8X8)
+ if (xd->ab_index != 0)
+ return;
+#endif
+
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8) {
+#else
if (bsize > BLOCK_SIZE_SB8X8) {
+#endif
int pl;
xd->left_seg_context = cm->left_seg_context + (mi_row & MI_MASK);
xd->above_seg_context = cm->above_seg_context + mi_col;
@@ -901,8 +950,13 @@
}
// update partition context
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8 &&
+ (bsize == BLOCK_SIZE_SB8X8 || partition != PARTITION_SPLIT)) {
+#else
if (bsize > BLOCK_SIZE_SB8X8 &&
(bsize == BLOCK_SIZE_MB16X16 || partition != PARTITION_SPLIT)) {
+#endif
set_partition_seg_context(cm, xd, mi_row, mi_col);
update_partition_context(xd, subsize, bsize);
}
@@ -1228,16 +1282,6 @@
FILE *vpxlogc = 0;
#endif
-static void put_delta_q(vp9_writer *bc, int delta_q) {
- if (delta_q != 0) {
- vp9_write_bit(bc, 1);
- vp9_write_literal(bc, abs(delta_q), 4);
- vp9_write_bit(bc, delta_q < 0);
- } else {
- vp9_write_bit(bc, 0);
- }
-}
-
static void decide_kf_ymode_entropy(VP9_COMP *cpi) {
int mode_cost[MB_MODE_COUNT];
int bestcost = INT_MAX;
@@ -1284,9 +1328,21 @@
}
}
-static void encode_loopfilter(MACROBLOCKD *xd, vp9_writer *w) {
+static void encode_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd, vp9_writer *w) {
int i;
+ // Encode the loop filter level and type
+ vp9_write_literal(w, pc->filter_level, 6);
+ vp9_write_literal(w, pc->sharpness_level, 3);
+#if CONFIG_LOOP_DERING
+ if (pc->dering_enabled) {
+ vp9_write_bit(w, 1);
+ vp9_write_literal(w, pc->dering_enabled - 1, 4);
+ } else {
+ vp9_write_bit(w, 0);
+ }
+#endif
+
// Write out loop filter deltas applied at the MB level based on mode or
// ref frame (if they are enabled).
vp9_write_bit(w, xd->mode_ref_lf_delta_enabled);
@@ -1340,6 +1396,24 @@
}
}
+static void put_delta_q(vp9_writer *bc, int delta_q) {
+ if (delta_q != 0) {
+ vp9_write_bit(bc, 1);
+ vp9_write_literal(bc, abs(delta_q), 4);
+ vp9_write_bit(bc, delta_q < 0);
+ } else {
+ vp9_write_bit(bc, 0);
+ }
+}
+
+static void encode_quantization(VP9_COMMON *pc, vp9_writer *w) {
+ vp9_write_literal(w, pc->base_qindex, QINDEX_BITS);
+ put_delta_q(w, pc->y_dc_delta_q);
+ put_delta_q(w, pc->uv_dc_delta_q);
+ put_delta_q(w, pc->uv_ac_delta_q);
+}
+
+
static void encode_segmentation(VP9_COMP *cpi, vp9_writer *w) {
int i, j;
VP9_COMMON *const pc = &cpi->common;
@@ -1478,30 +1552,9 @@
// error resilient mode
vp9_write_bit(&header_bc, pc->error_resilient_mode);
- // lossless mode: note this needs to be before loopfilter
- vp9_write_bit(&header_bc, cpi->mb.e_mbd.lossless);
+ encode_loopfilter(pc, xd, &header_bc);
- // Encode the loop filter level and type
- vp9_write_literal(&header_bc, pc->filter_level, 6);
- vp9_write_literal(&header_bc, pc->sharpness_level, 3);
-#if CONFIG_LOOP_DERING
- if (pc->dering_enabled) {
- vp9_write_bit(&header_bc, 1);
- vp9_write_literal(&header_bc, pc->dering_enabled - 1, 4);
- } else {
- vp9_write_bit(&header_bc, 0);
- }
-#endif
-
- encode_loopfilter(xd, &header_bc);
-
- // Frame Q baseline quantizer index
- vp9_write_literal(&header_bc, pc->base_qindex, QINDEX_BITS);
-
- // Transmit Dc, Second order and Uv quantizer delta information
- put_delta_q(&header_bc, pc->y_dc_delta_q);
- put_delta_q(&header_bc, pc->uv_dc_delta_q);
- put_delta_q(&header_bc, pc->uv_ac_delta_q);
+ encode_quantization(pc, &header_bc);
// When there is a key frame all reference buffers are updated using the new key frame
if (pc->frame_type != KEY_FRAME) {
diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h
index 4426148..d3851b4 100644
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -140,7 +140,12 @@
// TODO(jingning): Need to refactor the structure arrays that buffers the
// coding mode decisions of each partition type.
- PICK_MODE_CONTEXT sb8_context[4][4][4];
+#if CONFIG_AB4X4
+ PICK_MODE_CONTEXT ab4x4_context[4][4][4];
+ PICK_MODE_CONTEXT sb8x4_context[4][4][4];
+ PICK_MODE_CONTEXT sb4x8_context[4][4][4];
+#endif
+ PICK_MODE_CONTEXT sb8x8_context[4][4][4];
PICK_MODE_CONTEXT sb8x16_context[4][4][2];
PICK_MODE_CONTEXT sb16x8_context[4][4][2];
PICK_MODE_CONTEXT mb_context[4][4];
@@ -153,6 +158,9 @@
PICK_MODE_CONTEXT sb64_context;
int partition_cost[NUM_PARTITION_CONTEXTS][PARTITION_TYPES];
+#if CONFIG_AB4X4
+ BLOCK_SIZE_TYPE b_partitioning[4][4][4];
+#endif
BLOCK_SIZE_TYPE mb_partitioning[4][4];
BLOCK_SIZE_TYPE sb_partitioning[4];
BLOCK_SIZE_TYPE sb64_partitioning;
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 2edeb78..69ab147 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -361,8 +361,8 @@
assert(mb_mode_index < MAX_MODES);
assert(mi->mbmi.ref_frame < MAX_REF_FRAMES);
#endif
- assert(mi->mbmi.sb_type == bsize);
+ assert(mi->mbmi.sb_type == bsize);
// Restore the coding context of the MB to that that was in place
// when the mode was picked for it
for (y = 0; y < bh; y++) {
@@ -640,6 +640,12 @@
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8)
+ if (xd->ab_index != 0)
+ return;
+#endif
+
set_offsets(cpi, mi_row, mi_col, bsize);
xd->mode_info_context->mbmi.sb_type = bsize;
if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
@@ -718,7 +724,14 @@
} else if (bsize >= BLOCK_SIZE_MB16X16) {
xd->mb_index = idx;
} else {
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8)
+ xd->b_index = idx;
+ else
+ xd->ab_index = idx;
+#else
xd->b_index = idx;
+#endif
}
}
@@ -749,7 +762,15 @@
case BLOCK_SIZE_SB8X16:
return &x->sb8x16_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X8:
- return &x->sb8_context[xd->sb_index][xd->mb_index][xd->b_index];
+ return &x->sb8x8_context[xd->sb_index][xd->mb_index][xd->b_index];
+#if CONFIG_AB4X4
+ case BLOCK_SIZE_SB8X4:
+ return &x->sb8x4_context[xd->sb_index][xd->mb_index][xd->b_index];
+ case BLOCK_SIZE_SB4X8:
+ return &x->sb4x8_context[xd->sb_index][xd->mb_index][xd->b_index];
+ case BLOCK_SIZE_AB4X4:
+ return &x->ab4x4_context[xd->sb_index][xd->mb_index][xd->b_index];
+#endif
default:
assert(0);
return NULL;
@@ -766,6 +787,10 @@
return &x->sb_partitioning[xd->sb_index];
case BLOCK_SIZE_MB16X16:
return &x->mb_partitioning[xd->sb_index][xd->mb_index];
+#if CONFIG_AB4X4
+ case BLOCK_SIZE_SB8X8:
+ return &x->b_partitioning[xd->sb_index][xd->mb_index][xd->b_index];
+#endif
default:
assert(0);
return NULL;
@@ -833,14 +858,20 @@
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
BLOCK_SIZE_TYPE c1 = BLOCK_SIZE_SB8X8;
- const int bsl = mi_width_log2(bsize), bs = 1 << (bsl - 1);
+ const int bsl = mi_width_log2(bsize), bs = (1 << bsl) / 2;
int bwl, bhl;
int UNINITIALIZED_IS_SAFE(pl);
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
- if (bsize > BLOCK_SIZE_SB8X8) {
+#if CONFIG_AB4X4
+ c1 = BLOCK_SIZE_AB4X4;
+ if (bsize >= BLOCK_SIZE_SB8X8)
+#else
+ if (bsize > BLOCK_SIZE_SB8X8)
+#endif
+ {
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
c1 = *(get_sb_partitioning(x, bsize));
@@ -849,8 +880,18 @@
bwl = mi_width_log2(c1), bhl = mi_height_log2(c1);
if (bsl == bwl && bsl == bhl) {
+#if CONFIG_AB4X4
+ if (output_enabled && bsize >= BLOCK_SIZE_SB8X8) {
+ if (bsize > BLOCK_SIZE_SB8X8 ||
+ (bsize == BLOCK_SIZE_SB8X8 && c1 == bsize))
+ cpi->partition_count[pl][PARTITION_NONE]++;
+ else
+ cpi->partition_count[pl][PARTITION_SPLIT]++;
+ }
+#else
if (output_enabled && bsize > BLOCK_SIZE_SB8X8)
cpi->partition_count[pl][PARTITION_NONE]++;
+#endif
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, -1);
} else if (bsl == bhl && bsl > bwl) {
if (output_enabled)
@@ -867,14 +908,7 @@
int i;
assert(bwl < bsl && bhl < bsl);
- if (bsize == BLOCK_SIZE_SB64X64) {
- subsize = BLOCK_SIZE_SB32X32;
- } else if (bsize == BLOCK_SIZE_SB32X32) {
- subsize = BLOCK_SIZE_MB16X16;
- } else {
- assert(bsize == BLOCK_SIZE_MB16X16);
- subsize = BLOCK_SIZE_SB8X8;
- }
+ subsize = get_subsize(bsize, PARTITION_SPLIT);
if (output_enabled)
cpi->partition_count[pl][PARTITION_SPLIT]++;
@@ -888,8 +922,13 @@
}
}
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8 &&
+ (bsize == BLOCK_SIZE_SB8X8 || bsl == bwl || bsl == bhl)) {
+#else
if (bsize > BLOCK_SIZE_SB8X8 &&
(bsize == BLOCK_SIZE_MB16X16 || bsl == bwl || bsl == bhl)) {
+#endif
set_partition_seg_context(cm, xd, mi_row, mi_col);
update_partition_context(xd, c1, bsize);
}
@@ -907,7 +946,7 @@
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
int bsl = b_width_log2(bsize), bs = 1 << bsl;
- int msl = mi_height_log2(bsize), ms = 1 << msl;
+ int ms = bs / 2;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
TOKENEXTRA *tp_orig = *tp;
@@ -915,6 +954,15 @@
BLOCK_SIZE_TYPE subsize;
int srate = INT_MAX, sdist = INT_MAX;
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8)
+ if (xd->ab_index != 0) {
+ *rate = 0;
+ *dist = 0;
+ return;
+ }
+#endif
+
assert(mi_height_log2(bsize) == mi_width_log2(bsize));
// buffer the above/left context information of the block in search.
