#include "i965_encoder_utils.h"
#include "gen6_mfc.h"
#include "gen6_vme.h"
+#include "gen9_mfc.h"
#include "intel_media.h"
-#define BRC_CLIP(x, min, max) \
- { \
- x = ((x > (max)) ? (max) : ((x < (min)) ? (min) : x)); \
- }
-
-#define BRC_P_B_QP_DIFF 4
-#define BRC_I_P_QP_DIFF 2
-#define BRC_I_B_QP_DIFF (BRC_I_P_QP_DIFF + BRC_P_B_QP_DIFF)
-
-#define BRC_PWEIGHT 0.6 /* weight if P slice with comparison to I slice */
-#define BRC_BWEIGHT 0.25 /* weight if B slice with comparison to I slice */
-
-#define BRC_QP_MAX_CHANGE 5 /* maximum qp modification */
-#define BRC_CY 0.1 /* weight for */
-#define BRC_CX_UNDERFLOW 5.
-#define BRC_CX_OVERFLOW -4.
-
-#define BRC_PI_0_5 1.5707963267948966192313216916398
-
#ifndef HAVE_LOG2F
#define log2f(x) (logf(x)/(float)M_LN2)
#endif
static void
intel_mfc_bit_rate_control_context_init(struct encode_state *encode_state,
- struct gen6_mfc_context *mfc_context)
+ struct intel_encoder_context *encoder_context)
{
- VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
- int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
- int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
- float fps = pSequenceParameter->time_scale * 0.5 / pSequenceParameter->num_units_in_tick ;
- int inter_mb_size = pSequenceParameter->bits_per_second * 1.0 / (fps+4.0) / width_in_mbs / height_in_mbs;
- int intra_mb_size = inter_mb_size * 5.0;
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
int i;
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].target_mb_size = intra_mb_size;
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].target_frame_size = intra_mb_size * width_in_mbs * height_in_mbs;
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].target_mb_size = inter_mb_size;
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].target_frame_size = inter_mb_size * width_in_mbs * height_in_mbs;
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].target_mb_size = inter_mb_size;
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].target_frame_size = inter_mb_size * width_in_mbs * height_in_mbs;
-
for(i = 0 ; i < 3; i++) {
- mfc_context->bit_rate_control_context[i].QpPrimeY = 26;
mfc_context->bit_rate_control_context[i].MaxQpNegModifier = 6;
mfc_context->bit_rate_control_context[i].MaxQpPosModifier = 6;
mfc_context->bit_rate_control_context[i].GrowInit = 6;
mfc_context->bit_rate_control_context[i].Correct[4] = 4;
mfc_context->bit_rate_control_context[i].Correct[5] = 8;
}
-
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].TargetSizeInWord = (intra_mb_size + 16)/ 16;
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].TargetSizeInWord = (inter_mb_size + 16)/ 16;
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].TargetSizeInWord = (inter_mb_size + 16)/ 16;
-
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].MaxSizeInWord = mfc_context->bit_rate_control_context[SLICE_TYPE_I].TargetSizeInWord * 1.5;
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].MaxSizeInWord = mfc_context->bit_rate_control_context[SLICE_TYPE_P].TargetSizeInWord * 1.5;
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].MaxSizeInWord = mfc_context->bit_rate_control_context[SLICE_TYPE_B].TargetSizeInWord * 1.5;
}
static void intel_mfc_brc_init(struct encode_state *encode_state,
struct intel_encoder_context* encoder_context)
{
struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
- VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
- VAEncMiscParameterBuffer* pMiscParamHRD = (VAEncMiscParameterBuffer*)encode_state->misc_param[VAEncMiscParameterTypeHRD]->buffer;
- VAEncMiscParameterHRD* pParameterHRD = (VAEncMiscParameterHRD*)pMiscParamHRD->data;
- double bitrate = pSequenceParameter->bits_per_second;
- double framerate = (double)pSequenceParameter->time_scale /(2 * (double)pSequenceParameter->num_units_in_tick);
- int inum = 1, pnum = 0, bnum = 0; /* Gop structure: number of I, P, B frames in the Gop. */
- int intra_period = pSequenceParameter->intra_period;
- int ip_period = pSequenceParameter->ip_period;
- double qp1_size = 0.1 * 8 * 3 * (pSequenceParameter->picture_width_in_mbs<<4) * (pSequenceParameter->picture_height_in_mbs<<4)/2;
- double qp51_size = 0.001 * 8 * 3 * (pSequenceParameter->picture_width_in_mbs<<4) * (pSequenceParameter->picture_height_in_mbs<<4)/2;
- double bpf;
-
- if (pSequenceParameter->ip_period) {
- pnum = (intra_period + ip_period - 1)/ip_period - 1;
- bnum = intra_period - inum - pnum;
- }
+ double bitrate, framerate;
+ double frame_per_bits = 8 * 3 * encoder_context->frame_width_in_pixel * encoder_context->frame_height_in_pixel / 2;
+ double qp1_size = 0.1 * frame_per_bits;
+ double qp51_size = 0.001 * frame_per_bits;
+ int min_qp = MAX(1, encoder_context->brc.min_qp);
+ double bpf, factor, hrd_factor;
+ int inum = encoder_context->brc.num_iframes_in_gop,
+ pnum = encoder_context->brc.num_pframes_in_gop,
+ bnum = encoder_context->brc.num_bframes_in_gop; /* Gop structure: number of I, P, B frames in the Gop. */
+ int intra_period = encoder_context->brc.gop_size;
+ int i;
+
+ if (encoder_context->layer.num_layers > 1)
+ qp1_size = 0.15 * frame_per_bits;
mfc_context->brc.mode = encoder_context->rate_control_mode;
- mfc_context->brc.target_frame_size[SLICE_TYPE_I] = (int)((double)((bitrate * intra_period)/framerate) /
- (double)(inum + BRC_PWEIGHT * pnum + BRC_BWEIGHT * bnum));
- mfc_context->brc.target_frame_size[SLICE_TYPE_P] = BRC_PWEIGHT * mfc_context->brc.target_frame_size[SLICE_TYPE_I];
- mfc_context->brc.target_frame_size[SLICE_TYPE_B] = BRC_BWEIGHT * mfc_context->brc.target_frame_size[SLICE_TYPE_I];
+ mfc_context->hrd.violation_noted = 0;
- mfc_context->brc.gop_nums[SLICE_TYPE_I] = inum;
- mfc_context->brc.gop_nums[SLICE_TYPE_P] = pnum;
- mfc_context->brc.gop_nums[SLICE_TYPE_B] = bnum;
+ for (i = 0; i < encoder_context->layer.num_layers; i++) {
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_I] = 26;
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P] = 26;
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_B] = 26;
- bpf = mfc_context->brc.bits_per_frame = bitrate/framerate;
+ if (i == 0) {
+ bitrate = encoder_context->brc.bits_per_second[0];
+ framerate = (double)encoder_context->brc.framerate[0].num / (double)encoder_context->brc.framerate[0].den;
+ } else {
+ bitrate = (encoder_context->brc.bits_per_second[i] - encoder_context->brc.bits_per_second[i - 1]);
+ framerate = ((double)encoder_context->brc.framerate[i].num / (double)encoder_context->brc.framerate[i].den) -
+ ((double)encoder_context->brc.framerate[i - 1].num / (double)encoder_context->brc.framerate[i - 1].den);
+ }
- mfc_context->hrd.buffer_size = (double)pParameterHRD->buffer_size;
- mfc_context->hrd.current_buffer_fullness =
- (double)(pParameterHRD->initial_buffer_fullness < mfc_context->hrd.buffer_size)?
