[Libva] [PATCH 20/31] ENC: ME kernel for AVC encoder
Sean V Kelley
seanvk at posteo.de
Tue Jan 10 23:38:02 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 | 441 +++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 441 insertions(+)
diff --git a/src/gen9_avc_encoder.c b/src/gen9_avc_encoder.c
index af581fc7..d64694ca 100755
--- a/src/gen9_avc_encoder.c
+++ b/src/gen9_avc_encoder.c
@@ -3771,3 +3771,444 @@ gen9_avc_kernel_mbenc(VADriverContextP ctx,
&media_object_walker_param);
return VA_STATUS_SUCCESS;
}
+
+/*
+me kernle related function
+*/
+static void
+gen9_avc_set_curbe_me(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct i965_gpe_context *gpe_context,
+ struct intel_encoder_context *encoder_context,
+ void * param)
+{
+ gen9_avc_me_curbe_data *curbe_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;
+
+ VAEncSliceParameterBufferH264 * slice_param = avc_state->slice_param[0];
+
+ struct me_param * curbe_param = (struct me_param *)param ;
+ unsigned char use_mv_from_prev_step = 0;
+ unsigned char write_distortions = 0;
+ unsigned char qp_prime_y = 0;
+ unsigned char me_method = gen9_avc_p_me_method[generic_state->preset];
+ unsigned char seach_table_idx = 0;
+ unsigned char mv_shift_factor = 0, prev_mv_read_pos_factor = 0;
+ unsigned int downscaled_width_in_mb, downscaled_height_in_mb;
+ unsigned int scale_factor = 0;
+
+ qp_prime_y = avc_state->pic_param->pic_init_qp + slice_param->slice_qp_delta;
+ switch(curbe_param->hme_type)
+ {
+ case INTEL_ENC_HME_4x :
+ {
+ use_mv_from_prev_step = (generic_state->b16xme_enabled)? 1:0;
+ write_distortions = 1;
+ mv_shift_factor = 2;
+ scale_factor = 4;
+ prev_mv_read_pos_factor = 0;
+ break;
+ }
+ case INTEL_ENC_HME_16x :
+ {
+ use_mv_from_prev_step = (generic_state->b32xme_enabled)? 1:0;
+ write_distortions = 0;
+ mv_shift_factor = 2;
+ scale_factor = 16;
+ prev_mv_read_pos_factor = 1;
+ break;
+ }
+ case INTEL_ENC_HME_32x :
+ {
+ use_mv_from_prev_step = 0;
+ write_distortions = 0;
+ mv_shift_factor = 1;
+ scale_factor = 32;
+ prev_mv_read_pos_factor = 0;
+ break;
+ }
+ default:
+ assert(0);
+
+ }
+ curbe_cmd = i965_gpe_context_map_curbe(gpe_context);
+
+ if (!curbe_cmd)
+ return;
+
+ downscaled_width_in_mb = ALIGN(generic_state->frame_width_in_pixel/scale_factor,16)/16;
+ downscaled_height_in_mb = ALIGN(generic_state->frame_height_in_pixel/scale_factor,16)/16;
+
+ memcpy(curbe_cmd,gen9_avc_me_curbe_init_data,sizeof(gen9_avc_me_curbe_data));
+
+ curbe_cmd->dw3.sub_pel_mode = 3;
+ if(avc_state->field_scaling_output_interleaved)
+ {
+ /*frame set to zero,field specified*/
+ curbe_cmd->dw3.src_access = 0;
+ curbe_cmd->dw3.ref_access = 0;
+ curbe_cmd->dw7.src_field_polarity = 0;
+ }
+ curbe_cmd->dw4.picture_height_minus1 = downscaled_height_in_mb - 1;
+ curbe_cmd->dw4.picture_width = downscaled_width_in_mb;
+ curbe_cmd->dw5.qp_prime_y = qp_prime_y;
+
+ curbe_cmd->dw6.use_mv_from_prev_step = use_mv_from_prev_step;
