Merge "removed the idct rtcd idct calls" into experimental
diff --git a/vp9/decoder/decodframe.c b/vp9/decoder/decodframe.c
index a579894..625053e 100644
--- a/vp9/decoder/decodframe.c
+++ b/vp9/decoder/decodframe.c
@@ -288,13 +288,7 @@
xd->eobs[i] = 0;
}
- if (tx_size == TX_16X16) {
- eobtotal = vp9_decode_mb_tokens_16x16(pbi, xd, bc);
- } else if (tx_size == TX_8X8) {
- eobtotal = vp9_decode_mb_tokens_8x8(pbi, xd, bc);
- } else {
- eobtotal = vp9_decode_mb_tokens_4x4(pbi, xd, bc);
- }
+ eobtotal = vp9_decode_mb_tokens(pbi, xd, bc);
if (eobtotal == 0) { // skip loopfilter
xd->mode_info_context->mbmi.mb_skip_coeff = 1;
continue;
@@ -392,12 +386,8 @@
xd->block[i].eob = 0;
xd->eobs[i] = 0;
}
- if (tx_size == TX_16X16) {
- eobtotal = vp9_decode_mb_tokens_16x16(pbi, xd, bc);
- } else if (tx_size == TX_8X8) {
- eobtotal = vp9_decode_mb_tokens_8x8(pbi, xd, bc);
- } else if (mode != B_PRED) {
- eobtotal = vp9_decode_mb_tokens_4x4(pbi, xd, bc);
+ if (mode != B_PRED) {
+ eobtotal = vp9_decode_mb_tokens(pbi, xd, bc);
}
}
diff --git a/vp9/decoder/detokenize.c b/vp9/decoder/detokenize.c
index 1194aa8..b6823e7 100644
--- a/vp9/decoder/detokenize.c
+++ b/vp9/decoder/detokenize.c
@@ -124,14 +124,14 @@
PLANE_TYPE type,
TX_TYPE tx_type,
int seg_eob, INT16 *qcoeff_ptr,
- const int *const scan, int block_type,
+ const int *const scan, TX_SIZE txfm_size,
const int *coef_bands) {
FRAME_CONTEXT *const fc = &dx->common.fc;
int pt, c = (type == PLANE_TYPE_Y_NO_DC);
vp9_prob (*coef_probs)[PREV_COEF_CONTEXTS][ENTROPY_NODES], *prob;
unsigned int (*coef_counts)[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS];
- switch (block_type) {
+ switch (txfm_size) {
default:
case TX_4X4:
if (tx_type == DCT_DCT) {
@@ -246,8 +246,7 @@
return c;
}
-
-int get_eob(MACROBLOCKD* const xd, int segment_id, int eob_max) {
+static int get_eob(MACROBLOCKD* const xd, int segment_id, int eob_max) {
int active = vp9_segfeature_active(xd, segment_id, SEG_LVL_EOB);
int eob = vp9_get_segdata(xd, segment_id, SEG_LVL_EOB);
@@ -257,132 +256,117 @@
}
-int vp9_decode_mb_tokens_16x16(VP9D_COMP* const pbi,
- MACROBLOCKD* const xd,
- BOOL_DECODER* const bc) {
+static int vp9_decode_mb_tokens_16x16(VP9D_COMP* const pbi,
+ MACROBLOCKD* const xd,
+ BOOL_DECODER* const bc) {
ENTROPY_CONTEXT* const A = (ENTROPY_CONTEXT *)xd->above_context;
ENTROPY_CONTEXT* const L = (ENTROPY_CONTEXT *)xd->left_context;
-
unsigned short* const eobs = xd->eobs;
- PLANE_TYPE type;
- int c, i, eobtotal = 0, seg_eob;
const int segment_id = xd->mode_info_context->mbmi.segment_id;
- INT16 *qcoeff_ptr = &xd->qcoeff[0];
- TX_TYPE tx_type = get_tx_type(xd, &xd->block[0]);
-
- type = PLANE_TYPE_Y_WITH_DC;
- seg_eob = get_eob(xd, segment_id, 256);
+ int c, i, eobtotal = 0, seg_eob;
// Luma block
- {
- const int* const scan = vp9_default_zig_zag1d_16x16;
- eobs[0] = c = decode_coefs(pbi, xd, bc, A, L, type,
- tx_type, seg_eob, qcoeff_ptr,
- scan, TX_16X16, vp9_coef_bands_16x16);
- A[1] = A[2] = A[3] = A[0];
- L[1] = L[2] = L[3] = L[0];
- eobtotal += c;
- }
+ eobs[0] = c = decode_coefs(pbi, xd, bc, A, L, PLANE_TYPE_Y_WITH_DC,
+ get_tx_type(xd, &xd->block[0]),
+ get_eob(xd, segment_id, 256),
+ xd->qcoeff, vp9_default_zig_zag1d_16x16,
+ TX_16X16, vp9_coef_bands_16x16);
+ A[1] = A[2] = A[3] = A[0];
+ L[1] = L[2] = L[3] = L[0];
+ eobtotal += c;
// 8x8 chroma blocks
- qcoeff_ptr += 256;