- pParameterHRD->initial_buffer_fullness: mfc_context->hrd.buffer_size/2.;
- mfc_context->hrd.target_buffer_fullness = (double)mfc_context->hrd.buffer_size/2.;
- mfc_context->hrd.buffer_capacity = (double)mfc_context->hrd.buffer_size/qp1_size;
- mfc_context->hrd.violation_noted = 0;
+ if (mfc_context->brc.mode == VA_RC_VBR && encoder_context->brc.target_percentage[i])
+ bitrate = bitrate * encoder_context->brc.target_percentage[i] / 100;
- if ((bpf > qp51_size) && (bpf < qp1_size)) {
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY = 51 - 50*(bpf - qp51_size)/(qp1_size - qp51_size);
- }
- else if (bpf >= qp1_size)
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY = 1;
- else if (bpf <= qp51_size)
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY = 51;
+ if (i == encoder_context->layer.num_layers - 1)
+ factor = 1.0;
+ else {
+ factor = ((double)encoder_context->brc.framerate[i].num / (double)encoder_context->brc.framerate[i].den) /
+ ((double)encoder_context->brc.framerate[i - 1].num / (double)encoder_context->brc.framerate[i - 1].den);
+ }
+
+ hrd_factor = (double)bitrate / encoder_context->brc.bits_per_second[encoder_context->layer.num_layers - 1];
+
+ mfc_context->hrd.buffer_size[i] = (unsigned int)(encoder_context->brc.hrd_buffer_size * hrd_factor);
+ mfc_context->hrd.current_buffer_fullness[i] =
+ (double)(encoder_context->brc.hrd_initial_buffer_fullness < encoder_context->brc.hrd_buffer_size) ?
+ encoder_context->brc.hrd_initial_buffer_fullness : encoder_context->brc.hrd_buffer_size / 2.;
+ mfc_context->hrd.current_buffer_fullness[i] *= hrd_factor;
+ mfc_context->hrd.target_buffer_fullness[i] = (double)encoder_context->brc.hrd_buffer_size * hrd_factor / 2.;
+ mfc_context->hrd.buffer_capacity[i] = (double)encoder_context->brc.hrd_buffer_size * hrd_factor / qp1_size;
+
+ if (encoder_context->layer.num_layers > 1) {
+ if (i == 0) {
+ intra_period = (int)(encoder_context->brc.gop_size * factor);
+ inum = 1;
+ pnum = (int)(encoder_context->brc.num_pframes_in_gop * factor);
+ bnum = intra_period - inum - pnum;
+ } else {
+ intra_period = (int)(encoder_context->brc.gop_size * factor) - intra_period;
+ inum = 0;
+ pnum = (int)(encoder_context->brc.num_pframes_in_gop * factor) - pnum;
+ bnum = intra_period - inum - pnum;
+ }
+ }
+
+ mfc_context->brc.gop_nums[i][SLICE_TYPE_I] = inum;
+ mfc_context->brc.gop_nums[i][SLICE_TYPE_P] = pnum;
+ mfc_context->brc.gop_nums[i][SLICE_TYPE_B] = bnum;
+
+ mfc_context->brc.target_frame_size[i][SLICE_TYPE_I] = (int)((double)((bitrate * intra_period)/framerate) /
+ (double)(inum + BRC_PWEIGHT * pnum + BRC_BWEIGHT * bnum));
+ mfc_context->brc.target_frame_size[i][SLICE_TYPE_P] = BRC_PWEIGHT * mfc_context->brc.target_frame_size[i][SLICE_TYPE_I];
+ mfc_context->brc.target_frame_size[i][SLICE_TYPE_B] = BRC_BWEIGHT * mfc_context->brc.target_frame_size[i][SLICE_TYPE_I];
+
+ bpf = mfc_context->brc.bits_per_frame[i] = bitrate/framerate;
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY = mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY;
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY = mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY;
+ if (encoder_context->brc.initial_qp) {
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_I] = encoder_context->brc.initial_qp;
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P] = encoder_context->brc.initial_qp;
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_B] = encoder_context->brc.initial_qp;
+ } else {
+ if ((bpf > qp51_size) && (bpf < qp1_size)) {
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P] = 51 - 50*(bpf - qp51_size)/(qp1_size - qp51_size);
+ }
+ else if (bpf >= qp1_size)
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P] = 1;
+ else if (bpf <= qp51_size)
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P] = 51;
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY, 1, 51);
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY, 1, 51);
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY, 1, 51);
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_I] = mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P];
+ mfc_context->brc.qp_prime_y[i][SLICE_TYPE_B] = mfc_context->brc.qp_prime_y[i][SLICE_TYPE_I];
+ }
+
+ BRC_CLIP(mfc_context->brc.qp_prime_y[i][SLICE_TYPE_I], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[i][SLICE_TYPE_P], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[i][SLICE_TYPE_B], min_qp, 51);
+ }
}
int intel_mfc_update_hrd(struct encode_state *encode_state,
- struct gen6_mfc_context *mfc_context,
+ struct intel_encoder_context *encoder_context,
int frame_bits)
{
- double prev_bf = mfc_context->hrd.current_buffer_fullness;
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ int layer_id = encoder_context->layer.curr_frame_layer_id;
+ double prev_bf = mfc_context->hrd.current_buffer_fullness[layer_id];
- mfc_context->hrd.current_buffer_fullness -= frame_bits;
+ mfc_context->hrd.current_buffer_fullness[layer_id] -= frame_bits;
- if (mfc_context->hrd.buffer_size > 0 && mfc_context->hrd.current_buffer_fullness <= 0.) {
- mfc_context->hrd.current_buffer_fullness = prev_bf;
+ if (mfc_context->hrd.buffer_size[layer_id] > 0 && mfc_context->hrd.current_buffer_fullness[layer_id] <= 0.) {
+ mfc_context->hrd.current_buffer_fullness[layer_id] = prev_bf;
return BRC_UNDERFLOW;
}
- mfc_context->hrd.current_buffer_fullness += mfc_context->brc.bits_per_frame;
- if (mfc_context->hrd.buffer_size > 0 && mfc_context->hrd.current_buffer_fullness > mfc_context->hrd.buffer_size) {
+ mfc_context->hrd.current_buffer_fullness[layer_id] += mfc_context->brc.bits_per_frame[layer_id];
+ if (mfc_context->hrd.buffer_size[layer_id] > 0 && mfc_context->hrd.current_buffer_fullness[layer_id] > mfc_context->hrd.buffer_size[layer_id]) {
if (mfc_context->brc.mode == VA_RC_VBR)
- mfc_context->hrd.current_buffer_fullness = mfc_context->hrd.buffer_size;
+ mfc_context->hrd.current_buffer_fullness[layer_id] = mfc_context->hrd.buffer_size[layer_id];
else {
- mfc_context->hrd.current_buffer_fullness = prev_bf;
+ mfc_context->hrd.current_buffer_fullness[layer_id] = prev_bf;
return BRC_OVERFLOW;
}
}
return BRC_NO_HRD_VIOLATION;
}
-int intel_mfc_brc_postpack(struct encode_state *encode_state,
- struct gen6_mfc_context *mfc_context,
- int frame_bits)
+static int intel_mfc_brc_postpack_cbr(struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context,
+ int frame_bits)
{
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
gen6_brc_status sts = BRC_NO_HRD_VIOLATION;
VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
int slicetype = intel_avc_enc_slice_type_fixup(pSliceParameter->slice_type);