@@ -932,7 +980,11 @@
sizeof(PARTITION_CONTEXT) * ms);
// PARTITION_SPLIT
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8) {
+#else
if (bsize >= BLOCK_SIZE_MB16X16) {
+#endif
int r4 = 0, d4 = 0;
subsize = get_subsize(bsize, PARTITION_SPLIT);
*(get_sb_partitioning(x, bsize)) = subsize;
@@ -948,18 +1000,26 @@
*(get_sb_index(xd, subsize)) = i;
rd_pick_partition(cpi, tp, mi_row + y_idx, mi_col + x_idx, subsize,
&r, &d);
+
r4 += r;
d4 += d;
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
+#if CONFIG_AB4X4
+ if (r4 < INT_MAX)
+ r4 += x->partition_cost[pl][PARTITION_SPLIT];
+#else
r4 += x->partition_cost[pl][PARTITION_SPLIT];
-
+#endif
+ assert(r4 >= 0);
+ assert(d4 >= 0);
srate = r4;
sdist = d4;
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
+ // TODO(jingning): need to enable 4x8 and 8x4 partition coding
// PARTITION_HORZ
if ((mi_col + ms <= cm->mi_cols) && (mi_row + (ms >> 1) <= cm->mi_rows) &&
(bsize >= BLOCK_SIZE_MB16X16)) {
@@ -1036,7 +1096,11 @@
int r, d;
pick_sb_modes(cpi, mi_row, mi_col, tp, &r, &d, bsize,
get_block_context(x, bsize));
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8) {
+#else
if (bsize >= BLOCK_SIZE_MB16X16) {
+#endif
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
r += x->partition_cost[pl][PARTITION_NONE];
@@ -1046,21 +1110,28 @@
RDCOST(x->rdmult, x->rddiv, srate, sdist)) {
srate = r;
sdist = d;
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8)
+#else
if (bsize >= BLOCK_SIZE_MB16X16)
+#endif
*(get_sb_partitioning(x, bsize)) = bsize;
}
}
- assert(srate < INT_MAX && sdist < INT_MAX);
*rate = srate;
*dist = sdist;
- encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize);
+ if (srate < INT_MAX && sdist < INT_MAX)
+ encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize);
- if (bsize == BLOCK_SIZE_SB64X64)
+ if (bsize == BLOCK_SIZE_SB64X64) {
assert(tp_orig < *tp);
- else
+ assert(srate < INT_MAX);
+ assert(sdist < INT_MAX);
+ } else {
assert(tp_orig == *tp);
+ }
}
static void encode_sb_row(VP9_COMP *cpi, int mi_row,
@@ -1198,10 +1269,10 @@
vp9_zero(cpi->coef_counts_32x32);
vp9_zero(cm->fc.eob_branch_counts);
- cpi->mb.e_mbd.lossless = (cm->base_qindex == 0 &&
- cm->y_dc_delta_q == 0 &&
- cm->uv_dc_delta_q == 0 &&
- cm->uv_ac_delta_q == 0);
+ cpi->mb.e_mbd.lossless = cm->base_qindex == 0 &&
+ cm->y_dc_delta_q == 0 &&
+ cm->uv_dc_delta_q == 0 &&
+ cm->uv_ac_delta_q == 0;
switch_lossless_mode(cpi, cpi->mb.e_mbd.lossless);
vp9_frame_init_quantizer(cpi);
@@ -1587,7 +1658,11 @@
}
#endif
+#if CONFIG_AB4X4
+ if (xd->mode_info_context->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
+#else
if (xd->mode_info_context->mbmi.sb_type > BLOCK_SIZE_SB8X8) {
+#endif
++cpi->sb_ymode_count[m];
} else {
++cpi->ymode_count[m];
@@ -1672,13 +1747,17 @@
vp9_update_zbin_extra(cpi, x);
}
+#if CONFIG_AB4X4
+ if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME &&
+ bsize < BLOCK_SIZE_SB8X8) {
+#else
if (xd->mode_info_context->mbmi.mode == I4X4_PRED) {
assert(bsize == BLOCK_SIZE_SB8X8 &&
xd->mode_info_context->mbmi.txfm_size == TX_4X4);
-
- vp9_encode_intra4x4mby(x, bsize);
- vp9_build_intra_predictors_sbuv_s(&x->e_mbd, bsize);
- vp9_encode_sbuv(cm, x, bsize);
+#endif
+ vp9_encode_intra4x4mby(x, BLOCK_SIZE_SB8X8);
+ vp9_build_intra_predictors_sbuv_s(&x->e_mbd, BLOCK_SIZE_SB8X8);
+ vp9_encode_sbuv(cm, x, BLOCK_SIZE_SB8X8);
if (output_enabled)
sum_intra_stats(cpi, x);
@@ -1714,15 +1793,22 @@
? &cpi->common.yv12_fb[second_ref_fb_idx] : NULL,
mi_row, mi_col, xd->scale_factor, xd->scale_factor_uv);
- vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+ vp9_build_inter_predictors_sb(xd, mi_row, mi_col,
+ (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize);
}
+#if CONFIG_AB4X4
+ if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME &&
+ bsize < BLOCK_SIZE_SB8X8) {
+#else
if (xd->mode_info_context->mbmi.mode == I4X4_PRED) {
assert(bsize == BLOCK_SIZE_SB8X8);
- vp9_tokenize_sb(cpi, &x->e_mbd, t, !output_enabled, bsize);
+#endif
+ vp9_tokenize_sb(cpi, &x->e_mbd, t, !output_enabled, BLOCK_SIZE_SB8X8);
} else if (!x->skip) {
- vp9_encode_sb(cm, x, bsize);
- vp9_tokenize_sb(cpi, &x->e_mbd, t, !output_enabled, bsize);
+ vp9_encode_sb(cm, x, (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize);
+ vp9_tokenize_sb(cpi, &x->e_mbd, t, !output_enabled,
+ (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize);
} else {
// FIXME(rbultje): not tile-aware (mi - 1)
int mb_skip_context =
@@ -1731,7 +1817,8 @@
xd->mode_info_context->mbmi.mb_skip_coeff = 1;
if (output_enabled)
cpi->skip_true_count[mb_skip_context]++;
- vp9_reset_sb_tokens_context(xd, bsize);
+ vp9_reset_sb_tokens_context(xd,
+ (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize);
}
// copy skip flag on all mb_mode_info contexts in this SB
@@ -1761,8 +1848,12 @@
sz = TX_16X16;
if (sz == TX_16X16 && bsize < BLOCK_SIZE_MB16X16)
sz = TX_8X8;
+#if CONFIG_AB4X4
+ if (sz == TX_8X8 && bsize < BLOCK_SIZE_SB8X8)
+#else
if (sz == TX_8X8 && (xd->mode_info_context->mbmi.mode == SPLITMV ||
xd->mode_info_context->mbmi.mode == I4X4_PRED))
+#endif
sz = TX_4X4;
for (y = 0; y < bh; y++) {
diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index e4002d6..221de74 100644
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -425,6 +425,7 @@
VP9_COMMON *cm;
MACROBLOCK *x;
struct optimize_ctx *ctx;
+ int *wip_txfrm_size; // for "work in progress" only... will remove once done
};
static void xform_quant(int plane, int block, BLOCK_SIZE_TYPE bsize,
@@ -493,6 +494,7 @@
int ss_txfrm_size, void *arg) {
struct encode_b_args* const args = arg;
MACROBLOCK* const x = args->x;
+ int *wip_txfrm_size = args->wip_txfrm_size;
MACROBLOCKD* const xd = &x->e_mbd;
const int bw = 4 << (b_width_log2(bsize) - xd->plane[plane].subsampling_x);
const int raster_block = txfrm_block_to_raster_block(xd, bsize, plane,
@@ -500,6 +502,10 @@
int16_t* const diff = raster_block_offset_int16(xd, bsize, plane,
raster_block,
xd->plane[plane].diff);
+ uint8_t* const dst = raster_block_offset_uint8(xd, bsize, plane,
+ raster_block,
+ xd->plane[plane].dst.buf,
+ xd->plane[plane].dst.stride);
TX_TYPE tx_type = DCT_DCT;
xform_quant(plane, block, bsize, ss_txfrm_size, arg);
@@ -509,18 +515,21 @@
switch (ss_txfrm_size / 2) {
case TX_32X32:
- vp9_short_idct32x32(BLOCK_OFFSET(xd->plane[plane].dqcoeff, block, 16),
- diff, bw * 2);
+ vp9_short_idct32x32_add(BLOCK_OFFSET(xd->plane[plane].dqcoeff,
+ block, 16), dst, xd->plane[plane].dst.stride);
+ *wip_txfrm_size = 32;
break;
case TX_16X16:
tx_type = plane == 0 ? get_tx_type_16x16(xd, raster_block) : DCT_DCT;
if (tx_type == DCT_DCT) {
- vp9_short_idct16x16(BLOCK_OFFSET(xd->plane[plane].dqcoeff, block, 16),
- diff, bw * 2);
+ vp9_short_idct16x16_add(BLOCK_OFFSET(xd->plane[plane].dqcoeff,
+ block, 16), dst, xd->plane[plane].dst.stride);
} else {
- vp9_short_iht16x16(BLOCK_OFFSET(xd->plane[plane].dqcoeff, block, 16),
- diff, bw, tx_type);
+ vp9_short_iht16x16_add(BLOCK_OFFSET(xd->plane[plane].dqcoeff,
+ block, 16), dst, xd->plane[plane].dst.stride,
+ tx_type);
}
+ *wip_txfrm_size = 16;
break;
case TX_8X8:
tx_type = plane == 0 ? get_tx_type_8x8(xd, raster_block) : DCT_DCT;
@@ -531,6 +540,7 @@
vp9_short_iht8x8(BLOCK_OFFSET(xd->plane[plane].dqcoeff, block, 16),
diff, bw, tx_type);
}
+ *wip_txfrm_size = 8;
break;
case TX_4X4:
tx_type = plane == 0 ? get_tx_type_4x4(xd, raster_block) : DCT_DCT;
@@ -544,6 +554,7 @@
vp9_short_iht4x4(BLOCK_OFFSET(xd->plane[plane].dqcoeff, block, 16),
diff, bw, tx_type);
}
+ *wip_txfrm_size = 4;
break;
}
}
@@ -551,7 +562,7 @@
void vp9_xform_quant_sby(VP9_COMMON *const cm, MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD* const xd = &x->e_mbd;
- struct encode_b_args arg = {cm, x, NULL};
+ struct encode_b_args arg = {cm, x, NULL, NULL};
foreach_transformed_block_in_plane(xd, bsize, 0,
xform_quant, &arg);
@@ -560,7 +571,7 @@
void vp9_xform_quant_sbuv(VP9_COMMON *const cm, MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD* const xd = &x->e_mbd;
- struct encode_b_args arg = {cm, x, NULL};
+ struct encode_b_args arg = {cm, x, NULL, NULL};
foreach_transformed_block_uv(xd, bsize, xform_quant, &arg);
}
@@ -569,7 +580,8 @@
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD* const xd = &x->e_mbd;
struct optimize_ctx ctx;
- struct encode_b_args arg = {cm, x, &ctx};
+ int wip_txfrm_size = 0;
+ struct encode_b_args arg = {cm, x, &ctx, &wip_txfrm_size};
vp9_subtract_sby(x, bsize);
if (x->optimize)
@@ -577,15 +589,16 @@
foreach_transformed_block_in_plane(xd, bsize, 0,
encode_block, &arg);
-
- vp9_recon_sby(xd, bsize);
+ if (wip_txfrm_size < 32)
+ vp9_recon_sby(xd, bsize);
}
void vp9_encode_sbuv(VP9_COMMON *const cm, MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD* const xd = &x->e_mbd;
