[Libva] [PATCH 15/31] ENC: add BRC init/reset kernel for AVC RC logic
Sean V Kelley
seanvk at posteo.de
Tue Jan 10 23:37:57 UTC 2017
From: Pengfei Qu <Pengfei.Qu at intel.com>
Signed-off-by: Pengfei Qu <Pengfei.Qu at intel.com>
Reviewed-by: Sean V Kelley <seanvk at posteo.de>
---
src/gen9_avc_encoder.c | 218 +++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 218 insertions(+)
diff --git a/src/gen9_avc_encoder.c b/src/gen9_avc_encoder.c
index 878345ee..fd042c5f 100755
--- a/src/gen9_avc_encoder.c
+++ b/src/gen9_avc_encoder.c
@@ -1930,3 +1930,221 @@ gen9_avc_init_brc_const_data_old(VADriverContextP ctx,
i965_unmap_gpe_resource(gpe_resource);
}
+static void
+gen9_avc_set_curbe_brc_init_reset(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct i965_gpe_context *gpe_context,
+ struct intel_encoder_context *encoder_context,
+ void * param)
+{
+ gen9_avc_brc_init_reset_curbe_data *cmd;
+ struct encoder_vme_mfc_context * vme_context = (struct encoder_vme_mfc_context *)encoder_context->vme_context;
+ struct generic_enc_codec_state * generic_state = (struct generic_enc_codec_state * )vme_context->generic_enc_state;
+ struct avc_enc_state * avc_state = (struct avc_enc_state * )vme_context->private_enc_state;
+ double input_bits_per_frame = 0;
+ double bps_ratio = 0;
+ VAEncSequenceParameterBufferH264 * seq_param = avc_state->seq_param;
+ struct avc_param common_param;
+
+ cmd = i965_gpe_context_map_curbe(gpe_context);
+
+ memcpy(cmd,&gen9_avc_brc_init_reset_curbe_init_data,sizeof(gen9_avc_brc_init_reset_curbe_data));
+
+ memset(&common_param,0,sizeof(common_param));
+ common_param.frame_width_in_pixel = generic_state->frame_width_in_pixel;
+ common_param.frame_height_in_pixel = generic_state->frame_height_in_pixel;
+ common_param.frame_width_in_mbs = generic_state->frame_width_in_mbs;
+ common_param.frame_height_in_mbs = generic_state->frame_height_in_mbs;
+ common_param.frames_per_100s = generic_state->frames_per_100s;
+ common_param.vbv_buffer_size_in_bit = generic_state->vbv_buffer_size_in_bit;
+ common_param.target_bit_rate = generic_state->target_bit_rate;
+
+ cmd->dw0.profile_level_max_frame = i965_avc_get_profile_level_max_frame(&common_param,seq_param->level_idc);
+ cmd->dw1.init_buf_full_in_bits = generic_state->init_vbv_buffer_fullness_in_bit;
+ cmd->dw2.buf_size_in_bits = generic_state->vbv_buffer_size_in_bit;
+ cmd->dw3.average_bit_rate = generic_state->target_bit_rate * 1000;
+ cmd->dw4.max_bit_rate = generic_state->max_bit_rate * 1000;
+ cmd->dw8.gop_p = (generic_state->gop_ref_distance)?((generic_state->gop_size -1)/generic_state->gop_ref_distance):0;
+ cmd->dw9.gop_b = (generic_state->gop_size - 1 - cmd->dw8.gop_p);
+ cmd->dw9.frame_width_in_bytes = generic_state->frame_width_in_pixel;
+ cmd->dw10.frame_height_in_bytes = generic_state->frame_height_in_pixel;
+ cmd->dw12.no_slices = avc_state->slice_num;
+
+ //VUI
+ if(seq_param->vui_parameters_present_flag && generic_state->internal_rate_mode != INTEL_BRC_AVBR )
+ {
+ cmd->dw4.max_bit_rate = cmd->dw4.max_bit_rate;
+ if(generic_state->internal_rate_mode == INTEL_BRC_CBR)
+ {
+ cmd->dw3.average_bit_rate = cmd->dw4.max_bit_rate;
+
+ }
+
+ }
+ cmd->dw6.frame_rate_m = generic_state->frames_per_100s;
+ cmd->dw7.frame_rate_d = 100;
+ cmd->dw8.brc_flag = 0;
+ cmd->dw8.brc_flag |= (generic_state->mb_brc_enabled)? 0 : 0x8000;