- int qpi = mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY;
- int qpp = mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY;
- int qpb = mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY;
+ int curr_frame_layer_id, next_frame_layer_id;
+ int qpi, qpp, qpb;
int qp; // quantizer of previously encoded slice of current type
int qpn; // predicted quantizer for next frame of current type in integer format
double qpf; // predicted quantizer for next frame of current type in float format
double delta_qp; // QP correction
+ int min_qp = MAX(1, encoder_context->brc.min_qp);
int target_frame_size, frame_size_next;
/* Notes:
* x - how far we are from HRD buffer borders
double x, y;
double frame_size_alpha;
- qp = mfc_context->bit_rate_control_context[slicetype].QpPrimeY;
+ if (encoder_context->layer.num_layers < 2 || encoder_context->layer.size_frame_layer_ids == 0) {
+ curr_frame_layer_id = 0;
+ next_frame_layer_id = 0;
+ } else {
+ curr_frame_layer_id = encoder_context->layer.curr_frame_layer_id;
+ next_frame_layer_id = encoder_context->layer.frame_layer_ids[encoder_context->num_frames_in_sequence % encoder_context->layer.size_frame_layer_ids];
+ }
+
+ /* checking wthether HRD compliance first */
+ sts = intel_mfc_update_hrd(encode_state, encoder_context, frame_bits);
+
+ if (sts == BRC_NO_HRD_VIOLATION) { // no HRD violation
+ /* nothing */
+ } else {
+ next_frame_layer_id = curr_frame_layer_id;
+ }
+
+ mfc_context->brc.bits_prev_frame[curr_frame_layer_id] = frame_bits;
+ frame_bits = mfc_context->brc.bits_prev_frame[next_frame_layer_id];
+
+ mfc_context->brc.prev_slice_type[curr_frame_layer_id] = slicetype;
+ slicetype = mfc_context->brc.prev_slice_type[next_frame_layer_id];
+
+ /* 0 means the next frame is the first frame of next layer */
+ if (frame_bits == 0)
+ return sts;
+
+ qpi = mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_I];
+ qpp = mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_P];
+ qpb = mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_B];
+
+ qp = mfc_context->brc.qp_prime_y[next_frame_layer_id][slicetype];
- target_frame_size = mfc_context->brc.target_frame_size[slicetype];
- if (mfc_context->hrd.buffer_capacity < 5)
+ target_frame_size = mfc_context->brc.target_frame_size[next_frame_layer_id][slicetype];
+ if (mfc_context->hrd.buffer_capacity[next_frame_layer_id] < 5)
frame_size_alpha = 0;
else
- frame_size_alpha = (double)mfc_context->brc.gop_nums[slicetype];
+ frame_size_alpha = (double)mfc_context->brc.gop_nums[next_frame_layer_id][slicetype];
if (frame_size_alpha > 30) frame_size_alpha = 30;
frame_size_next = target_frame_size + (double)(target_frame_size - frame_bits) /
(double)(frame_size_alpha + 1.);
if (qpn == qp) {
/* setting qpn we round qpf making mistakes: now we are trying to compensate this */
- mfc_context->brc.qpf_rounding_accumulator += qpf - qpn;
- if (mfc_context->brc.qpf_rounding_accumulator > 1.0) {
+ mfc_context->brc.qpf_rounding_accumulator[next_frame_layer_id] += qpf - qpn;
+ if (mfc_context->brc.qpf_rounding_accumulator[next_frame_layer_id] > 1.0) {
qpn++;
- mfc_context->brc.qpf_rounding_accumulator = 0.;
- } else if (mfc_context->brc.qpf_rounding_accumulator < -1.0) {
+ mfc_context->brc.qpf_rounding_accumulator[next_frame_layer_id] = 0.;
+ } else if (mfc_context->brc.qpf_rounding_accumulator[next_frame_layer_id] < -1.0) {
qpn--;
- mfc_context->brc.qpf_rounding_accumulator = 0.;
+ mfc_context->brc.qpf_rounding_accumulator[next_frame_layer_id] = 0.;
}
}
/* making sure that QP is not changing too fast */
/* making sure that with QP predictions we did do not leave QPs range */
BRC_CLIP(qpn, 1, 51);
- /* checking wthether HRD compliance is still met */
- sts = intel_mfc_update_hrd(encode_state, mfc_context, frame_bits);
-
/* calculating QP delta as some function*/
- x = mfc_context->hrd.target_buffer_fullness - mfc_context->hrd.current_buffer_fullness;
+ x = mfc_context->hrd.target_buffer_fullness[next_frame_layer_id] - mfc_context->hrd.current_buffer_fullness[next_frame_layer_id];
if (x > 0) {
- x /= mfc_context->hrd.target_buffer_fullness;
- y = mfc_context->hrd.current_buffer_fullness;
+ x /= mfc_context->hrd.target_buffer_fullness[next_frame_layer_id];
+ y = mfc_context->hrd.current_buffer_fullness[next_frame_layer_id];
}
else {
- x /= (mfc_context->hrd.buffer_size - mfc_context->hrd.target_buffer_fullness);
- y = mfc_context->hrd.buffer_size - mfc_context->hrd.current_buffer_fullness;
+ x /= (mfc_context->hrd.buffer_size[next_frame_layer_id] - mfc_context->hrd.target_buffer_fullness[next_frame_layer_id]);
+ y = mfc_context->hrd.buffer_size[next_frame_layer_id] - mfc_context->hrd.current_buffer_fullness[next_frame_layer_id];
}
if (y < 0.01) y = 0.01;
if (x > 1) x = 1;
qpn = (int)(qpn + delta_qp + 0.5);
/* making sure that with QP predictions we did do not leave QPs range */
- BRC_CLIP(qpn, 1, 51);
+ BRC_CLIP(qpn, min_qp, 51);
if (sts == BRC_NO_HRD_VIOLATION) { // no HRD violation
/* correcting QPs of slices of other types */
if (slicetype == SLICE_TYPE_P) {
if (abs(qpn + BRC_P_B_QP_DIFF - qpb) > 2)
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY += (qpn + BRC_P_B_QP_DIFF - qpb) >> 1;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_B] += (qpn + BRC_P_B_QP_DIFF - qpb) >> 1;
if (abs(qpn - BRC_I_P_QP_DIFF - qpi) > 2)
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY += (qpn - BRC_I_P_QP_DIFF - qpi) >> 1;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_I] += (qpn - BRC_I_P_QP_DIFF - qpi) >> 1;
} else if (slicetype == SLICE_TYPE_I) {
if (abs(qpn + BRC_I_B_QP_DIFF - qpb) > 4)
- mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY += (qpn + BRC_I_B_QP_DIFF - qpb) >> 2;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_B] += (qpn + BRC_I_B_QP_DIFF - qpb) >> 2;
if (abs(qpn + BRC_I_P_QP_DIFF - qpp) > 2)
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY += (qpn + BRC_I_P_QP_DIFF - qpp) >> 2;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_P] += (qpn + BRC_I_P_QP_DIFF - qpp) >> 2;
} else { // SLICE_TYPE_B
if (abs(qpn - BRC_P_B_QP_DIFF - qpp) > 2)
- mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY += (qpn - BRC_P_B_QP_DIFF - qpp) >> 1;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_P] += (qpn - BRC_P_B_QP_DIFF - qpp) >> 1;
if (abs(qpn - BRC_I_B_QP_DIFF - qpi) > 4)
- mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY += (qpn - BRC_I_B_QP_DIFF - qpi) >> 2;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_I] += (qpn - BRC_I_B_QP_DIFF - qpi) >> 2;
}
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_I].QpPrimeY, 1, 51);
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_P].QpPrimeY, 1, 51);