- type = PLANE_TYPE_UV;
- tx_type = DCT_DCT;
seg_eob = get_eob(xd, segment_id, 64);
for (i = 16; i < 24; i += 4) {
ENTROPY_CONTEXT* const a = A + vp9_block2above_8x8[i];
ENTROPY_CONTEXT* const l = L + vp9_block2left_8x8[i];
- const int* const scan = vp9_default_zig_zag1d_8x8;
- eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, type,
- tx_type, seg_eob, qcoeff_ptr,
- scan, TX_8X8, vp9_coef_bands_8x8);
+ eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, PLANE_TYPE_UV,
+ DCT_DCT, seg_eob, xd->block[i].qcoeff,
+ vp9_default_zig_zag1d_8x8,
+ TX_8X8, vp9_coef_bands_8x8);
a[1] = a[0];
l[1] = l[0];
-
eobtotal += c;
- qcoeff_ptr += 64;
}
+
+ // no Y2 block
vpx_memset(&A[8], 0, sizeof(A[8]));
vpx_memset(&L[8], 0, sizeof(L[8]));
+
return eobtotal;
}
-int vp9_decode_mb_tokens_8x8(VP9D_COMP* const pbi,
- MACROBLOCKD* const xd,
- BOOL_DECODER* const bc) {
+static int vp9_decode_mb_tokens_8x8(VP9D_COMP* const pbi,
+ MACROBLOCKD* const xd,
+ BOOL_DECODER* const bc) {
ENTROPY_CONTEXT *const A = (ENTROPY_CONTEXT *)xd->above_context;
ENTROPY_CONTEXT *const L = (ENTROPY_CONTEXT *)xd->left_context;
-
unsigned short *const eobs = xd->eobs;
PLANE_TYPE type;
int c, i, eobtotal = 0, seg_eob;
const int segment_id = xd->mode_info_context->mbmi.segment_id;
- INT16 *qcoeff_ptr = &xd->qcoeff[0];
- TX_TYPE tx_type = DCT_DCT;
- int bufthred = (xd->mode_info_context->mbmi.mode == I8X8_PRED ||
- xd->mode_info_context->mbmi.mode == SPLITMV) ? 16 : 24;
+ // 2nd order DC block
if (xd->mode_info_context->mbmi.mode != B_PRED &&
xd->mode_info_context->mbmi.mode != SPLITMV &&
xd->mode_info_context->mbmi.mode != I8X8_PRED) {
ENTROPY_CONTEXT *const a = A + vp9_block2above_8x8[24];
ENTROPY_CONTEXT *const l = L + vp9_block2left_8x8[24];
- const int *const scan = vp9_default_zig_zag1d;
- type = PLANE_TYPE_Y2;
- seg_eob = get_eob(xd, segment_id, 4);
- eobs[24] = c = decode_coefs(pbi, xd, bc, a, l, type,
- tx_type, seg_eob, qcoeff_ptr + 24 * 16,
- scan, TX_8X8, vp9_coef_bands);
-
+ eobs[24] = c = decode_coefs(pbi, xd, bc, a, l, PLANE_TYPE_Y2,
+ DCT_DCT, get_eob(xd, segment_id, 4),
+ xd->block[24].qcoeff,
+ vp9_default_zig_zag1d, TX_8X8, vp9_coef_bands);
eobtotal += c - 4;
-
type = PLANE_TYPE_Y_NO_DC;
- } else
+ } else {
type = PLANE_TYPE_Y_WITH_DC;
-
- seg_eob = get_eob(xd, segment_id, 64);
-
- for (i = 0; i < bufthred ; i += 4) {
- ENTROPY_CONTEXT *const a = A + vp9_block2above_8x8[i];
- ENTROPY_CONTEXT *const l = L + vp9_block2left_8x8[i];
- const int *const scan = vp9_default_zig_zag1d_8x8;
- tx_type = DCT_DCT;
-
- if (i == 16)
- type = PLANE_TYPE_UV;
- if (type == PLANE_TYPE_Y_WITH_DC) {
- tx_type = get_tx_type(xd, xd->block + i);
- }
-
- eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, type,
- tx_type, seg_eob, qcoeff_ptr,
- scan, TX_8X8, vp9_coef_bands_8x8);
- a[1] = a[0];
- l[1] = l[0];
-
- eobtotal += c;
- qcoeff_ptr += 64;
}
- if (bufthred == 16) {
- type = PLANE_TYPE_UV;
- tx_type = DCT_DCT;
- seg_eob = get_eob(xd, segment_id, 16);
+ // luma blocks
+ seg_eob = get_eob(xd, segment_id, 64);
+ for (i = 0; i < 16; i += 4) {
+ ENTROPY_CONTEXT *const a = A + vp9_block2above_8x8[i];
+ ENTROPY_CONTEXT *const l = L + vp9_block2left_8x8[i];
- // use 4x4 transform for U, V components in I8X8 prediction mode
+ eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, type,
+ type == PLANE_TYPE_Y_WITH_DC ?