struct optimize_ctx ctx;
- struct encode_b_args arg = {cm, x, &ctx};
+ int wip_txfrm_size = 0;
+ struct encode_b_args arg = {cm, x, &ctx, &wip_txfrm_size};
vp9_subtract_sbuv(x, bsize);
if (x->optimize)
@@ -593,20 +606,35 @@
foreach_transformed_block_uv(xd, bsize, encode_block, &arg);
- vp9_recon_sbuv(xd, bsize);
+ if (wip_txfrm_size < 16)
+ vp9_recon_sbuv(xd, bsize);
}
void vp9_encode_sb(VP9_COMMON *const cm, MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD* const xd = &x->e_mbd;
struct optimize_ctx ctx;
- struct encode_b_args arg = {cm, x, &ctx};
+ int wip_txfrm_size = 0;
+ struct encode_b_args arg = {cm, x, &ctx, &wip_txfrm_size};
vp9_subtract_sb(x, bsize);
if (x->optimize)
vp9_optimize_init(xd, bsize, &ctx);
-
+#if 0
foreach_transformed_block(xd, bsize, encode_block, &arg);
vp9_recon_sb(xd, bsize);
+#else
+ // wip version... will use foreach_transformed_block when done
+ foreach_transformed_block_in_plane(xd, bsize, 0,
+ encode_block, &arg);
+ if (wip_txfrm_size < 16)
+ vp9_recon_sby(xd, bsize);
+ wip_txfrm_size = 0;
+
+ foreach_transformed_block_uv(xd, bsize, encode_block, &arg);
+
+ if (wip_txfrm_size < 16)
+ vp9_recon_sbuv(xd, bsize);
+#endif
}
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 7223851..367ebec 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -47,7 +47,7 @@
#define KF_MB_INTRA_MIN 150
#define GF_MB_INTRA_MIN 100
-#define DOUBLE_DIVIDE_CHECK(X) ((X)<0?(X)-.000001:(X)+.000001)
+#define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
#define POW1 (double)cpi->oxcf.two_pass_vbrbias/100.0
#define POW2 (double)cpi->oxcf.two_pass_vbrbias/100.0
@@ -78,8 +78,8 @@
// Resets the first pass file to the given position using a relative seek from the current position
-static void reset_fpf_position(VP9_COMP *cpi, FIRSTPASS_STATS *Position) {
- cpi->twopass.stats_in = Position;
+static void reset_fpf_position(VP9_COMP *cpi, FIRSTPASS_STATS *position) {
+ cpi->twopass.stats_in = position;
}
static int lookup_next_frame_stats(VP9_COMP *cpi, FIRSTPASS_STATS *next_frame) {
@@ -252,17 +252,11 @@
// Calculate a modified Error used in distributing bits between easier and harder frames
static double calculate_modified_err(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
- double av_err = (cpi->twopass.total_stats.ssim_weighted_pred_err /
- cpi->twopass.total_stats.count);
- double this_err = this_frame->ssim_weighted_pred_err;
- double modified_err;
-
- if (this_err > av_err)
- modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW1);
- else
- modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW2);
-
- return modified_err;
+ const FIRSTPASS_STATS *const stats = &cpi->twopass.total_stats;
+ const double av_err = stats->ssim_weighted_pred_err / stats->count;
+ const double this_err = this_frame->ssim_weighted_pred_err;
+ return av_err * pow(this_err / DOUBLE_DIVIDE_CHECK(av_err),
+ this_err > av_err ? POW1 : POW2);
}
static const double weight_table[256] = {
@@ -328,20 +322,14 @@
static int frame_max_bits(VP9_COMP *cpi) {
// Max allocation for a single frame based on the max section guidelines
// passed in and how many bits are left.
- int max_bits;
-
// For VBR base this on the bits and frames left plus the
// two_pass_vbrmax_section rate passed in by the user.
- max_bits = (int) (((double) cpi->twopass.bits_left
- / (cpi->twopass.total_stats.count - (double) cpi->common
- .current_video_frame))
- * ((double) cpi->oxcf.two_pass_vbrmax_section / 100.0));
+ const double max_bits = (1.0 * cpi->twopass.bits_left /
+ (cpi->twopass.total_stats.count - cpi->common.current_video_frame)) *
+ (cpi->oxcf.two_pass_vbrmax_section / 100.0);
// Trap case where we are out of bits.
- if (max_bits < 0)
- max_bits = 0;
-
- return max_bits;
+ return MAX((int)max_bits, 0);
}
void vp9_init_first_pass(VP9_COMP *cpi) {
@@ -534,6 +522,8 @@
xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
xd->left_available = (mb_col != 0);
+ xd->mode_info_context->mbmi.sb_type = BLOCK_SIZE_MB16X16;
+
// do intra 16x16 prediction
this_error = vp9_encode_intra(cpi, x, use_dc_pred);
@@ -780,7 +770,7 @@
// swap frame pointers so last frame refers to the frame we just compressed
swap_yv12(lst_yv12, new_yv12);
- vp8_yv12_extend_frame_borders(lst_yv12);
+ vp9_extend_frame_borders(lst_yv12, cm->subsampling_x, cm->subsampling_y);
// Special case for the first frame. Copy into the GF buffer as a second reference.
if (cm->current_video_frame == 0)
@@ -854,26 +844,18 @@
double err_divisor,
double pt_low,
double pt_high,
- int Q) {
- double power_term;
- double error_term = err_per_mb / err_divisor;
- double correction_factor;
+ int q) {
+ const double error_term = err_per_mb / err_divisor;
// Adjustment based on actual quantizer to power term.
- power_term = (vp9_convert_qindex_to_q(Q) * 0.01) + pt_low;
- power_term = (power_term > pt_high) ? pt_high : power_term;
+ const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.01 + pt_low,
+ pt_high);
// Calculate correction factor
if (power_term < 1.0)
assert(error_term >= 0.0);
- correction_factor = pow(error_term, power_term);
- // Clip range
- correction_factor =
- (correction_factor < 0.05)
- ? 0.05 : (correction_factor > 5.0) ? 5.0 : correction_factor;
-
- return correction_factor;
+ return fclamp(pow(error_term, power_term), 0.05, 5.0);
}
// Given a current maxQ value sets a range for future values.
@@ -882,10 +864,8 @@
// (now uses the actual quantizer) but has not been tuned.
static void adjust_maxq_qrange(VP9_COMP *cpi) {
int i;
- double q;
-
// Set the max corresponding to cpi->avg_q * 2.0
- q = cpi->avg_q * 2.0;
+ double q = cpi->avg_q * 2.0;
cpi->twopass.maxq_max_limit = cpi->worst_quality;
for (i = cpi->best_quality; i <= cpi->worst_quality; i++) {
cpi->twopass.maxq_max_limit = i;
@@ -906,12 +886,11 @@
static int estimate_max_q(VP9_COMP *cpi,
FIRSTPASS_STATS *fpstats,
int section_target_bandwitdh) {
- int Q;
+ int q;
int num_mbs = cpi->common.MBs;
int target_norm_bits_per_mb;
- double section_err = (fpstats->coded_error / fpstats->count);
- double sr_err_diff;
+ double section_err = fpstats->coded_error / fpstats->count;
double sr_correction;
double err_per_mb = section_err / num_mbs;
double err_correction_factor;
@@ -920,92 +899,74 @@
if (section_target_bandwitdh <= 0)
return cpi->twopass.maxq_max_limit; // Highest value allowed
- target_norm_bits_per_mb =
- (section_target_bandwitdh < (1 << 20))
- ? (512 * section_target_bandwitdh) / num_mbs
- : 512 * (section_target_bandwitdh / num_mbs);
+ target_norm_bits_per_mb = section_target_bandwitdh < (1 << 20)
+ ? (512 * section_target_bandwitdh) / num_mbs
+ : 512 * (section_target_bandwitdh / num_mbs);
// Look at the drop in prediction quality between the last frame
// and the GF buffer (which contained an older frame).
if (fpstats->sr_coded_error > fpstats->coded_error) {
- sr_err_diff =
- (fpstats->sr_coded_error - fpstats->coded_error) /
- (fpstats->count * cpi->common.MBs);
- sr_correction = (sr_err_diff / 32.0);
- sr_correction = pow(sr_correction, 0.25);
- if (sr_correction < 0.75)
- sr_correction = 0.75;
- else if (sr_correction > 1.25)
- sr_correction = 1.25;
+ double sr_err_diff = (fpstats->sr_coded_error - fpstats->coded_error) /
+ (fpstats->count * cpi->common.MBs);
+ sr_correction = fclamp(pow(sr_err_diff / 32.0, 0.25), 0.75, 1.25);
} else {
sr_correction = 0.75;
}
// Calculate a corrective factor based on a rolling ratio of bits spent
// vs target bits
- if ((cpi->rolling_target_bits > 0) &&
- (cpi->active_worst_quality < cpi->worst_quality)) {
- double rolling_ratio;
-
- rolling_ratio = (double)cpi->rolling_actual_bits /
- (double)cpi->rolling_target_bits;
+ if (cpi->rolling_target_bits > 0 &&
+ cpi->active_worst_quality < cpi->worst_quality) {
+ double rolling_ratio = (double)cpi->rolling_actual_bits /
+ (double)cpi->rolling_target_bits;
if (rolling_ratio < 0.95)
cpi->twopass.est_max_qcorrection_factor -= 0.005;
else if (rolling_ratio > 1.05)
cpi->twopass.est_max_qcorrection_factor += 0.005;
- cpi->twopass.est_max_qcorrection_factor =
- (cpi->twopass.est_max_qcorrection_factor < 0.1)
- ? 0.1
- : (cpi->twopass.est_max_qcorrection_factor > 10.0)
- ? 10.0 : cpi->twopass.est_max_qcorrection_factor;
+ cpi->twopass.est_max_qcorrection_factor = fclamp(
+ cpi->twopass.est_max_qcorrection_factor, 0.1, 10.0);
}
// Corrections for higher compression speed settings
// (reduced compression expected)
- if (cpi->compressor_speed == 1) {
- if (cpi->oxcf.cpu_used <= 5)
- speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
- else
- speed_correction = 1.25;
- }
+ if (cpi->compressor_speed == 1)
+ speed_correction = cpi->oxcf.cpu_used <= 5 ?