- BRC_CLIP(mfc_context->bit_rate_control_context[SLICE_TYPE_B].QpPrimeY, 1, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_I], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_P], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[next_frame_layer_id][SLICE_TYPE_B], min_qp, 51);
} else if (sts == BRC_UNDERFLOW) { // underflow
if (qpn <= qp) qpn = qp + 1;
if (qpn > 51) {
}
} else if (sts == BRC_OVERFLOW) {
if (qpn >= qp) qpn = qp - 1;
- if (qpn < 1) { // < 0 (?) overflow with minQP
- qpn = 1;
+ if (qpn < min_qp) { // overflow with minQP
+ qpn = min_qp;
sts = BRC_OVERFLOW_WITH_MIN_QP; // bit stuffing to be done
}
}
- mfc_context->bit_rate_control_context[slicetype].QpPrimeY = qpn;
+ mfc_context->brc.qp_prime_y[next_frame_layer_id][slicetype] = qpn;
return sts;
}
+static int intel_mfc_brc_postpack_vbr(struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context,
+ int frame_bits)
+{
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ gen6_brc_status sts;
+ VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int slice_type = intel_avc_enc_slice_type_fixup(pSliceParameter->slice_type);
+ int *qp = mfc_context->brc.qp_prime_y[0];
+ int min_qp = MAX(1, encoder_context->brc.min_qp);
+ int qp_delta, large_frame_adjustment;
+
+ // This implements a simple reactive VBR rate control mode for single-layer H.264. The primary
+ // aim here is to avoid the problematic behaviour that the CBR rate controller displays on
+ // scene changes, where the QP can get pushed up by a large amount in a short period and
+ // compromise the quality of following frames to a very visible degree.
+ // The main idea, then, is to try to keep the HRD buffering above the target level most of the
+ // time, so that when a large frame is generated (on a scene change or when the stream
+ // complexity increases) we have plenty of slack to be able to encode the more difficult region
+ // without compromising quality immediately on the following frames. It is optimistic about
+ // the complexity of future frames, so even after generating one or more large frames on a
+ // significant change it will try to keep the QP at its current level until the HRD buffer
+ // bounds force a change to maintain the intended rate.
+
+ sts = intel_mfc_update_hrd(encode_state, encoder_context, frame_bits);
+
+ // This adjustment is applied to increase the QP by more than we normally would if a very
+ // large frame is encountered and we are in danger of running out of slack.
+ large_frame_adjustment = rint(2.0 * log(frame_bits / mfc_context->brc.target_frame_size[0][slice_type]));
+
+ if (sts == BRC_UNDERFLOW) {
+ // The frame is far too big and we don't have the bits available to send it, so it will
+ // have to be re-encoded at a higher QP.
+ qp_delta = +2;
+ if (frame_bits > mfc_context->brc.target_frame_size[0][slice_type])
+ qp_delta += large_frame_adjustment;
+ } else if (sts == BRC_OVERFLOW) {
+ // The frame is very small and we are now overflowing the HRD buffer. Currently this case
+ // does not occur because we ignore overflow in VBR mode.
+ assert(0 && "Overflow in VBR mode");
+ } else if (frame_bits <= mfc_context->brc.target_frame_size[0][slice_type]) {
+ // The frame is smaller than the average size expected for this frame type.
+ if (mfc_context->hrd.current_buffer_fullness[0] >
+ (mfc_context->hrd.target_buffer_fullness[0] + mfc_context->hrd.buffer_size[0]) / 2.0) {
+ // We currently have lots of bits available, so decrease the QP slightly for the next
+ // frame.
+ qp_delta = -1;
+ } else {
+ // The HRD buffer fullness is increasing, so do nothing. (We may be under the target
+ // level here, but are moving in the right direction.)
+ qp_delta = 0;
+ }
+ } else {
+ // The frame is larger than the average size expected for this frame type.
+ if (mfc_context->hrd.current_buffer_fullness[0] > mfc_context->hrd.target_buffer_fullness[0]) {
+ // We are currently over the target level, so do nothing.
+ qp_delta = 0;
+ } else if (mfc_context->hrd.current_buffer_fullness[0] > mfc_context->hrd.target_buffer_fullness[0] / 2.0) {
+ // We are under the target level, but not critically. Increase the QP by one step if
+ // continuing like this would underflow soon (currently within one second).
+ if (mfc_context->hrd.current_buffer_fullness[0] /
+ (double)(frame_bits - mfc_context->brc.target_frame_size[0][slice_type] + 1) <
+ ((double)encoder_context->brc.framerate[0].num / (double)encoder_context->brc.framerate[0].den))
+ qp_delta = +1;
+ else
+ qp_delta = 0;
+ } else {
+ // We are a long way under the target level. Always increase the QP, possibly by a
+ // larger amount dependent on how big the frame we just made actually was.
+ qp_delta = +1 + large_frame_adjustment;
+ }
+ }
+
+ switch (slice_type) {
+ case SLICE_TYPE_I:
+ qp[SLICE_TYPE_I] += qp_delta;
+ qp[SLICE_TYPE_P] = qp[SLICE_TYPE_I] + BRC_I_P_QP_DIFF;
+ qp[SLICE_TYPE_B] = qp[SLICE_TYPE_I] + BRC_I_B_QP_DIFF;
+ break;
+ case SLICE_TYPE_P:
+ qp[SLICE_TYPE_P] += qp_delta;
+ qp[SLICE_TYPE_I] = qp[SLICE_TYPE_P] - BRC_I_P_QP_DIFF;
+ qp[SLICE_TYPE_B] = qp[SLICE_TYPE_P] + BRC_P_B_QP_DIFF;
+ break;
+ case SLICE_TYPE_B:
+ qp[SLICE_TYPE_B] += qp_delta;
+ qp[SLICE_TYPE_I] = qp[SLICE_TYPE_B] - BRC_I_B_QP_DIFF;
+ qp[SLICE_TYPE_P] = qp[SLICE_TYPE_B] - BRC_P_B_QP_DIFF;
+ break;
+ }
+ BRC_CLIP(mfc_context->brc.qp_prime_y[0][SLICE_TYPE_I], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[0][SLICE_TYPE_P], min_qp, 51);
+ BRC_CLIP(mfc_context->brc.qp_prime_y[0][SLICE_TYPE_B], min_qp, 51);
+
+ if (sts == BRC_UNDERFLOW && qp[slice_type] == 51)
+ sts = BRC_UNDERFLOW_WITH_MAX_QP;
+ if (sts == BRC_OVERFLOW && qp[slice_type] == min_qp)
+ sts = BRC_OVERFLOW_WITH_MIN_QP;
+
+ return sts;
+}
+
+int intel_mfc_brc_postpack(struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context,
+ int frame_bits)
+{
+ switch (encoder_context->rate_control_mode) {
+ case VA_RC_CBR:
+ return intel_mfc_brc_postpack_cbr(encode_state, encoder_context, frame_bits);
+ case VA_RC_VBR:
+ return intel_mfc_brc_postpack_vbr(encode_state, encoder_context, frame_bits);
+ }
+ assert(0 && "Invalid RC mode");
+}
+
static void intel_mfc_hrd_context_init(struct encode_state *encode_state,
struct intel_encoder_context *encoder_context)
{
struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
- VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
unsigned int rate_control_mode = encoder_context->rate_control_mode;
- int target_bit_rate = pSequenceParameter->bits_per_second;
+ int target_bit_rate = encoder_context->brc.bits_per_second[encoder_context->layer.num_layers - 1];
// current we only support CBR mode.