+
+
+ if(generic_state->internal_rate_mode == INTEL_BRC_CBR)
+ { //CBR
+ cmd->dw4.max_bit_rate = cmd->dw3.average_bit_rate;
+ cmd->dw8.brc_flag = cmd->dw8.brc_flag |INTEL_ENCODE_BRCINIT_ISCBR;
+
+ }else if(generic_state->internal_rate_mode == INTEL_BRC_VBR)
+ {//VBR
+ if(cmd->dw4.max_bit_rate < cmd->dw3.average_bit_rate)
+ {
+ cmd->dw4.max_bit_rate = cmd->dw3.average_bit_rate << 1;
+ }
+ cmd->dw8.brc_flag = cmd->dw8.brc_flag |INTEL_ENCODE_BRCINIT_ISVBR;
+
+ }else if(generic_state->internal_rate_mode == INTEL_BRC_AVBR)
+ { //AVBR
+ cmd->dw4.max_bit_rate =cmd->dw3.average_bit_rate;
+ cmd->dw8.brc_flag = cmd->dw8.brc_flag |INTEL_ENCODE_BRCINIT_ISAVBR;
+
+ }
+ //igonre icq/vcm/qvbr
+
+ cmd->dw10.avbr_accuracy = generic_state->avbr_curracy;
+ cmd->dw11.avbr_convergence = generic_state->avbr_convergence;
+
+ //frame bits
+ input_bits_per_frame = (double)(cmd->dw4.max_bit_rate) * (double)(cmd->dw7.frame_rate_d)/(double)(cmd->dw6.frame_rate_m);;
+
+ if(cmd->dw2.buf_size_in_bits == 0)
+ {
+ cmd->dw2.buf_size_in_bits = (unsigned int)(input_bits_per_frame * 4);
+ }
+
+ if(cmd->dw1.init_buf_full_in_bits == 0)
+ {
+ cmd->dw1.init_buf_full_in_bits = cmd->dw2.buf_size_in_bits * 7/8;
+ }
+ if(cmd->dw1.init_buf_full_in_bits < (unsigned int)(input_bits_per_frame * 2))
+ {
+ cmd->dw1.init_buf_full_in_bits = (unsigned int)(input_bits_per_frame * 2);
+ }
+ if(cmd->dw1.init_buf_full_in_bits > cmd->dw2.buf_size_in_bits)
+ {
+ cmd->dw1.init_buf_full_in_bits = cmd->dw2.buf_size_in_bits;
+ }
+
+ //AVBR
+ if(generic_state->internal_rate_mode == INTEL_BRC_AVBR)
+ {
+ cmd->dw2.buf_size_in_bits = 2 * generic_state->target_bit_rate * 1000;
+ cmd->dw1.init_buf_full_in_bits = (unsigned int)(3 * cmd->dw2.buf_size_in_bits/4);
+
+ }
+
+ bps_ratio = input_bits_per_frame / (cmd->dw2.buf_size_in_bits/30.0);
+ bps_ratio = (bps_ratio < 0.1)? 0.1:(bps_ratio > 3.5)?3.5:bps_ratio;
+
+
+ cmd->dw16.deviation_threshold_0_pand_b = (unsigned int)(-50 * pow(0.90,bps_ratio));
+ cmd->dw16.deviation_threshold_1_pand_b = (unsigned int)(-50 * pow(0.66,bps_ratio));
+ cmd->dw16.deviation_threshold_2_pand_b = (unsigned int)(-50 * pow(0.46,bps_ratio));
+ cmd->dw16.deviation_threshold_3_pand_b = (unsigned int)(-50 * pow(0.3, bps_ratio));
+ cmd->dw17.deviation_threshold_4_pand_b = (unsigned int)(50 * pow(0.3, bps_ratio));
+ cmd->dw17.deviation_threshold_5_pand_b = (unsigned int)(50 * pow(0.46, bps_ratio));
+ cmd->dw17.deviation_threshold_6_pand_b = (unsigned int)(50 * pow(0.7, bps_ratio));
+ cmd->dw17.deviation_threshold_7_pand_b = (unsigned int)(50 * pow(0.9, bps_ratio));
+ cmd->dw18.deviation_threshold_0_vbr = (unsigned int)(-50 * pow(0.9, bps_ratio));
+ cmd->dw18.deviation_threshold_1_vbr = (unsigned int)(-50 * pow(0.7, bps_ratio));
+ cmd->dw18.deviation_threshold_2_vbr = (unsigned int)(-50 * pow(0.5, bps_ratio));
+ cmd->dw18.deviation_threshold_3_vbr = (unsigned int)(-50 * pow(0.3, bps_ratio));
+ cmd->dw19.deviation_threshold_4_vbr = (unsigned int)(100 * pow(0.4, bps_ratio));
+ cmd->dw19.deviation_threshold_5_vbr = (unsigned int)(100 * pow(0.5, bps_ratio));
+ cmd->dw19.deviation_threshold_6_vbr = (unsigned int)(100 * pow(0.75,bps_ratio));
+ cmd->dw19.deviation_threshold_7_vbr = (unsigned int)(100 * pow(0.9, bps_ratio));
+ cmd->dw20.deviation_threshold_0_i = (unsigned int)(-50 * pow(0.8, bps_ratio));