+ 1.04 + (cpi->oxcf.cpu_used * 0.04) :
+ 1.25;
// Try and pick a max Q that will be high enough to encode the
// content at the given rate.
- for (Q = cpi->twopass.maxq_min_limit; Q < cpi->twopass.maxq_max_limit; Q++) {
+ for (q = cpi->twopass.maxq_min_limit; q < cpi->twopass.maxq_max_limit; q++) {
int bits_per_mb_at_this_q;
- err_correction_factor =
- calc_correction_factor(err_per_mb, ERR_DIVISOR, 0.4, 0.90, Q) *
- sr_correction * speed_correction *
- cpi->twopass.est_max_qcorrection_factor;
+ err_correction_factor = calc_correction_factor(err_per_mb,
+ ERR_DIVISOR, 0.4, 0.90, q) *
+ sr_correction * speed_correction *
+ cpi->twopass.est_max_qcorrection_factor;
-
- bits_per_mb_at_this_q =
- vp9_bits_per_mb(INTER_FRAME, Q, err_correction_factor);
+ bits_per_mb_at_this_q = vp9_bits_per_mb(INTER_FRAME, q,
+ err_correction_factor);
if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
break;
}
// Restriction on active max q for constrained quality mode.
- if ((cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
- (Q < cpi->cq_target_quality)) {
- Q = cpi->cq_target_quality;
- }
+ if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY &&
+ q < cpi->cq_target_quality)
+ q = cpi->cq_target_quality;
// Adjust maxq_min_limit and maxq_max_limit limits based on
// average q observed in clip for non kf/gf/arf frames
// Give average a chance to settle though.
// PGW TODO.. This code is broken for the extended Q range
- if ((cpi->ni_frames >
- ((int)cpi->twopass.total_stats.count >> 8)) &&
- (cpi->ni_frames > 25)) {
+ if (cpi->ni_frames > ((int)cpi->twopass.total_stats.count >> 8) &&
+ cpi->ni_frames > 25)
adjust_maxq_qrange(cpi);
- }
- return Q;
+ return q;
}
// For cq mode estimate a cq level that matches the observed
@@ -1013,7 +974,7 @@
static int estimate_cq(VP9_COMP *cpi,
FIRSTPASS_STATS *fpstats,
int section_target_bandwitdh) {
- int Q;
+ int q;
int num_mbs = cpi->common.MBs;
int target_norm_bits_per_mb;
@@ -1064,29 +1025,29 @@
clip_iifactor = 0.80;
// Try and pick a Q that can encode the content at the given rate.
- for (Q = 0; Q < MAXQ; Q++) {
+ for (q = 0; q < MAXQ; q++) {
int bits_per_mb_at_this_q;
// Error per MB based correction factor
err_correction_factor =
- calc_correction_factor(err_per_mb, 100.0, 0.4, 0.90, Q) *
+ calc_correction_factor(err_per_mb, 100.0, 0.4, 0.90, q) *
sr_correction * speed_correction * clip_iifactor;
bits_per_mb_at_this_q =
- vp9_bits_per_mb(INTER_FRAME, Q, err_correction_factor);
+ vp9_bits_per_mb(INTER_FRAME, q, err_correction_factor);
if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
break;
}
// Clip value to range "best allowed to (worst allowed - 1)"
- Q = select_cq_level(Q);
- if (Q >= cpi->worst_quality)
- Q = cpi->worst_quality - 1;
- if (Q < cpi->best_quality)
- Q = cpi->best_quality;
+ q = select_cq_level(q);
+ if (q >= cpi->worst_quality)
+ q = cpi->worst_quality - 1;
+ if (q < cpi->best_quality)
+ q = cpi->best_quality;
- return Q;
+ return q;
}
@@ -1117,9 +1078,8 @@
// encoded in the second pass is a guess. However the sum duration is not.
// Its calculated based on the actual durations of all frames from the first
// pass.
- vp9_new_frame_rate(cpi,
- 10000000.0 * cpi->twopass.total_stats.count /
- cpi->twopass.total_stats.duration);
+ vp9_new_frame_rate(cpi, 10000000.0 * cpi->twopass.total_stats.count /
+ cpi->twopass.total_stats.duration);
cpi->output_frame_rate = cpi->oxcf.frame_rate;
cpi->twopass.bits_left = (int64_t)(cpi->twopass.total_stats.duration *
@@ -1191,9 +1151,8 @@
// Look at the observed drop in prediction quality between the last frame
// and the GF buffer (which contains an older frame).
- mb_sr_err_diff =
- (next_frame->sr_coded_error - next_frame->coded_error) /
- (cpi->common.MBs);
+ mb_sr_err_diff = (next_frame->sr_coded_error - next_frame->coded_error) /
+ cpi->common.MBs;
if (mb_sr_err_diff <= 512.0) {
second_ref_decay = 1.0 - (mb_sr_err_diff / 512.0);
second_ref_decay = pow(second_ref_decay, 0.5);
@@ -1225,9 +1184,9 @@
// Break clause to detect very still sections after motion
// For example a static image after a fade or other transition
// instead of a clean scene cut.
- if ((frame_interval > MIN_GF_INTERVAL) &&
- (loop_decay_rate >= 0.999) &&
- (last_decay_rate < 0.9)) {
+ if (frame_interval > MIN_GF_INTERVAL &&
+ loop_decay_rate >= 0.999 &&
+ last_decay_rate < 0.9) {
int j;
FIRSTPASS_STATS *position = cpi->twopass.stats_in;
FIRSTPASS_STATS tmp_next_frame;
@@ -1271,10 +1230,9 @@
// are reasonably well predicted by an earlier (pre flash) frame.
// The recovery after a flash is indicated by a high pcnt_second_ref
// comapred to pcnt_inter.
- if ((next_frame.pcnt_second_ref > next_frame.pcnt_inter) &&
- (next_frame.pcnt_second_ref >= 0.5)) {
+ if (next_frame.pcnt_second_ref > next_frame.pcnt_inter &&
+ next_frame.pcnt_second_ref >= 0.5)
flash_detected = 1;
- }
}
return flash_detected;
@@ -1356,13 +1314,9 @@
return frame_boost;
}
-static int calc_arf_boost(
- VP9_COMP *cpi,
- int offset,
- int f_frames,
- int b_frames,
- int *f_boost,
- int *b_boost) {
+static int calc_arf_boost(VP9_COMP *cpi, int offset,
+ int f_frames, int b_frames,
+ int *f_boost, int *b_boost) {
FIRSTPASS_STATS this_frame;
int i;
@@ -1392,8 +1346,7 @@
// Cumulative effect of prediction quality decay
if (!flash_detected) {
- decay_accumulator =
- decay_accumulator * get_prediction_decay_rate(cpi, &this_frame);
+ decay_accumulator *= get_prediction_decay_rate(cpi, &this_frame);
decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
? MIN_DECAY_FACTOR : decay_accumulator;
}
@@ -1429,10 +1382,9 @@
// Cumulative effect of prediction quality decay
if (!flash_detected) {
- decay_accumulator =
- decay_accumulator * get_prediction_decay_rate(cpi, &this_frame);
+ decay_accumulator *= get_prediction_decay_rate(cpi, &this_frame);
decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
- ? MIN_DECAY_FACTOR : decay_accumulator;
+ ? MIN_DECAY_FACTOR : decay_accumulator;
}
boost_score += (decay_accumulator *
@@ -1871,26 +1823,20 @@
for (i = 0;
i <= (cpi->source_alt_ref_pending && cpi->common.frame_type != KEY_FRAME);
++i) {
- int boost;
int allocation_chunks;
- int Q =
- (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
+ int q = cpi->oxcf.fixed_q < 0 ? cpi->last_q[INTER_FRAME]
+ : cpi->oxcf.fixed_q;
int gf_bits;
- boost = (cpi->gfu_boost * vp9_gfboost_qadjust(Q)) / 100;
+ int boost = (cpi->gfu_boost * vp9_gfboost_qadjust(q)) / 100;
// Set max and minimum boost and hence minimum allocation
- if (boost > ((cpi->baseline_gf_interval + 1) * 200))
- boost = ((cpi->baseline_gf_interval + 1) * 200);
- else if (boost < 125)
- boost = 125;
+ boost = clamp(boost, 125, (cpi->baseline_gf_interval + 1) * 200);
if (cpi->source_alt_ref_pending && i == 0)
- allocation_chunks =
- ((cpi->baseline_gf_interval + 1) * 100) + boost;
+ allocation_chunks = ((cpi->baseline_gf_interval + 1) * 100) + boost;
else
- allocation_chunks =
- (cpi->baseline_gf_interval * 100) + (boost - 100);
+ allocation_chunks = (cpi->baseline_gf_interval * 100) + (boost - 100);
// Prevent overflow
if (boost > 1023) {
@@ -1901,41 +1847,34 @@
// Calculate the number of bits to be spent on the gf or arf based on
// the boost number
- gf_bits = (int)((double)boost *
- (cpi->twopass.gf_group_bits /
- (double)allocation_chunks));
+ gf_bits = (int)((double)boost * (cpi->twopass.gf_group_bits /
+ (double)allocation_chunks));
// If the frame that is to be boosted is simpler than the average for
// the gf/arf group then use an alternative calculation
// based on the error score of the frame itself
if (mod_frame_err < gf_group_err / (double)cpi->baseline_gf_interval) {
- double alt_gf_grp_bits;
- int alt_gf_bits;
-
- alt_gf_grp_bits =
+ double alt_gf_grp_bits =
(double)cpi->twopass.kf_group_bits *
(mod_frame_err * (double)cpi->baseline_gf_interval) /
DOUBLE_DIVIDE_CHECK(cpi->twopass.kf_group_error_left);
- alt_gf_bits = (int)((double)boost * (alt_gf_grp_bits /
+ int alt_gf_bits = (int)((double)boost * (alt_gf_grp_bits /
(double)allocation_chunks));
- if (gf_bits > alt_gf_bits) {
+ if (gf_bits > alt_gf_bits)
gf_bits = alt_gf_bits;
- }
}
// Else if it is harder than other frames in the group make sure it at
// least receives an allocation in keeping with its relative error
// score, otherwise it may be worse off than an "un-boosted" frame
else {
- int alt_gf_bits =
- (int)((double)cpi->twopass.kf_group_bits *
- mod_frame_err /
- DOUBLE_DIVIDE_CHECK(cpi->twopass.kf_group_error_left));
+ int alt_gf_bits = (int)((double)cpi->twopass.kf_group_bits *
+ mod_frame_err /
+ DOUBLE_DIVIDE_CHECK(cpi->twopass.kf_group_error_left));
- if (alt_gf_bits > gf_bits) {
+ if (alt_gf_bits > gf_bits)
gf_bits = alt_gf_bits;
- }
}
// Dont allow a negative value for gf_bits
@@ -1983,14 +1922,11 @@
// despite (MIN_GF_INTERVAL) and would cause a divide by 0 in the
// calculation of alt_extra_bits.