if (rate_control_mode == VA_RC_CBR) {
mfc_context->vui_hrd.i_bit_rate_value = target_bit_rate >> 10;
- mfc_context->vui_hrd.i_cpb_size_value = (target_bit_rate * 8) >> 10;
- mfc_context->vui_hrd.i_initial_cpb_removal_delay = mfc_context->vui_hrd.i_cpb_size_value * 0.5 * 1024 / target_bit_rate * 90000;
+ mfc_context->vui_hrd.i_initial_cpb_removal_delay = ((target_bit_rate * 8) >> 10) * 0.5 * 1024 / target_bit_rate * 90000;
mfc_context->vui_hrd.i_cpb_removal_delay = 2;
mfc_context->vui_hrd.i_frame_number = 0;
return 1;
}
-/*
- * Check whether the parameters related with CBR are updated and decide whether
- * it needs to reinitialize the configuration related with CBR.
- * Currently it will check the following parameters:
- * bits_per_second
- * frame_rate
- * gop_configuration(intra_period, ip_period, intra_idr_period)
- */
-static bool intel_mfc_brc_updated_check(struct encode_state *encode_state,
- struct intel_encoder_context *encoder_context)
-{
- unsigned int rate_control_mode = encoder_context->rate_control_mode;
- struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
- double cur_fps, cur_bitrate;
- VAEncSequenceParameterBufferH264 *pSequenceParameter;
-
-
- if (rate_control_mode != VA_RC_CBR) {
- return false;
- }
-
- pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
-
- cur_bitrate = pSequenceParameter->bits_per_second;
- cur_fps = (double)pSequenceParameter->time_scale /
- (2 * (double)pSequenceParameter->num_units_in_tick);
-
- if ((cur_bitrate == mfc_context->brc.saved_bps) &&
- (cur_fps == mfc_context->brc.saved_fps) &&
- (pSequenceParameter->intra_period == mfc_context->brc.saved_intra_period) &&
- (pSequenceParameter->intra_idr_period == mfc_context->brc.saved_idr_period) &&
- (pSequenceParameter->intra_period == mfc_context->brc.saved_intra_period)) {
- /* the parameters related with CBR are not updaetd */
- return false;
- }
-
- mfc_context->brc.saved_ip_period = pSequenceParameter->ip_period;
- mfc_context->brc.saved_intra_period = pSequenceParameter->intra_period;
- mfc_context->brc.saved_idr_period = pSequenceParameter->intra_idr_period;
- mfc_context->brc.saved_fps = cur_fps;
- mfc_context->brc.saved_bps = cur_bitrate;
- return true;
-}
-
void intel_mfc_brc_prepare(struct encode_state *encode_state,
struct intel_encoder_context *encoder_context)
{
unsigned int rate_control_mode = encoder_context->rate_control_mode;
- struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
-
- if (rate_control_mode == VA_RC_CBR) {
- bool brc_updated;
- assert(encoder_context->codec != CODEC_MPEG2);
- brc_updated = intel_mfc_brc_updated_check(encode_state, encoder_context);
+ if (encoder_context->codec != CODEC_H264 &&
+ encoder_context->codec != CODEC_H264_MVC)
+ return;
+ if (rate_control_mode != VA_RC_CQP) {
/*Programing bit rate control */
- if ((mfc_context->bit_rate_control_context[SLICE_TYPE_I].MaxSizeInWord == 0) ||
- brc_updated) {
- intel_mfc_bit_rate_control_context_init(encode_state, mfc_context);
+ if (encoder_context->brc.need_reset) {
+ intel_mfc_bit_rate_control_context_init(encode_state, encoder_context);
intel_mfc_brc_init(encode_state, encoder_context);
}
/*Programing HRD control */
- if ((mfc_context->vui_hrd.i_cpb_size_value == 0) || brc_updated )
+ if (encoder_context->brc.need_reset)
intel_mfc_hrd_context_init(encode_state, encoder_context);
}
}
-static int intel_avc_find_skipemulcnt(unsigned char *buf, int bits_length)
-{
- int i, found;
- int leading_zero_cnt, byte_length, zero_byte;
- int nal_unit_type;
- int skip_cnt = 0;
-
-#define NAL_UNIT_TYPE_MASK 0x1f
-#define HW_MAX_SKIP_LENGTH 15
-
- byte_length = ALIGN(bits_length, 32) >> 3;
-
-
- leading_zero_cnt = 0;
- found = 0;
- for(i = 0; i < byte_length - 4; i++) {
- if (((buf[i] == 0) && (buf[i + 1] == 0) && (buf[i + 2] == 1)) ||
- ((buf[i] == 0) && (buf[i + 1] == 0) && (buf[i + 2] == 0) && (buf[i + 3] == 1))) {
- found = 1;
- break;
- }
- leading_zero_cnt++;
- }
- if (!found) {
- /* warning message is complained. But anyway it will be inserted. */
- WARN_ONCE("Invalid packed header data. "
- "Can't find the 000001 start_prefix code\n");
- return 0;
- }
- i = leading_zero_cnt;
-
- zero_byte = 0;
- if (!((buf[i] == 0) && (buf[i + 1] == 0) && (buf[i + 2] == 1)))
- zero_byte = 1;
-
- skip_cnt = leading_zero_cnt + zero_byte + 3;
-
- /* the unit header byte is accounted */
- nal_unit_type = (buf[skip_cnt]) & NAL_UNIT_TYPE_MASK;
- skip_cnt += 1;
-
- if (nal_unit_type == 14 || nal_unit_type == 20 || nal_unit_type == 21) {
- /* more unit header bytes are accounted for MVC/SVC */
- skip_cnt += 3;
- }
- if (skip_cnt > HW_MAX_SKIP_LENGTH) {
- WARN_ONCE("Too many leading zeros are padded for packed data. "
- "It is beyond the HW range.!!!\n");
- }
- return skip_cnt;
-}
-
void intel_mfc_avc_pipeline_header_programing(VADriverContextP ctx,
struct encode_state *encode_state,
struct intel_encoder_context *encoder_context,
if ( obj_surface->private_data == NULL) {
gen6_avc_surface = calloc(sizeof(GenAvcSurface), 1);
+ assert(gen6_avc_surface);
gen6_avc_surface->dmv_top =
dri_bo_alloc(i965->intel.bufmgr,
"Buffer",
if ( obj_surface->private_data == NULL) {
gen6_avc_surface = calloc(sizeof(GenAvcSurface), 1);
+ assert(gen6_avc_surface);
gen6_avc_surface->dmv_top =
dri_bo_alloc(i965->intel.bufmgr,
"Buffer",
#define QP_MAX 52
+#define VP8_QP_MAX 128
static float intel_lambda_qp(int qp)
return lambdaf;
}
-
-void intel_vme_update_mbmv_cost(VADriverContextP ctx,
- struct encode_state *encode_state,
- struct intel_encoder_context *encoder_context)
+static
+void intel_h264_calc_mbmvcost_qp(int qp,
+ int slice_type,
+ uint8_t *vme_state_message)
{
- struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
- struct gen6_vme_context *vme_context = encoder_context->vme_context;
- VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
- VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
- int qp, m_cost, j, mv_count;
- uint8_t *vme_state_message = (uint8_t *)(vme_context->vme_state_message);
+ int m_cost, j, mv_count;
float lambda, m_costf;
- int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
-
-
- if (encoder_context->rate_control_mode == VA_RC_CQP)
- qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
- else
- qp = mfc_context->bit_rate_control_context[slice_type].QpPrimeY;
-
- if (vme_state_message == NULL)
- return;
-
assert(qp <= QP_MAX);