+ cmd->dw20.deviation_threshold_1_i = (unsigned int)(-50 * pow(0.6, bps_ratio));
+ cmd->dw20.deviation_threshold_2_i = (unsigned int)(-50 * pow(0.34,bps_ratio));
+ cmd->dw20.deviation_threshold_3_i = (unsigned int)(-50 * pow(0.2, bps_ratio));
+ cmd->dw21.deviation_threshold_4_i = (unsigned int)(50 * pow(0.2, bps_ratio));
+ cmd->dw21.deviation_threshold_5_i = (unsigned int)(50 * pow(0.4, bps_ratio));
+ cmd->dw21.deviation_threshold_6_i = (unsigned int)(50 * pow(0.66, bps_ratio));
+ cmd->dw21.deviation_threshold_7_i = (unsigned int)(50 * pow(0.9, bps_ratio));
+
+ cmd->dw22.sliding_window_size = generic_state->window_size;
+
+ i965_gpe_context_unmap_curbe(gpe_context);
+
+ return;
+}
+
+static void
+gen9_avc_send_surface_brc_init_reset(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct i965_gpe_context *gpe_context,
+ struct intel_encoder_context *encoder_context,
+ void * param_mbenc)
+{
+ struct encoder_vme_mfc_context * vme_context = (struct encoder_vme_mfc_context *)encoder_context->vme_context;
+ struct gen9_avc_encoder_context * avc_ctx = (struct gen9_avc_encoder_context * )vme_context->private_enc_ctx;
+
+ gen9_add_buffer_gpe_surface(ctx,
+ gpe_context,
+ &avc_ctx->res_brc_history_buffer,
+ 0,
+ avc_ctx->res_brc_history_buffer.size,
+ 0,
+ GEN9_AVC_BRC_INIT_RESET_HISTORY_INDEX);
+
+ gen9_add_buffer_2d_gpe_surface(ctx,
+ gpe_context,
+ &avc_ctx->res_brc_dist_data_surface,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_BRC_INIT_RESET_DISTORTION_INDEX);
+
+ return;
+}
+
+static VAStatus
+gen9_avc_kernel_brc_init_reset(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ struct encoder_vme_mfc_context * vme_context = (struct encoder_vme_mfc_context *)encoder_context->vme_context;
+ struct gen9_avc_encoder_context * avc_ctx = (struct gen9_avc_encoder_context * )vme_context->private_enc_ctx;
+ struct generic_enc_codec_state * generic_state = (struct generic_enc_codec_state * )vme_context->generic_enc_state;
+ struct generic_encoder_context * generic_ctx = (struct generic_encoder_context * )vme_context->generic_enc_ctx;
+
+ struct i965_gpe_context *gpe_context;
+ struct gpe_media_object_parameter media_object_param;
+ struct gpe_media_object_inline_data media_object_inline_data;
+ int media_function = 0;
+ int kernel_idx = GEN9_AVC_KERNEL_BRC_INIT;
+
+ media_function = INTEL_MEDIA_STATE_BRC_INIT_RESET;
+
+ if(generic_state->brc_inited)
+ kernel_idx = GEN9_AVC_KERNEL_BRC_RESET;
+
+ gpe_context = &(avc_ctx->context_brc.gpe_contexts[kernel_idx]);
+
+ gen8_gpe_context_init(ctx, gpe_context);
+ gen9_gpe_reset_binding_table(ctx, gpe_context);
+
+ generic_ctx->pfn_set_curbe_brc_init_reset(ctx,encode_state,gpe_context,encoder_context,NULL);
+
+ generic_ctx->pfn_send_brc_init_reset_surface(ctx,encode_state,gpe_context,encoder_context,NULL);
+
+ gen8_gpe_setup_interface_data(ctx, gpe_context);
+
+ memset(&media_object_param, 0, sizeof(media_object_param));
+ memset(&media_object_inline_data, 0, sizeof(media_object_inline_data));
+ media_object_param.pinline_data = &media_object_inline_data;
+ media_object_param.inline_size = sizeof(media_object_inline_data);
+
+ gen9_avc_run_kernel_media_object(ctx, encoder_context,
+ gpe_context,
+ media_function,
+ &media_object_param);
+
+ return VA_STATUS_SUCCESS;
+}
--
2.11.0
More information about the Libva
mailing list