if (cpi->baseline_gf_interval >= 3) {
- int boost = (cpi->source_alt_ref_pending)
- ? b_boost : cpi->gfu_boost;
+ const int boost = cpi->source_alt_ref_pending ? b_boost : cpi->gfu_boost;
if (boost >= 150) {
- int pct_extra;
int alt_extra_bits;
-
- pct_extra = (boost - 100) / 50;
+ int pct_extra = (boost - 100) / 50;
pct_extra = (pct_extra > 20) ? 20 : pct_extra;
alt_extra_bits = (int)((cpi->twopass.gf_group_bits * pct_extra) / 100);
@@ -2071,33 +2007,21 @@
// Make a damped adjustment to the active max q.
static int adjust_active_maxq(int old_maxqi, int new_maxqi) {
int i;
- int ret_val = new_maxqi;
- double old_q;
- double new_q;
- double target_q;
-
- old_q = vp9_convert_qindex_to_q(old_maxqi);
- new_q = vp9_convert_qindex_to_q(new_maxqi);
-
- target_q = ((old_q * 7.0) + new_q) / 8.0;
+ const double old_q = vp9_convert_qindex_to_q(old_maxqi);
+ const double new_q = vp9_convert_qindex_to_q(new_maxqi);
+ const double target_q = ((old_q * 7.0) + new_q) / 8.0;
if (target_q > old_q) {
- for (i = old_maxqi; i <= new_maxqi; i++) {
- if (vp9_convert_qindex_to_q(i) >= target_q) {
- ret_val = i;
- break;
- }
- }
+ for (i = old_maxqi; i <= new_maxqi; i++)
+ if (vp9_convert_qindex_to_q(i) >= target_q)
+ return i;
} else {
- for (i = old_maxqi; i >= new_maxqi; i--) {
- if (vp9_convert_qindex_to_q(i) <= target_q) {
- ret_val = i;
- break;
- }
- }
+ for (i = old_maxqi; i >= new_maxqi; i--)
+ if (vp9_convert_qindex_to_q(i) <= target_q)
+ return i;
}
- return ret_val;
+ return new_maxqi;
}
void vp9_second_pass(VP9_COMP *cpi) {
@@ -2111,9 +2035,8 @@
double this_frame_intra_error;
double this_frame_coded_error;
- if (!cpi->twopass.stats_in) {
+ if (!cpi->twopass.stats_in)
return;
- }
vp9_clear_system_state();
@@ -2123,12 +2046,8 @@
// Set a cq_level in constrained quality mode.
if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) {
- int est_cq;
-
- est_cq =
- estimate_cq(cpi,
- &cpi->twopass.total_left_stats,
- (int)(cpi->twopass.bits_left / frames_left));
+ int est_cq = estimate_cq(cpi, &cpi->twopass.total_left_stats,
+ (int)(cpi->twopass.bits_left / frames_left));
cpi->cq_target_quality = cpi->oxcf.cq_level;
if (est_cq > cpi->cq_target_quality)
@@ -2139,14 +2058,12 @@
cpi->twopass.maxq_max_limit = cpi->worst_quality;
cpi->twopass.maxq_min_limit = cpi->best_quality;
- tmp_q = estimate_max_q(
- cpi,
- &cpi->twopass.total_left_stats,
- (int)(cpi->twopass.bits_left / frames_left));
+ tmp_q = estimate_max_q(cpi, &cpi->twopass.total_left_stats,
+ (int)(cpi->twopass.bits_left / frames_left));
- cpi->active_worst_quality = tmp_q;
- cpi->ni_av_qi = tmp_q;
- cpi->avg_q = vp9_convert_qindex_to_q(tmp_q);
+ cpi->active_worst_quality = tmp_q;
+ cpi->ni_av_qi = tmp_q;
+ cpi->avg_q = vp9_convert_qindex_to_q(tmp_q);
#ifndef ONE_SHOT_Q_ESTIMATE
// Limit the maxq value returned subsequently.
@@ -2404,9 +2321,9 @@
if (cpi->oxcf.auto_key
&& lookup_next_frame_stats(cpi, &next_frame) != EOF) {
// Normal scene cut check
- if (test_candidate_kf(cpi, &last_frame, this_frame, &next_frame)) {
+ if (test_candidate_kf(cpi, &last_frame, this_frame, &next_frame))
break;
- }
+
// How fast is prediction quality decaying
loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
@@ -2416,19 +2333,14 @@
// quality since the last GF or KF.
recent_loop_decay[i % 8] = loop_decay_rate;
decay_accumulator = 1.0;
- for (j = 0; j < 8; j++) {
- decay_accumulator = decay_accumulator * recent_loop_decay[j];
- }
+ for (j = 0; j < 8; j++)
+ decay_accumulator *= recent_loop_decay[j];
// Special check for transition or high motion followed by a
// to a static scene.
- if (detect_transition_to_still(cpi, i,
- (cpi->key_frame_frequency - i),
- loop_decay_rate,
- decay_accumulator)) {
+ if (detect_transition_to_still(cpi, i, cpi->key_frame_frequency - i,
+ loop_decay_rate, decay_accumulator))
break;
- }
-
// Step on to the next frame
cpi->twopass.frames_to_key++;
diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c
index 610d733..2d3fea9 100644
--- a/vp9/encoder/vp9_onyx_if.c
+++ b/vp9/encoder/vp9_onyx_if.c
@@ -487,12 +487,14 @@
}
#if CONFIG_IMPLICIT_SEGMENTATION
-static void configure_implicit_segmentation(VP9_COMP *cpi) {
+static double implict_seg_q_modifiers[MAX_MB_SEGMENTS] =
+ {1.0, 0.95, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
+static void configure_implicit_segmentation(VP9_COMP *cpi, int frame_qindex) {
VP9_COMMON *cm = &cpi->common;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
int i;
int qi_delta;
- double q_target = cpi->active_worst_quality * 1.10;
+ double q_baseline = vp9_convert_qindex_to_q(frame_qindex);
// Set the flags to allow implicit segment update but disallow explicit update
xd->segmentation_enabled = 1;
@@ -507,13 +509,19 @@
// Clear down the segment features.
vp9_clearall_segfeatures(xd);
+ xd->update_mb_segmentation_data = 0;
+
+ // Update the segment data if it is an arf or non overlay gf.
+ } else if (cpi->refresh_alt_ref_frame ||
+ (cpi->refresh_golden_frame && !cpi->is_src_frame_alt_ref)) {
xd->update_mb_segmentation_data = 1;
// Enable use of q deltas on segments 1 and up
+ // Segment 0 is treated as a neutral segment with no changes
for (i = 1; i < MAX_MB_SEGMENTS; ++i) {
- qi_delta = compute_qdelta(cpi, cpi->active_worst_quality, q_target);
+ qi_delta = compute_qdelta(cpi, q_baseline,
+ implict_seg_q_modifiers[i] * q_baseline);
vp9_set_segdata(xd, i, SEG_LVL_ALT_Q, qi_delta);
- q_target *= 0.95;
vp9_enable_segfeature(xd, i, SEG_LVL_ALT_Q);
}
@@ -728,7 +736,7 @@
#if CONFIG_IMPLICIT_SEGMENTATION
sf->static_segmentation = 0;
#else
- sf->static_segmentation = 1;
+ sf->static_segmentation = 0;
#endif
#endif
sf->splitmode_breakout = 0;
@@ -747,7 +755,7 @@
#if CONFIG_IMPLICIT_SEGMENTATION
sf->static_segmentation = 0;
#else
- sf->static_segmentation = 1;
+ sf->static_segmentation = 0;
#endif
#endif
sf->splitmode_breakout = 1;
@@ -1589,11 +1597,15 @@
vp9_sub_pixel_avg_variance8x8, NULL, NULL, NULL,
vp9_sad8x8x3, vp9_sad8x8x8, vp9_sad8x8x4d)
- BFP(BLOCK_4X8, NULL, vp9_variance4x8, NULL,
- NULL, NULL, NULL, NULL, NULL, NULL, NULL)
+ BFP(BLOCK_8X4, vp9_sad8x4, vp9_variance8x4, vp9_sub_pixel_variance8x4,
+ vp9_sub_pixel_avg_variance8x4, NULL, NULL,
+ NULL, NULL, NULL,
+ vp9_sad8x4x4d)
- BFP(BLOCK_8X4, NULL, vp9_variance8x4, NULL,
- NULL, NULL, NULL, NULL, NULL, NULL, NULL)
+ BFP(BLOCK_4X8, vp9_sad4x8, vp9_variance4x8, vp9_sub_pixel_variance4x8,
+ vp9_sub_pixel_avg_variance4x8, NULL, NULL,
+ NULL, NULL, NULL,
+ vp9_sad4x8x4d)
BFP(BLOCK_4X4, vp9_sad4x4, vp9_variance4x4, vp9_sub_pixel_variance4x4,
vp9_sub_pixel_avg_variance4x4, NULL, NULL, NULL,
@@ -2128,49 +2140,31 @@
const int in_h = src_fb->y_crop_height;
const int out_w = dst_fb->y_crop_width;
const int out_h = dst_fb->y_crop_height;
- int x, y;
+ int x, y, i;
+
+ uint8_t *srcs[3] = {src_fb->y_buffer, src_fb->u_buffer, src_fb->v_buffer};
+ int src_strides[3] = {src_fb->y_stride, src_fb->uv_stride, src_fb->uv_stride};
+
+ uint8_t *dsts[3] = {dst_fb->y_buffer, dst_fb->u_buffer, dst_fb->v_buffer};
+ int dst_strides[3] = {dst_fb->y_stride, dst_fb->uv_stride, dst_fb->uv_stride};
for (y = 0; y < out_h; y += 16) {
for (x = 0; x < out_w; x += 16) {
- int x_q4 = x * 16 * in_w / out_w;
- int y_q4 = y * 16 * in_h / out_h;
- uint8_t *src = src_fb->y_buffer + y * in_h / out_h * src_fb->y_stride +
- x * in_w / out_w;
- uint8_t *dst = dst_fb->y_buffer + y * dst_fb->y_stride + x;
- int src_stride = src_fb->y_stride;
- int dst_stride = dst_fb->y_stride;
+ for (i = 0; i < MAX_MB_PLANE; ++i) {