lambda = intel_lambda_qp(qp);
+
+ m_cost = lambda;
+ vme_state_message[MODE_CHROMA_INTRA] = 0;
+ vme_state_message[MODE_REFID_COST] = intel_format_lutvalue(m_cost, 0x8f);
+
if (slice_type == SLICE_TYPE_I) {
vme_state_message[MODE_INTRA_16X16] = 0;
m_cost = lambda * 4;
vme_state_message[MODE_INTER_BWD] = intel_format_lutvalue(m_cost, 0x6f);
}
}
+ return;
+}
+
+void intel_vme_update_mbmv_cost(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int qp;
+ uint8_t *vme_state_message = (uint8_t *)(vme_context->vme_state_message);
+
+ int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+
+ if (encoder_context->rate_control_mode == VA_RC_CQP)
+ qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
+ else
+ qp = mfc_context->brc.qp_prime_y[encoder_context->layer.curr_frame_layer_id][slice_type];
+
+ if (vme_state_message == NULL)
+ return;
+
+ intel_h264_calc_mbmvcost_qp(qp, slice_type, vme_state_message);
}
void intel_vme_vp8_update_mbmv_cost(VADriverContextP ctx,
float lambda, m_costf;
int is_key_frame = !pic_param->pic_flags.bits.frame_type;
+ int slice_type = (is_key_frame ? SLICE_TYPE_I : SLICE_TYPE_P);
if (vme_state_message == NULL)
return;
- lambda = intel_lambda_qp(q_matrix->quantization_index[0] >> 1);
+ if (encoder_context->rate_control_mode == VA_RC_CQP)
+ qp = q_matrix->quantization_index[0];
+ else
+ qp = mfc_context->brc.qp_prime_y[encoder_context->layer.curr_frame_layer_id][slice_type];
+
+ lambda = intel_lambda_qp(qp * QP_MAX / VP8_QP_MAX);
+
+ m_cost = lambda;
+ vme_state_message[MODE_CHROMA_INTRA] = intel_format_lutvalue(m_cost, 0x8f);
+
if (is_key_frame) {
vme_state_message[MODE_INTRA_16X16] = 0;
m_cost = lambda * 16;
vme_state_message[MODE_INTRA_4X4] = intel_format_lutvalue(m_cost, 0x8f);
+ m_cost = lambda * 3;
+ vme_state_message[MODE_INTRA_NONPRED] = intel_format_lutvalue(m_cost, 0x6f);
} else {
m_cost = 0;
vme_state_message[MODE_INTER_MV0] = intel_format_lutvalue(m_cost, 0x6f);
mv_count++;
}
- if (q_matrix->quantization_index[0] < 32 ) {
+ if (qp < 92 ) {
vme_state_message[MODE_INTRA_16X16] = 0x4a;
vme_state_message[MODE_INTRA_4X4] = 0x4a;
vme_state_message[MODE_INTRA_NONPRED] = 0x4a;
vme_state_message[MODE_INTER_16X8] = 0x4a;
vme_state_message[MODE_INTER_8X8] = 0x4a;
vme_state_message[MODE_INTER_4X4] = 0x4a;
+ vme_state_message[MODE_INTER_BWD] = 0;
return;
}
m_costf = lambda * 10;
m_cost = lambda * 24;
vme_state_message[MODE_INTRA_4X4] = intel_format_lutvalue(m_cost, 0x8f);
+ m_costf = lambda * 3.5;
+ m_cost = m_costf;
+ vme_state_message[MODE_INTRA_NONPRED] = intel_format_lutvalue(m_cost, 0x6f);
+
m_costf = lambda * 2.5;
m_cost = m_costf;
vme_state_message[MODE_INTER_16X16] = intel_format_lutvalue(m_cost, 0x8f);
int mb_row;
int s;
unsigned int *command_ptr;
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int qp,qp_mb,qp_index;
+ int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+
+ if (encoder_context->rate_control_mode == VA_RC_CQP)
+ qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
+ else
+ qp = mfc_context->brc.qp_prime_y[encoder_context->layer.curr_frame_layer_id][slice_type];
#define USE_SCOREBOARD (1 << 21)
}
}
- *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (9 - 2));
*command_ptr++ = kernel;
*command_ptr++ = USE_SCOREBOARD;
/* Indirect data */
/*inline data */
*command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
*command_ptr++ = ((1 << 18) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+ /* QP occupies one byte */
+ if (vme_context->roi_enabled) {
+ qp_index = y_inner * mb_width + x_inner;
+ qp_mb = *(vme_context->qp_per_mb + qp_index);
+ } else
+ qp_mb = qp;
+ *command_ptr++ = qp_mb;
x_inner -= 2;
y_inner += 1;
}
}
}
- *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (9 - 2));
*command_ptr++ = kernel;
*command_ptr++ = USE_SCOREBOARD;
/* Indirect data */
/*inline data */
*command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
*command_ptr++ = ((1 << 18) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+ /* qp occupies one byte */
+ if (vme_context->roi_enabled) {
+ qp_index = y_inner * mb_width + x_inner;
+ qp_mb = *(vme_context->qp_per_mb + qp_index);
+ } else
+ qp_mb = qp;
+ *command_ptr++ = qp_mb;
x_inner -= 2;
y_inner += 1;
return;
}
+void
+intel_h264_initialize_mbmv_cost(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ struct i965_driver_data *i965 = i965_driver_data(ctx);
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int qp;
+ dri_bo *bo;
+ uint8_t *cost_table;
+
+ int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+
+
+ if (slice_type == SLICE_TYPE_I) {
+ if (vme_context->i_qp_cost_table)
+ return;
+ } else if (slice_type == SLICE_TYPE_P) {
+ if (vme_context->p_qp_cost_table)
+ return;
+ } else {
+ if (vme_context->b_qp_cost_table)
+ return;
+ }
+
+ /* It is enough to allocate 32 bytes for each qp. */
+ bo = dri_bo_alloc(i965->intel.bufmgr,
+ "cost_table ",
+ QP_MAX * 32,
+ 64);
+
+ dri_bo_map(bo, 1);
+ assert(bo->virtual);
+ cost_table = (uint8_t *)(bo->virtual);
+ for (qp = 0; qp < QP_MAX; qp++) {
+ intel_h264_calc_mbmvcost_qp(qp, slice_type, cost_table);
+ cost_table += 32;
+ }
+
+ dri_bo_unmap(bo);
+
+ if (slice_type == SLICE_TYPE_I) {
+ vme_context->i_qp_cost_table = bo;
+ } else if (slice_type == SLICE_TYPE_P) {
+ vme_context->p_qp_cost_table = bo;
+ } else {
+ vme_context->b_qp_cost_table = bo;
+ }
+
+ vme_context->cost_table_size = QP_MAX * 32;
+ return;
+}
+
+extern void
+intel_h264_setup_cost_surface(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context,
+ unsigned long binding_table_offset,
+ unsigned long surface_state_offset)
+{
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ dri_bo *bo;
+
+
+ struct i965_buffer_surface cost_table;
+
+ int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+
+
+ if (slice_type == SLICE_TYPE_I) {
+ bo = vme_context->i_qp_cost_table;
+ } else if (slice_type == SLICE_TYPE_P) {
+ bo = vme_context->p_qp_cost_table;
+ } else {
+ bo = vme_context->b_qp_cost_table;
+ }
+
+ cost_table.bo = bo;
+ cost_table.num_blocks = QP_MAX;
+ cost_table.pitch = 16;
+ cost_table.size_block = 32;
+
+ vme_context->vme_buffer_suface_setup(ctx,
+ &vme_context->gpe_context,
+ &cost_table,
+ binding_table_offset,
+ surface_state_offset);
+}
+
+/*
+ * the idea of conversion between qp and qstep comes from scaling process
+ * of transform coeff for Luma component in H264 spec.
