Use proper bit-depth path in non-rd encoding
In non-rd pickmode file, LBD and HBD path weren't written properly,
which resulted in LBD encoder calling some HBD functions. This CL
fixed that, and brought encoder speedup with minimal BDRate drop.
This change applied for high-bit-depth = 1 build.
Borg tests showed 2.4% - 4% speedup.
STATS_CHANGED for rt speed 7 to 10
Change-Id: I493882592fa59318ff74d2c55ac78627f13aa300
diff --git a/av1/encoder/nonrd_pickmode.c b/av1/encoder/nonrd_pickmode.c
index 4633cd3..4b12353 100644
--- a/av1/encoder/nonrd_pickmode.c
+++ b/av1/encoder/nonrd_pickmode.c
@@ -757,13 +757,14 @@
int eob_cost = 0;
const int bw = 4 * num_4x4_w;
const int bh = 4 * num_4x4_h;
+ const int use_hbd = is_cur_buf_hbd(xd);
(void)mi_row;
(void)mi_col;
(void)cpi;
#if CONFIG_AV1_HIGHBITDEPTH
- if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+ if (use_hbd) {
aom_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
p->src.stride, pd->dst.buf, pd->dst.stride,
x->e_mbd.bd);
@@ -784,14 +785,15 @@
if (c < max_blocks_wide) {
const SCAN_ORDER *const scan_order = &av1_scan_orders[tx_size][DCT_DCT];
const int block_offset = BLOCK_OFFSET(block);
+ int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
+ int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
+ int16_t *const low_dqcoeff = (int16_t *)p->dqcoeff + block_offset;
#if CONFIG_AV1_HIGHBITDEPTH
tran_low_t *const coeff = p->coeff + block_offset;
tran_low_t *const qcoeff = p->qcoeff + block_offset;
tran_low_t *const dqcoeff = p->dqcoeff + block_offset;
#else
- int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
- int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
- int16_t *const low_dqcoeff = (int16_t *)p->dqcoeff + block_offset;
+ (void)use_hbd;
#endif
uint16_t *const eob = &p->eobs[block];
const int diff_stride = bw;
@@ -805,28 +807,53 @@
case TX_32X32:
assert(0); // Not used
break;
+
#if CONFIG_AV1_HIGHBITDEPTH
case TX_16X16:
- aom_hadamard_16x16(src_diff, diff_stride, coeff);
- av1_quantize_fp(coeff, 16 * 16, p->zbin_QTX, p->round_fp_QTX,
- p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
- dqcoeff, p->dequant_QTX, eob, scan_order->scan,
- scan_order->iscan);
+ if (use_hbd) {
+ aom_hadamard_16x16(src_diff, diff_stride, coeff);
+ av1_quantize_fp(coeff, 16 * 16, p->zbin_QTX, p->round_fp_QTX,
+ p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
+ dqcoeff, p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ } else {
+ aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
+ av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX,
+ p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
+ p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ }
break;
case TX_8X8:
- aom_hadamard_8x8(src_diff, diff_stride, coeff);
- av1_quantize_fp(coeff, 8 * 8, p->zbin_QTX, p->round_fp_QTX,
- p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
- dqcoeff, p->dequant_QTX, eob, scan_order->scan,
- scan_order->iscan);
+ if (use_hbd) {
+ aom_hadamard_8x8(src_diff, diff_stride, coeff);
+ av1_quantize_fp(coeff, 8 * 8, p->zbin_QTX, p->round_fp_QTX,
+ p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
+ dqcoeff, p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ } else {
+ aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
+ av1_quantize_lp(low_coeff, 8 * 8, p->round_fp_QTX,
+ p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
+ p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ }
break;
default:
assert(tx_size == TX_4X4);
- aom_fdct4x4(src_diff, coeff, diff_stride);
- av1_quantize_fp(coeff, 4 * 4, p->zbin_QTX, p->round_fp_QTX,
- p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
- dqcoeff, p->dequant_QTX, eob, scan_order->scan,
- scan_order->iscan);
+ if (use_hbd) {
+ aom_fdct4x4(src_diff, coeff, diff_stride);
+ av1_quantize_fp(coeff, 4 * 4, p->zbin_QTX, p->round_fp_QTX,
+ p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
+ dqcoeff, p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ } else {
+ aom_fdct4x4_lp(src_diff, low_coeff, diff_stride);
+ av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX,
+ p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
+ p->dequant_QTX, eob, scan_order->scan,
+ scan_order->iscan);
+ }
break;
#else
case TX_16X16:
@@ -876,18 +903,32 @@
const int block_offset = BLOCK_OFFSET(block);
uint16_t *const eob = &p->eobs[block];
#if CONFIG_AV1_HIGHBITDEPTH
- int64_t dummy;
- tran_low_t *const coeff = p->coeff + block_offset;
- tran_low_t *const qcoeff = p->qcoeff + block_offset;
- tran_low_t *const dqcoeff = p->dqcoeff + block_offset;
+ if (use_hbd) {
+ int64_t dummy;
+ tran_low_t *const coeff = p->coeff + block_offset;
+ tran_low_t *const qcoeff = p->qcoeff + block_offset;
+ tran_low_t *const dqcoeff = p->dqcoeff + block_offset;
- if (*eob == 1)
- this_rdc->rate += (int)abs(qcoeff[0]);
- else if (*eob > 1)
- this_rdc->rate += aom_satd(qcoeff, step << 4);
+ if (*eob == 1)
+ this_rdc->rate += (int)abs(qcoeff[0]);
+ else if (*eob > 1)
+ this_rdc->rate += aom_satd(qcoeff, step << 4);
- this_rdc->dist +=
- av1_block_error(coeff, dqcoeff, step << 4, &dummy) >> 2;
+ this_rdc->dist +=
+ av1_block_error(coeff, dqcoeff, step << 4, &dummy) >> 2;
+ } else {
+ int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
+ int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
+ int16_t *const low_dqcoeff = (int16_t *)p->dqcoeff + block_offset;
+
+ if (*eob == 1)
+ this_rdc->rate += (int)abs(low_qcoeff[0]);
+ else if (*eob > 1)
+ this_rdc->rate += aom_satd_lp(low_qcoeff, step << 4);
+
+ this_rdc->dist +=
+ av1_block_error_lp(low_coeff, low_dqcoeff, step << 4) >> 2;
+ }
#else
int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