+ const int factor = i == 0 ? 1 : 2;
+ const int x_q4 = x * (16 / factor) * in_w / out_w;
+ const int y_q4 = y * (16 / factor) * in_h / out_h;
+ const int src_stride = src_strides[i];
+ const int dst_stride = dst_strides[i];
+ uint8_t *src = srcs[i] + y / factor * in_h / out_h * src_stride +
+ x / factor * in_w / out_w;
+ uint8_t *dst = dsts[i] + y * dst_stride + x;
- vp9_convolve8(src, src_stride, dst, dst_stride,
- vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * in_w / out_w,
- vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * in_h / out_h,
- 16, 16);
-
- x_q4 >>= 1;
- y_q4 >>= 1;
- src_stride = src_fb->uv_stride;
- dst_stride = dst_fb->uv_stride;
-
- src = src_fb->u_buffer +
- y / 2 * in_h / out_h * src_fb->uv_stride +
- x / 2 * in_w / out_w;
- dst = dst_fb->u_buffer +
- y / 2 * dst_fb->uv_stride +
- x / 2;
- vp9_convolve8(src, src_stride, dst, dst_stride,
- vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * in_w / out_w,
- vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * in_h / out_h,
- 8, 8);
-
- src = src_fb->v_buffer +
- y / 2 * in_h / out_h * src_fb->uv_stride +
- x / 2 * in_w / out_w;
- dst = dst_fb->v_buffer +
- y / 2 * dst_fb->uv_stride +
- x / 2;
- vp9_convolve8(src, src_stride, dst, dst_stride,
- vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * in_w / out_w,
- vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * in_h / out_h,
- 8, 8);
+ vp9_convolve8(src, src_stride, dst, dst_stride,
+ vp9_sub_pel_filters_8[x_q4 & 0xf], 16 * in_w / out_w,
+ vp9_sub_pel_filters_8[y_q4 & 0xf], 16 * in_h / out_h,
+ 16 / factor, 16 / factor);
+ }
}
}
@@ -2457,7 +2451,8 @@
cm->dering_enabled);
}
- vp8_yv12_extend_frame_borders(cm->frame_to_show);
+ vp9_extend_frame_borders(cm->frame_to_show,
+ cm->subsampling_x, cm->subsampling_y);
}
@@ -2881,7 +2876,7 @@
break;
}
- vp9_denoise(cpi->Source, cpi->Source, l, 1, 0);
+ vp9_denoise(cpi->Source, cpi->Source, l);
}
#endif
@@ -2962,13 +2957,13 @@
cpi->common.frame_context_idx = cpi->refresh_alt_ref_frame;
vp9_setup_inter_frame(cpi);
}
+ }
#if CONFIG_IMPLICIT_SEGMENTATION
- if (!cm->error_resilient_mode && !cpi->sf.static_segmentation) {
- configure_implicit_segmentation(cpi);
- }
-#endif
+ if (!cm->error_resilient_mode && !cpi->sf.static_segmentation) {
+ configure_implicit_segmentation(cpi, q);
}
+#endif
// transform / motion compensation build reconstruction frame
#if CONFIG_MODELCOEFPROB
@@ -3878,16 +3873,8 @@
VP9BORDERINPIXELS);
// Calculate scaling factors for each of the 3 available references
- for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i) {
- if (cm->active_ref_idx[i] >= NUM_YV12_BUFFERS) {
- memset(&cm->active_ref_scale[i], 0, sizeof(cm->active_ref_scale[i]));
- } else {
- YV12_BUFFER_CONFIG *fb = &cm->yv12_fb[cm->active_ref_idx[i]];
- vp9_setup_scale_factors_for_frame(&cm->active_ref_scale[i],
- fb->y_crop_width, fb->y_crop_height,
- cm->width, cm->height);
- }
- }
+ for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i)
+ vp9_setup_scale_factors(cm, i);
vp9_setup_interp_filters(&cpi->mb.e_mbd, DEFAULT_INTERP_FILTER, cm);
@@ -3966,7 +3953,7 @@
double weight = 0;
#if CONFIG_POSTPROC
vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
- cm->filter_level * 10 / 6, 1, 0);
+ cm->filter_level * 10 / 6);
#endif
vp9_clear_system_state();
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 765a071..f928e7a 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -689,7 +689,11 @@
int *Distortion, int64_t best_rd) {
int i;
MACROBLOCKD *const xd = &mb->e_mbd;
+#if CONFIG_AB4X4
+ int cost = 0;
+#else
int cost = mb->mbmode_cost[xd->frame_type][I4X4_PRED];
+#endif
int distortion = 0;
int tot_rate_y = 0;
int64_t total_rd = 0;
@@ -719,7 +723,6 @@
total_rd += rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
t_above + x_idx, t_left + y_idx,
&r, &ry, &d);
-
cost += r;
distortion += d;
tot_rate_y += ry;
@@ -753,6 +756,13 @@
TX_SIZE UNINITIALIZED_IS_SAFE(best_tx);
int i;
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8) {
+ x->e_mbd.mode_info_context->mbmi.txfm_size = TX_4X4;
+ return best_rd;
+ }
+#endif
+
for (i = 0; i < NB_TXFM_MODES; i++)
txfm_cache[i] = INT64_MAX;
@@ -1086,6 +1096,50 @@
return r;
}
+static enum BlockSize get_block_size(int bw, int bh) {
+ if (bw == 4 && bh == 4)
+ return BLOCK_4X4;
+
+ if (bw == 4 && bh == 8)
+ return BLOCK_4X8;
+
+ if (bw == 8 && bh == 4)
+ return BLOCK_8X4;
+
+ if (bw == 8 && bh == 8)
+ return BLOCK_8X8;
+
+ if (bw == 8 && bh == 16)
+ return BLOCK_8X16;
+
+ if (bw == 16 && bh == 8)
+ return BLOCK_16X8;
+
+ if (bw == 16 && bh == 16)
+ return BLOCK_16X16;
+
+ if (bw == 32 && bh == 32)
+ return BLOCK_32X32;
+
+ if (bw == 32 && bh == 16)
+ return BLOCK_32X16;
+
+ if (bw == 16 && bh == 32)
+ return BLOCK_16X32;
+
+ if (bw == 64 && bh == 32)
+ return BLOCK_64X32;
+
+ if (bw == 32 && bh == 64)
+ return BLOCK_32X64;
+
+ if (bw == 64 && bh == 64)
+ return BLOCK_64X64;
+
+ assert(0);
+ return -1;
+}
+
static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x,
BEST_SEG_INFO *bsi,
int_mv seg_mvs[4][MAX_REF_FRAMES - 1]) {
@@ -1101,6 +1155,10 @@
int sbr = 0, sbd = 0;
int segmentyrate = 0;
int best_eobs[4] = { 0 };
+#if CONFIG_AB4X4
+ BLOCK_SIZE_TYPE bsize = mbmi->sb_type;
+ int bwl = b_width_log2(bsize), bhl = b_height_log2(bsize);
+#endif
vp9_variance_fn_ptr_t *v_fn_ptr;
@@ -1110,7 +1168,11 @@
vpx_memcpy(t_above, x->e_mbd.plane[0].above_context, sizeof(t_above));
vpx_memcpy(t_left, x->e_mbd.plane[0].left_context, sizeof(t_left));
+#if CONFIG_AB4X4
+ v_fn_ptr = &cpi->fn_ptr[get_block_size(4 << bwl, 4 << bhl)];
+#else
v_fn_ptr = &cpi->fn_ptr[BLOCK_4X4];
+#endif
// 64 makes this threshold really big effectively
// making it so that we very rarely check mvs on
@@ -1660,51 +1722,6 @@
frame_type, block_size);
}
-
-static enum BlockSize get_block_size(int bw, int bh) {
- if (bw == 4 && bh == 4)
- return BLOCK_4X4;
-
- if (bw == 4 && bh == 8)
- return BLOCK_4X8;
-
- if (bw == 8 && bh == 4)
- return BLOCK_8X4;
-
- if (bw == 8 && bh == 8)
- return BLOCK_8X8;
-
- if (bw == 8 && bh == 16)
- return BLOCK_8X16;
-
- if (bw == 16 && bh == 8)
- return BLOCK_16X8;
-
- if (bw == 16 && bh == 16)
- return BLOCK_16X16;
-
- if (bw == 32 && bh == 32)
- return BLOCK_32X32;
-
- if (bw == 32 && bh == 16)
- return BLOCK_32X16;
-
- if (bw == 16 && bh == 32)
- return BLOCK_16X32;
-
- if (bw == 64 && bh == 32)
- return BLOCK_64X32;
-
- if (bw == 32 && bh == 64)
- return BLOCK_32X64;
-
- if (bw == 64 && bh == 64)
- return BLOCK_64X64;
-
- assert(0);
- return -1;
-}
-
static void model_rd_from_var_lapndz(int var, int n, int qstep,
int *rate, int *dist) {
// This function models the rate and distortion for a Laplacian
@@ -1863,6 +1880,7 @@
struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
struct buf_2d backup_second_yv12[MAX_MB_PLANE] = {{0}};
struct buf_2d scaled_first_yv12;
+ int last_besterr[2] = {INT_MAX, INT_MAX};
if (scaled_ref_frame[0]) {
int i;
@@ -1897,11 +1915,9 @@
frame_mv[NEWMV][refs[0]].as_int = single_newmv[refs[0]].as_int;
frame_mv[NEWMV][refs[1]].as_int = single_newmv[refs[1]].as_int;
- // Iteration: joint search is done once for each ref frame.
- // Tried allowing search multiple times iteratively, and break out if
- // it couldn't find better mv. But tests didn't show noticeable
- // improvement.
- for (ite = 0; ite < 2; ite++) {
+ // Allow joint search multiple times iteratively for each ref frame, and
+ // break out the search loop if it couldn't find better mv.