+ * 2^(Qpy / 6 - 6)
+ * In order to avoid too small qstep, it is multiplied by 16.
+ */
+static float intel_h264_qp_qstep(int qp)
+{
+ float value, qstep;
+ value = qp;
+ value = value / 6 - 2;
+ qstep = powf(2, value);
+ return qstep;
+}
+
+static int intel_h264_qstep_qp(float qstep)
+{
+ float qp;
+
+ qp = 12.0f + 6.0f * log2f(qstep);
+
+ return floorf(qp);
+}
+
+/*
+ * Currently it is based on the following assumption:
+ * SUM(roi_area * 1 / roi_qstep) + non_area * 1 / nonroi_qstep =
+ * total_aread * 1 / baseqp_qstep
+ *
+ * qstep is the linearized quantizer of H264 quantizer
+ */
+typedef struct {
+ int row_start_in_mb;
+ int row_end_in_mb;
+ int col_start_in_mb;
+ int col_end_in_mb;
+
+ int width_mbs;
+ int height_mbs;
+
+ int roi_qp;
+} ROIRegionParam;
+
+static VAStatus
+intel_h264_enc_roi_cbr(VADriverContextP ctx,
+ int base_qp,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ int nonroi_qp;
+ int min_qp = MAX(1, encoder_context->brc.min_qp);
+ bool quickfill = 0;
+
+ ROIRegionParam param_regions[I965_MAX_NUM_ROI_REGIONS];
+ int num_roi = 0;
+ int i,j;
+
+ float temp;
+ float qstep_nonroi, qstep_base;
+ float roi_area, total_area, nonroi_area;
+ float sum_roi;
+
+ VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
+ int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
+ int mbs_in_picture = width_in_mbs * height_in_mbs;
+
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ VAStatus vaStatus = VA_STATUS_SUCCESS;
+
+ /* currently roi_value_is_qp_delta is the only supported mode of priority.
+ *
+ * qp_delta set by user is added to base_qp, which is then clapped by
+ * [base_qp-min_delta, base_qp+max_delta].
+ */
+ ASSERT_RET(encoder_context->brc.roi_value_is_qp_delta, VA_STATUS_ERROR_INVALID_PARAMETER);
+
+ num_roi = encoder_context->brc.num_roi;
+
+ /* when the base_qp is lower than 12, the quality is quite good based
+ * on the H264 test experience.
+ * In such case it is unnecessary to adjust the quality for ROI region.
+ */
+ if (base_qp <= 12) {
+ nonroi_qp = base_qp;
+ quickfill = 1;
+ goto qp_fill;
+ }
+
+ sum_roi = 0.0f;
+ roi_area = 0;
+ for (i = 0; i < num_roi; i++) {
+ int row_start, row_end, col_start, col_end;
+ int roi_width_mbs, roi_height_mbs;
+ int mbs_in_roi;
+ int roi_qp;
+ float qstep_roi;
+
+ col_start = encoder_context->brc.roi[i].left;
+ col_end = encoder_context->brc.roi[i].right;
+ row_start = encoder_context->brc.roi[i].top;
+ row_end = encoder_context->brc.roi[i].bottom;
+
+ col_start = col_start / 16;
+ col_end = (col_end + 15) / 16;
+ row_start = row_start / 16;
+ row_end = (row_end + 15) / 16;
+
+ roi_width_mbs = col_end - col_start;
+ roi_height_mbs = row_end - row_start;
+ mbs_in_roi = roi_width_mbs * roi_height_mbs;
+
+ param_regions[i].row_start_in_mb = row_start;
+ param_regions[i].row_end_in_mb = row_end;
+ param_regions[i].col_start_in_mb = col_start;
+ param_regions[i].col_end_in_mb = col_end;
+ param_regions[i].width_mbs = roi_width_mbs;
+ param_regions[i].height_mbs = roi_height_mbs;
+
+ roi_qp = base_qp + encoder_context->brc.roi[i].value;
+ BRC_CLIP(roi_qp, min_qp, 51);
+
+ param_regions[i].roi_qp = roi_qp;
+ qstep_roi = intel_h264_qp_qstep(roi_qp);
+
+ roi_area += mbs_in_roi;
+ sum_roi += mbs_in_roi / qstep_roi;
+ }
+
+ total_area = mbs_in_picture;
+ nonroi_area = total_area - roi_area;
+
+ qstep_base = intel_h264_qp_qstep(base_qp);
+ temp = (total_area / qstep_base - sum_roi);
+
+ if (temp < 0) {
+ nonroi_qp = 51;
+ } else {
+ qstep_nonroi = nonroi_area / temp;
+ nonroi_qp = intel_h264_qstep_qp(qstep_nonroi);
+ }
+
+ BRC_CLIP(nonroi_qp, min_qp, 51);
+
+qp_fill:
+ memset(vme_context->qp_per_mb, nonroi_qp, mbs_in_picture);
+ if (!quickfill) {
+ char *qp_ptr;
+
+ for (i = 0; i < num_roi; i++) {
+ for (j = param_regions[i].row_start_in_mb; j < param_regions[i].row_end_in_mb; j++) {
+ qp_ptr = vme_context->qp_per_mb + (j * width_in_mbs) + param_regions[i].col_start_in_mb;
+ memset(qp_ptr, param_regions[i].roi_qp, param_regions[i].width_mbs);
+ }
+ }
+ }
+ return vaStatus;
+}
+
+extern void
+intel_h264_enc_roi_config(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ char *qp_ptr;
+ int i, j;
+ struct i965_driver_data *i965 = i965_driver_data(ctx);
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
+ int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
+
+ int row_start, row_end, col_start, col_end;
+ int num_roi = 0;
+
+ vme_context->roi_enabled = 0;
+ /* Restriction: Disable ROI when multi-slice is enabled */
+ if (!encoder_context->context_roi || (encode_state->num_slice_params_ext > 1))
+ return;
+
+ vme_context->roi_enabled = !!encoder_context->brc.num_roi;
+
+ if (!vme_context->roi_enabled)
+ return;
+
+ if ((vme_context->saved_width_mbs != width_in_mbs) ||
+ (vme_context->saved_height_mbs != height_in_mbs)) {
+ free(vme_context->qp_per_mb);
+ vme_context->qp_per_mb = calloc(1, width_in_mbs * height_in_mbs);
+
+ vme_context->saved_width_mbs = width_in_mbs;
+ vme_context->saved_height_mbs = height_in_mbs;
+ assert(vme_context->qp_per_mb);
+ }
+ if (encoder_context->rate_control_mode == VA_RC_CBR) {
+ /*
+ * TODO: More complex Qp adjust needs to be added.