+ get_tx_type(xd, xd->block + i) : DCT_DCT,
+ seg_eob, xd->block[i].qcoeff,
+ vp9_default_zig_zag1d_8x8,
+ TX_8X8, vp9_coef_bands_8x8);
+ a[1] = a[0];
+ l[1] = l[0];
+ eobtotal += c;
+ }
+
+ // chroma blocks
+ if (xd->mode_info_context->mbmi.mode == I8X8_PRED ||
+ xd->mode_info_context->mbmi.mode == SPLITMV) {
+ // use 4x4 transform for U, V components in I8X8/splitmv prediction mode
+ seg_eob = get_eob(xd, segment_id, 16);
for (i = 16; i < 24; i++) {
ENTROPY_CONTEXT *const a = A + vp9_block2above[i];
ENTROPY_CONTEXT *const l = L + vp9_block2left[i];
- const int *scan = vp9_default_zig_zag1d;
- eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, type,
- tx_type, seg_eob, qcoeff_ptr,
- scan, TX_4X4, vp9_coef_bands);
-
+ eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, PLANE_TYPE_UV,
+ DCT_DCT, seg_eob, xd->block[i].qcoeff,
+ vp9_default_zig_zag1d, TX_4X4, vp9_coef_bands);
eobtotal += c;
- qcoeff_ptr += 16;
+ }
+ } else {
+ for (i = 16; i < 24; i += 4) {
+ ENTROPY_CONTEXT *const a = A + vp9_block2above_8x8[i];
+ ENTROPY_CONTEXT *const l = L + vp9_block2left_8x8[i];
+
+ eobs[i] = c = decode_coefs(pbi, xd, bc, a, l, PLANE_TYPE_UV,
+ DCT_DCT, seg_eob, xd->block[i].qcoeff,
+ vp9_default_zig_zag1d_8x8,
+ TX_8X8, vp9_coef_bands_8x8);
+ a[1] = a[0];
+ l[1] = l[0];
+ eobtotal += c;
}
}
@@ -435,9 +419,9 @@
return eobtotal;
}
-int vp9_decode_mb_tokens_4x4(VP9D_COMP* const dx,
- MACROBLOCKD* const xd,
- BOOL_DECODER* const bc) {
+static int vp9_decode_mb_tokens_4x4(VP9D_COMP* const dx,
+ MACROBLOCKD* const xd,
+ BOOL_DECODER* const bc) {
int i, eobtotal = 0;
PLANE_TYPE type;
@@ -456,3 +440,21 @@
return eobtotal + vp9_decode_mb_tokens_4x4_uv(dx, xd, bc);
}
+
+int vp9_decode_mb_tokens(VP9D_COMP* const dx,
+ MACROBLOCKD* const xd,
+ BOOL_DECODER* const bc) {
+ const TX_SIZE tx_size = xd->mode_info_context->mbmi.txfm_size;
+ int eobtotal;
+
+ if (tx_size == TX_16X16) {
+ eobtotal = vp9_decode_mb_tokens_16x16(dx, xd, bc);
+ } else if (tx_size == TX_8X8) {
+ eobtotal = vp9_decode_mb_tokens_8x8(dx, xd, bc);
+ } else {
+ assert(tx_size == TX_4X4);
+ eobtotal = vp9_decode_mb_tokens_4x4(dx, xd, bc);
+ }
+
+ return eobtotal;
+}
diff --git a/vp9/decoder/detokenize.h b/vp9/decoder/detokenize.h
index a8f78f4..9f00d29 100644
--- a/vp9/decoder/detokenize.h
+++ b/vp9/decoder/detokenize.h
@@ -20,16 +20,10 @@
BOOL_DECODER* const bc,
PLANE_TYPE type, int i);
-int vp9_decode_mb_tokens_4x4(VP9D_COMP* const, MACROBLOCKD* const,
- BOOL_DECODER* const);
+int vp9_decode_mb_tokens(VP9D_COMP* const, MACROBLOCKD* const,
+ BOOL_DECODER* const);
int vp9_decode_mb_tokens_4x4_uv(VP9D_COMP* const dx, MACROBLOCKD* const xd,
BOOL_DECODER* const bc);
-int vp9_decode_mb_tokens_8x8(VP9D_COMP* const, MACROBLOCKD* const,
- BOOL_DECODER* const);
-
-int vp9_decode_mb_tokens_16x16(VP9D_COMP* const, MACROBLOCKD* const,
- BOOL_DECODER* const);
-
#endif /* DETOKENIZE_H */
diff --git a/vp9/encoder/rdopt.c b/vp9/encoder/rdopt.c
index 5dee848..a3e5b75 100644
--- a/vp9/encoder/rdopt.c
+++ b/vp9/encoder/rdopt.c
@@ -4620,7 +4620,8 @@
#endif
// Test best rd so far against threshold for trying this mode.
- if (best_rd <= cpi->rd_threshes[mode_index]) {
+ if (best_rd <= cpi->rd_threshes[mode_index] ||
+ cpi->rd_threshes[mode_index] == INT_MAX) {
continue;
}