+ for (ite = 0; ite < 4; ite++) {
struct buf_2d ref_yv12[2] = {xd->plane[0].pre[0],
xd->plane[0].pre[1]};
int bestsme = INT_MAX;
@@ -1952,19 +1968,26 @@
int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
- vp9_find_best_sub_pixel_comp(x, &tmp_mv,
- &ref_mv[id],
- x->errorperbit,
- &cpi->fn_ptr[block_size],
- x->nmvjointcost, x->mvcost,
- &dis, &sse, second_pred,
- b_sz[bsize][0], b_sz[bsize][1]);
+ bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv,
+ &ref_mv[id],
+ x->errorperbit,
+ &cpi->fn_ptr[block_size],
+ x->nmvjointcost, x->mvcost,
+ &dis, &sse, second_pred,
+ b_sz[bsize][0],
+ b_sz[bsize][1]);
}
- frame_mv[NEWMV][refs[id]].as_int =
- xd->mode_info_context->bmi[0].as_mv[1].as_int = tmp_mv.as_int;
if (id)
xd->plane[0].pre[0] = scaled_first_yv12;
+
+ if (bestsme < last_besterr[id]) {
+ frame_mv[NEWMV][refs[id]].as_int =
+ xd->mode_info_context->bmi[0].as_mv[1].as_int = tmp_mv.as_int;
+ last_besterr[id] = bestsme;
+ } else {
+ break;
+ }
}
// restore the predictor
@@ -2302,7 +2325,11 @@
&dist_uv, &uv_skip,
(bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 :
bsize);
+#if CONFIG_AB4X4
+ if (bsize < BLOCK_SIZE_SB8X8)
+#else
if (bsize == BLOCK_SIZE_SB8X8)
+#endif
err4x4 = rd_pick_intra4x4mby_modes(cpi, x, &rate4x4_y,
&rate4x4_y_tokenonly,
&dist4x4_y, err);
@@ -2311,11 +2338,14 @@
*returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
vp9_cost_bit(vp9_get_pred_prob(cm, xd, PRED_MBSKIP), 1);
*returndist = dist_y + (dist_uv >> 2);
- memset(ctx->txfm_rd_diff, 0,
- sizeof(x->sb32_context[xd->sb_index].txfm_rd_diff));
+ memset(ctx->txfm_rd_diff, 0, sizeof(ctx->txfm_rd_diff));
xd->mode_info_context->mbmi.mode = mode;
xd->mode_info_context->mbmi.txfm_size = txfm_size;
+#if CONFIG_AB4X4
+ } else if (bsize < BLOCK_SIZE_SB8X8 && err4x4 < err) {
+#else
} else if (bsize == BLOCK_SIZE_SB8X8 && err4x4 < err) {
+#endif
*returnrate = rate4x4_y + rate_uv +
vp9_cost_bit(vp9_get_pred_prob(cm, xd, PRED_MBSKIP), 0);
*returndist = dist4x4_y + (dist_uv >> 2);
@@ -2457,7 +2487,9 @@
i++) {
mbmi->txfm_size = i;
rd_pick_intra_sbuv_mode(cpi, x, &rate_uv_intra[i], &rate_uv_tokenonly[i],
- &dist_uv[i], &skip_uv[i], bsize);
+ &dist_uv[i], &skip_uv[i],
+ (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 :
+ bsize);
mode_uv[i] = mbmi->uv_mode;
}
}
@@ -2487,6 +2519,7 @@
|| (cpi->ref_frame_flags & flag_list[ref_frame]))) {
continue;
}
+
if (cpi->speed > 0) {
if (!(ref_frame_mask & (1 << ref_frame))) {
continue;
@@ -2533,10 +2566,18 @@
mbmi->interp_filter = cm->mcomp_filter_type;
vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
+#if CONFIG_AB4X4
+ if (bsize >= BLOCK_SIZE_SB8X8 &&
+ (this_mode == I4X4_PRED || this_mode == SPLITMV))
+ continue;
+ if (bsize < BLOCK_SIZE_SB8X8 &&
+ !(this_mode == I4X4_PRED || this_mode == SPLITMV))
+ continue;
+#else
if (bsize != BLOCK_SIZE_SB8X8 &&
(this_mode == I4X4_PRED || this_mode == SPLITMV))
continue;
-
+#endif
if (comp_pred) {
if (ref_frame == ALTREF_FRAME) {
@@ -2599,7 +2640,6 @@
// Note the rate value returned here includes the cost of coding
// the I4X4_PRED mode : x->mbmode_cost[xd->frame_type][I4X4_PRED];
- assert(bsize == BLOCK_SIZE_SB8X8);
mbmi->txfm_size = TX_4X4;
rd_pick_intra4x4mby_modes(cpi, x, &rate, &rate_y,
&distortion_y, INT64_MAX);
@@ -2995,7 +3035,13 @@
}
}
-
+#if CONFIG_AB4X4
+ if (best_rd == INT64_MAX && bsize < BLOCK_SIZE_SB8X8) {
+ *returnrate = INT_MAX;
+ *returndistortion = INT_MAX;
+ return best_rd;
+ }
+#endif
assert((cm->mcomp_filter_type == SWITCHABLE) ||
(cm->mcomp_filter_type == best_mbmode.interp_filter) ||
diff --git a/vp9/encoder/vp9_sad_c.c b/vp9/encoder/vp9_sad_c.c
index b4cd193..994828f 100644
--- a/vp9/encoder/vp9_sad_c.c
+++ b/vp9/encoder/vp9_sad_c.c
@@ -156,6 +156,21 @@
return sad_mx_n_c(src_ptr, src_stride, ref_ptr, ref_stride, 8, 16);
}
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr,
+ int src_stride,
+ const uint8_t *ref_ptr,
+ int ref_stride,
+ unsigned int max_sad) {
+ return sad_mx_n_c(src_ptr, src_stride, ref_ptr, ref_stride, 8, 4);
+}
+
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr,
+ int src_stride,
+ const uint8_t *ref_ptr,
+ int ref_stride,
+ unsigned int max_sad) {
+ return sad_mx_n_c(src_ptr, src_stride, ref_ptr, ref_stride, 4, 8);
+}
unsigned int vp9_sad4x4_c(const uint8_t *src_ptr,
int src_stride,
@@ -563,6 +578,36 @@
ref_ptr[3], ref_stride, 0x7fffffff);
}
+void vp9_sad8x4x4d_c(const uint8_t *src_ptr,
+ int src_stride,
+ const uint8_t* const ref_ptr[],
+ int ref_stride,
+ unsigned int *sad_array) {
+ sad_array[0] = vp9_sad8x4(src_ptr, src_stride,
+ ref_ptr[0], ref_stride, 0x7fffffff);
+ sad_array[1] = vp9_sad8x4(src_ptr, src_stride,
+ ref_ptr[1], ref_stride, 0x7fffffff);
+ sad_array[2] = vp9_sad8x4(src_ptr, src_stride,
+ ref_ptr[2], ref_stride, 0x7fffffff);
+ sad_array[3] = vp9_sad8x4(src_ptr, src_stride,
+ ref_ptr[3], ref_stride, 0x7fffffff);
+}
+
+void vp9_sad4x8x4d_c(const uint8_t *src_ptr,
+ int src_stride,
+ const uint8_t* const ref_ptr[],
+ int ref_stride,
+ unsigned int *sad_array) {
+ sad_array[0] = vp9_sad4x8(src_ptr, src_stride,
+ ref_ptr[0], ref_stride, 0x7fffffff);
+ sad_array[1] = vp9_sad4x8(src_ptr, src_stride,
+ ref_ptr[1], ref_stride, 0x7fffffff);
+ sad_array[2] = vp9_sad4x8(src_ptr, src_stride,
+ ref_ptr[2], ref_stride, 0x7fffffff);
+ sad_array[3] = vp9_sad4x8(src_ptr, src_stride,
+ ref_ptr[3], ref_stride, 0x7fffffff);
+}
+
void vp9_sad4x4x4d_c(const uint8_t *src_ptr,
int src_stride,
const uint8_t* const ref_ptr[],
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 4420d49..50d849d 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -119,7 +119,12 @@
TOKENEXTRA *t = *tp; /* store tokens starting here */
const int eob = xd->plane[plane].eobs[block];
const int16_t *qcoeff_ptr = BLOCK_OFFSET(xd->plane[plane].qcoeff, block, 16);
+#if CONFIG_AB4X4
+ const BLOCK_SIZE_TYPE sb_type = (mbmi->sb_type < BLOCK_SIZE_SB8X8) ?