+ * Currently it is initialized to slice_qp.
+ */
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int qp;
+ int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+
+ qp = mfc_context->brc.qp_prime_y[encoder_context->layer.curr_frame_layer_id][slice_type];
+ intel_h264_enc_roi_cbr(ctx, qp, encode_state, encoder_context);
+
+ } else if (encoder_context->rate_control_mode == VA_RC_CQP){
+ VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
+ VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int qp;
+ int min_qp = MAX(1, encoder_context->brc.min_qp);
+
+ qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
+ memset(vme_context->qp_per_mb, qp, width_in_mbs * height_in_mbs);
+
+
+ for (j = num_roi; j ; j--) {
+ int qp_delta, qp_clip;
+
+ col_start = encoder_context->brc.roi[i].left;
+ col_end = encoder_context->brc.roi[i].right;
+ row_start = encoder_context->brc.roi[i].top;
+ row_end = encoder_context->brc.roi[i].bottom;
+
+ col_start = col_start / 16;
+ col_end = (col_end + 15) / 16;
+ row_start = row_start / 16;
+ row_end = (row_end + 15) / 16;
+
+ qp_delta = encoder_context->brc.roi[i].value;
+ qp_clip = qp + qp_delta;
+
+ BRC_CLIP(qp_clip, min_qp, 51);
+
+ for (i = row_start; i < row_end; i++) {
+ qp_ptr = vme_context->qp_per_mb + (i * width_in_mbs) + col_start;
+ memset(qp_ptr, qp_clip, (col_end - col_start));
+ }
+ }
+ } else {
+ /*
+ * TODO: Disable it for non CBR-CQP.
+ */
+ vme_context->roi_enabled = 0;
+ }
+
+ if (vme_context->roi_enabled && IS_GEN7(i965->intel.device_info))
+ encoder_context->soft_batch_force = 1;
+
+ return;
+}
+
/* HEVC */
static int
hevc_temporal_find_surface(VAPictureHEVC *curr_pic,
struct object_surface *obj_surface = NULL;
struct i965_driver_data *i965 = i965_driver_data(ctx);
VASurfaceID ref_surface_id;
+ VAEncSequenceParameterBufferHEVC *pSequenceParameter = (VAEncSequenceParameterBufferHEVC *)encode_state->seq_param_ext->buffer;
VAEncPictureParameterBufferHEVC *pic_param = (VAEncPictureParameterBufferHEVC *)encode_state->pic_param_ext->buffer;
VAEncSliceParameterBufferHEVC *slice_param = (VAEncSliceParameterBufferHEVC *)encode_state->slice_params_ext[0]->buffer;
int max_num_references;
VAPictureHEVC *curr_pic;
VAPictureHEVC *ref_list;
int ref_idx;
+ unsigned int is_hevc10 = 0;
+ GenHevcSurface *hevc_encoder_surface = NULL;
+
+ if((pSequenceParameter->seq_fields.bits.bit_depth_luma_minus8 > 0)
+ || (pSequenceParameter->seq_fields.bits.bit_depth_chroma_minus8 > 0))
+ is_hevc10 = 1;
if (list_index == 0) {
max_num_references = pic_param->num_ref_idx_l0_default_active_minus1 + 1;
obj_surface->bo) {
assert(ref_idx >= 0);
vme_context->used_reference_objects[list_index] = obj_surface;
+
+ if(is_hevc10){
+ hevc_encoder_surface = (GenHevcSurface *) obj_surface->private_data;
+ assert(hevc_encoder_surface);
+ obj_surface = hevc_encoder_surface->nv12_surface_obj;
+ }
vme_source_surface_state(ctx, surface_index, obj_surface, encoder_context);
vme_context->ref_index_in_mb[list_index] = (ref_idx << 24 |
ref_idx << 16 |
struct encode_state *encode_state,
struct intel_encoder_context *encoder_context)
{
- //struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
+ struct gen9_hcpe_context *mfc_context = encoder_context->mfc_context;
struct gen6_vme_context *vme_context = encoder_context->vme_context;
VAEncPictureParameterBufferHEVC *pic_param = (VAEncPictureParameterBufferHEVC *)encode_state->pic_param_ext->buffer;
VAEncSliceParameterBufferHEVC *slice_param = (VAEncSliceParameterBufferHEVC *)encode_state->slice_params_ext[0]->buffer;
+ VAEncSequenceParameterBufferHEVC *pSequenceParameter = (VAEncSequenceParameterBufferHEVC *)encode_state->seq_param_ext->buffer;
int qp, m_cost, j, mv_count;
uint8_t *vme_state_message = (uint8_t *)(vme_context->vme_state_message);
float lambda, m_costf;
int slice_type = slice_param->slice_type;
- /* to do for CBR*/
- //if (encoder_context->rate_control_mode == VA_RC_CQP)
qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
- //else
- //qp = mfc_context->bit_rate_control_context[slice_type].QpPrimeY;
+
+ if(encoder_context->rate_control_mode == VA_RC_CBR)
+ {
+ qp = mfc_context->bit_rate_control_context[slice_type].QpPrimeY;
+ if(slice_type == HEVC_SLICE_B) {
+ if(pSequenceParameter->ip_period == 1)
+ {
+ slice_type = HEVC_SLICE_P;
+ qp = mfc_context->bit_rate_control_context[HEVC_SLICE_P].QpPrimeY;
+
+ }else if(mfc_context->vui_hrd.i_frame_number % pSequenceParameter->ip_period == 1){
+ slice_type = HEVC_SLICE_P;
+ qp = mfc_context->bit_rate_control_context[HEVC_SLICE_P].QpPrimeY;
+ }
+ }
+
+ }
if (vme_state_message == NULL)
return;
assert(qp <= QP_MAX);
lambda = intel_lambda_qp(qp);
- if (slice_type == SLICE_TYPE_I) {
+ if (slice_type == HEVC_SLICE_I) {
vme_state_message[MODE_INTRA_16X16] = 0;
m_cost = lambda * 4;
vme_state_message[MODE_INTRA_8X8] = intel_format_lutvalue(m_cost, 0x8f);
m_costf = lambda * 3.5;
m_cost = m_costf;
vme_state_message[MODE_INTRA_NONPRED] = intel_format_lutvalue(m_cost, 0x6f);
- if (slice_type == SLICE_TYPE_P) {
+ if (slice_type == HEVC_SLICE_P) {
m_costf = lambda * 2.5;
m_cost = m_costf;
vme_state_message[MODE_INTER_16X16] = intel_format_lutvalue(m_cost, 0x8f);
OSDN Git Service