+ BLOCK_SIZE_SB8X8 : mbmi->sb_type;
+#else
const BLOCK_SIZE_TYPE sb_type = mbmi->sb_type;
+#endif
const int bwl = b_width_log2(sb_type);
const int off = block >> (2 * tx_size);
const int mod = bwl - tx_size - xd->plane[plane].subsampling_x;
diff --git a/vp9/encoder/vp9_variance_c.c b/vp9/encoder/vp9_variance_c.c
index fa53abd..e24a46b 100644
--- a/vp9/encoder/vp9_variance_c.c
+++ b/vp9/encoder/vp9_variance_c.c
@@ -820,3 +820,91 @@
comp_avg_pred(temp3, second_pred, 8, 16, temp2, 8);
return vp9_variance8x16_c(temp3, 8, dst_ptr, dst_pixels_per_line, sse);
}
+
+unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr,
+ int src_pixels_per_line,
+ int xoffset,
+ int yoffset,
+ const uint8_t *dst_ptr,
+ int dst_pixels_per_line,
+ unsigned int *sse) {
+ uint16_t fdata3[8 * 5]; // Temp data bufffer used in filtering
+ uint8_t temp2[20 * 16];
+ const int16_t *hfilter, *vfilter;
+
+ hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset);
+ vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset);
+
+ var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
+ 1, 5, 8, hfilter);
+ var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter);
+
+ return vp9_variance8x4_c(temp2, 8, dst_ptr, dst_pixels_per_line, sse);
+}
+
+unsigned int vp9_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr,
+ int src_pixels_per_line,
+ int xoffset,
+ int yoffset,
+ const uint8_t *dst_ptr,
+ int dst_pixels_per_line,
+ unsigned int *sse,
+ const uint8_t *second_pred) {
+ uint16_t fdata3[8 * 5]; // Temp data bufffer used in filtering
+ uint8_t temp2[20 * 16];
+ DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 8 * 4); // compound pred buffer
+ const int16_t *hfilter, *vfilter;
+
+ hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset);
+ vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset);
+
+ var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
+ 1, 5, 8, hfilter);
+ var_filter_block2d_bil_second_pass(fdata3, temp2, 8, 8, 4, 8, vfilter);
+ comp_avg_pred(temp3, second_pred, 8, 4, temp2, 8);
+ return vp9_variance8x4_c(temp3, 8, dst_ptr, dst_pixels_per_line, sse);
+}
+
+unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr,
+ int src_pixels_per_line,
+ int xoffset,
+ int yoffset,
+ const uint8_t *dst_ptr,
+ int dst_pixels_per_line,
+ unsigned int *sse) {
+ uint16_t fdata3[5 * 8]; // Temp data bufffer used in filtering
+ uint8_t temp2[20 * 16];
+ const int16_t *hfilter, *vfilter;
+
+ hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset);
+ vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset);
+
+ var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
+ 1, 17, 4, hfilter);
+ var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter);
+
+ return vp9_variance4x8_c(temp2, 4, dst_ptr, dst_pixels_per_line, sse);
+}
+
+unsigned int vp9_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr,
+ int src_pixels_per_line,
+ int xoffset,
+ int yoffset,
+ const uint8_t *dst_ptr,
+ int dst_pixels_per_line,
+ unsigned int *sse,
+ const uint8_t *second_pred) {
+ uint16_t fdata3[5 * 8]; // Temp data bufffer used in filtering
+ uint8_t temp2[20 * 16];
+ DECLARE_ALIGNED_ARRAY(16, uint8_t, temp3, 4 * 8); // compound pred buffer
+ const int16_t *hfilter, *vfilter;
+
+ hfilter = VP9_BILINEAR_FILTERS_2TAP(xoffset);
+ vfilter = VP9_BILINEAR_FILTERS_2TAP(yoffset);
+
+ var_filter_block2d_bil_first_pass(src_ptr, fdata3, src_pixels_per_line,
+ 1, 17, 4, hfilter);
+ var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 8, 4, vfilter);
+ comp_avg_pred(temp3, second_pred, 4, 8, temp2, 4);
+ return vp9_variance4x8_c(temp3, 4, dst_ptr, dst_pixels_per_line, sse);
+}
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index 45609da..9326165 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -543,31 +543,6 @@
return VPX_CODEC_OK;
}
-static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
- YV12_BUFFER_CONFIG *yv12) {
- vpx_codec_err_t res = VPX_CODEC_OK;
- yv12->y_buffer = img->planes[VPX_PLANE_Y];
- yv12->u_buffer = img->planes[VPX_PLANE_U];
- yv12->v_buffer = img->planes[VPX_PLANE_V];
-
- yv12->y_crop_width = img->d_w;
- yv12->y_crop_height = img->d_h;
- yv12->y_width = img->d_w;
- yv12->y_height = img->d_h;
-
- yv12->uv_width = img->x_chroma_shift == 1 ? (1 + yv12->y_width) / 2
- : yv12->y_width;
- yv12->uv_height = img->y_chroma_shift == 1 ? (1 + yv12->y_height) / 2
- : yv12->y_height;
-
- yv12->y_stride = img->stride[VPX_PLANE_Y];
- yv12->uv_stride = img->stride[VPX_PLANE_U];
-
- yv12->border = (img->stride[VPX_PLANE_Y] - img->w) / 2;
- yv12->clrtype = REG_YUV;
- return res;
-}
-
static void pick_quickcompress_mode(vpx_codec_alg_priv_t *ctx,
unsigned long duration,
unsigned long deadline) {
diff --git a/vp9/vp9_dx_iface.c b/vp9/vp9_dx_iface.c
index 85022c9..811cea7 100644
--- a/vp9/vp9_dx_iface.c
+++ b/vp9/vp9_dx_iface.c
@@ -578,30 +578,6 @@
return res;
}
-static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
- YV12_BUFFER_CONFIG *yv12) {
- vpx_codec_err_t res = VPX_CODEC_OK;
- yv12->y_buffer = img->planes[VPX_PLANE_Y];
- yv12->u_buffer = img->planes[VPX_PLANE_U];
- yv12->v_buffer = img->planes[VPX_PLANE_V];
-
- yv12->y_crop_width = img->d_w;
- yv12->y_crop_height = img->d_h;
- yv12->y_width = img->d_w;
- yv12->y_height = img->d_h;
- yv12->uv_width = yv12->y_width / 2;
- yv12->uv_height = yv12->y_height / 2;
-
- yv12->y_stride = img->stride[VPX_PLANE_Y];
- yv12->uv_stride = img->stride[VPX_PLANE_U];
-
- yv12->border = (img->stride[VPX_PLANE_Y] - img->d_w) / 2;
- yv12->clrtype = (img->fmt == VPX_IMG_FMT_VPXI420 ||
- img->fmt == VPX_IMG_FMT_VPXYV12);
-
- return res;
-}
-
static vpx_codec_err_t vp9_set_reference(vpx_codec_alg_priv_t *ctx,
int ctr_id,
diff --git a/vp9/vp9_iface_common.h b/vp9/vp9_iface_common.h
index 96de5f5..84b4d39 100644
--- a/vp9/vp9_iface_common.h
+++ b/vp9/vp9_iface_common.h
@@ -37,11 +37,11 @@
img->planes[VPX_PLANE_Y] = yv12->y_buffer;
img->planes[VPX_PLANE_U] = yv12->u_buffer;
img->planes[VPX_PLANE_V] = yv12->v_buffer;
- img->planes[VPX_PLANE_ALPHA] = NULL;
+ img->planes[VPX_PLANE_ALPHA] = yv12->alpha_buffer;
img->stride[VPX_PLANE_Y] = yv12->y_stride;
img->stride[VPX_PLANE_U] = yv12->uv_stride;
img->stride[VPX_PLANE_V] = yv12->uv_stride;
- img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
+ img->stride[VPX_PLANE_ALPHA] = yv12->alpha_stride;
img->bps = bps;
img->user_priv = user_priv;
img->img_data = yv12->buffer_alloc;
@@ -49,4 +49,34 @@
img->self_allocd = 0;
}
+static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
+ YV12_BUFFER_CONFIG *yv12) {
+ yv12->y_buffer = img->planes[VPX_PLANE_Y];
+ yv12->u_buffer = img->planes[VPX_PLANE_U];
+ yv12->v_buffer = img->planes[VPX_PLANE_V];
+ yv12->alpha_buffer = img->planes[VPX_PLANE_ALPHA];
+
+ yv12->y_crop_width = img->d_w;
+ yv12->y_crop_height = img->d_h;
+ yv12->y_width = img->d_w;
+ yv12->y_height = img->d_h;
+
+ yv12->uv_width = img->x_chroma_shift == 1 ? (1 + yv12->y_width) / 2
+ : yv12->y_width;
+ yv12->uv_height = img->y_chroma_shift == 1 ? (1 + yv12->y_height) / 2
+ : yv12->y_height;
+
+ yv12->alpha_width = yv12->alpha_buffer ? img->d_w : 0;
+ yv12->alpha_height = yv12->alpha_buffer ? img->d_h : 0;
+
+ yv12->y_stride = img->stride[VPX_PLANE_Y];
+ yv12->uv_stride = img->stride[VPX_PLANE_U];
+ yv12->alpha_stride = yv12->alpha_buffer ? img->stride[VPX_PLANE_ALPHA] : 0;
+
+ yv12->border = (img->stride[VPX_PLANE_Y] - img->w) / 2;
+ yv12->clrtype = REG_YUV;
+
+ return VPX_CODEC_OK;
+}
+
#endif // VP9_VP9_IFACE_COMMON_H_
diff --git a/vpx_scale/generic/yv12config.c b/vpx_scale/generic/yv12config.c
index cd66f00..99e3543 100644
--- a/vpx_scale/generic/yv12config.c
+++ b/vpx_scale/generic/yv12config.c
@@ -76,12 +76,17 @@
ybf->uv_height = uv_height;
ybf->uv_stride = uv_stride;
+ ybf->alpha_width = 0;
+ ybf->alpha_height = 0;
+ ybf->alpha_stride = 0;
+
ybf->border = border;
ybf->frame_size = frame_size;
ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
ybf->u_buffer = ybf->buffer_alloc + yplane_size + (border / 2 * uv_stride) + border / 2;
ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size + (border / 2 * uv_stride) + border / 2;
+ ybf->alpha_buffer = NULL;
ybf->corrupted = 0; /* assume not currupted by errors */
return 0;
diff --git a/vpx_scale/generic/yv12extend.c b/vpx_scale/generic/yv12extend.c
index a322e0a..c38fb80 100644
--- a/vpx_scale/generic/yv12extend.c
+++ b/vpx_scale/generic/yv12extend.c
@@ -9,6 +9,7 @@
*/
#include <assert.h>
+#include "./vpx_config.h"
#include "vpx_scale/yv12config.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_scale/vpx_scale.h"
@@ -94,6 +95,36 @@
(ybf->border + ybf->y_width - ybf->y_crop_width + 1) / 2);
}
+#if CONFIG_VP9
+void vp9_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf,
+ int subsampling_x, int subsampling_y) {
+ const int c_w = (ybf->y_crop_width + subsampling_x) >> subsampling_x;
+ const int c_h = (ybf->y_crop_height + subsampling_y) >> subsampling_y;
+ const int c_et = ybf->border >> subsampling_y;
+ const int c_el = ybf->border >> subsampling_x;
+ const int c_eb = (ybf->border + ybf->y_height - ybf->y_crop_height +
+ subsampling_y) >> subsampling_y;
+ const int c_er = (ybf->border + ybf->y_width - ybf->y_crop_width +
+ subsampling_x) >> subsampling_x;
+
+ assert(ybf->y_height - ybf->y_crop_height < 16);
+ assert(ybf->y_width - ybf->y_crop_width < 16);
+ assert(ybf->y_height - ybf->y_crop_height >= 0);
+ assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+ extend_plane(ybf->y_buffer, ybf->y_stride,
+ ybf->y_crop_width, ybf->y_crop_height,
+ ybf->border, ybf->border,
+ ybf->border + ybf->y_height - ybf->y_crop_height,
+ ybf->border + ybf->y_width - ybf->y_crop_width);
+
+ extend_plane(ybf->u_buffer, ybf->uv_stride,
+ c_w, c_h, c_et, c_el, c_eb, c_er);
+
+ extend_plane(ybf->v_buffer, ybf->uv_stride,
+ c_w, c_h, c_et, c_el, c_eb, c_er);
+}
+#endif
/****************************************************************************
*
diff --git a/vpx_scale/vpx_scale_rtcd.sh b/vpx_scale/vpx_scale_rtcd.sh
index e2bade0..b4f8907 100644
--- a/vpx_scale/vpx_scale_rtcd.sh
+++ b/vpx_scale/vpx_scale_rtcd.sh
@@ -24,3 +24,8 @@
prototype void vp8_yv12_copy_y "struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc"
specialize vp8_yv12_copy_y neon
+
+if [ "$CONFIG_VP9" = "yes" ]; then
+ prototype void vp9_extend_frame_borders "struct yv12_buffer_config *ybf, int subsampling_x, int subsampling_y"
+ specialize vp9_extend_frame_borders
+fi
diff --git a/vpx_scale/yv12config.h b/vpx_scale/yv12config.h
index 85396c0..7b8bd85 100644
--- a/vpx_scale/yv12config.h
+++ b/vpx_scale/yv12config.h
@@ -52,9 +52,14 @@
int uv_stride;
/* int uvinternal_width; */
+ int alpha_width;
+ int alpha_height;
+ int alpha_stride;
+
uint8_t *y_buffer;
uint8_t *u_buffer;
uint8_t *v_buffer;
+ uint8_t *alpha_buffer;
uint8_t *buffer_alloc;
int buffer_alloc_sz;