+ curbe_cmd->dw6.write_distortions = write_distortions;
+ curbe_cmd->dw6.super_combine_dist = gen9_avc_super_combine_dist[generic_state->preset];
+ curbe_cmd->dw6.max_vmvr = i965_avc_get_max_mv_len(avc_state->seq_param->level_idc) * 4;//frame only
+
+ if(generic_state->frame_type == SLICE_TYPE_B)
+ {
+ curbe_cmd->dw1.bi_weight = 32;
+ curbe_cmd->dw13.num_ref_idx_l1_minus1 = slice_param->num_ref_idx_l1_active_minus1;
+ me_method = gen9_avc_b_me_method[generic_state->preset];
+ seach_table_idx = 1;
+ }
+
+ if(generic_state->frame_type == SLICE_TYPE_P ||
+ generic_state->frame_type == SLICE_TYPE_B )
+ curbe_cmd->dw13.num_ref_idx_l0_minus1 = slice_param->num_ref_idx_l0_active_minus1;
+
+ curbe_cmd->dw13.ref_streamin_cost = 5;
+ curbe_cmd->dw13.roi_enable = 0;
+
+ curbe_cmd->dw15.prev_mv_read_pos_factor = prev_mv_read_pos_factor;
+ curbe_cmd->dw15.mv_shift_factor = mv_shift_factor;
+
+ memcpy(&curbe_cmd->dw16,table_enc_search_path[seach_table_idx][me_method],14*sizeof(int));
+
+ curbe_cmd->dw32._4x_memv_output_data_surf_index = GEN9_AVC_ME_MV_DATA_SURFACE_INDEX;
+ curbe_cmd->dw33._16x_32x_memv_input_data_surf_index = (curbe_param->hme_type == INTEL_ENC_HME_32x)? GEN9_AVC_32XME_MV_DATA_SURFACE_INDEX:GEN9_AVC_16XME_MV_DATA_SURFACE_INDEX ;
+ curbe_cmd->dw34._4x_me_output_dist_surf_index = GEN9_AVC_ME_DISTORTION_SURFACE_INDEX;
+ curbe_cmd->dw35._4x_me_output_brc_dist_surf_index = GEN9_AVC_ME_BRC_DISTORTION_INDEX;
+ curbe_cmd->dw36.vme_fwd_inter_pred_surf_index = GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX;
+ curbe_cmd->dw37.vme_bdw_inter_pred_surf_index = GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX;
+ curbe_cmd->dw38.reserved = GEN9_AVC_ME_VDENC_STREAMIN_INDEX;
+
+ i965_gpe_context_unmap_curbe(gpe_context);
+ return;
+}
+
+static void
+gen9_avc_send_surface_me(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct i965_gpe_context *gpe_context,
+ struct intel_encoder_context *encoder_context,
+ void * param)
+{
+ struct i965_driver_data *i965 = i965_driver_data(ctx);
+
+ 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 gen9_avc_encoder_context * avc_ctx = (struct gen9_avc_encoder_context * )vme_context->private_enc_ctx;
+ struct avc_enc_state * avc_state = (struct avc_enc_state * )vme_context->private_enc_state;
+
+ struct object_surface *obj_surface, *input_surface;
+ struct gen9_surface_avc *avc_priv_surface;
+ struct i965_gpe_resource *gpe_resource;
+ struct me_param * curbe_param = (struct me_param *)param ;
+
+ VAEncSliceParameterBufferH264 * slice_param = avc_state->slice_param[0];
+ VASurfaceID surface_id;
+ int i = 0;
+
+ /* all scaled input surface stored in reconstructed_object*/
+ obj_surface = encode_state->reconstructed_object;
+ if (!obj_surface || !obj_surface->private_data)
+ return;
+ avc_priv_surface = obj_surface->private_data;
+
+
+ switch(curbe_param->hme_type)
+ {
+ case INTEL_ENC_HME_4x :
+ {
+ /*memv output 4x*/
+ gpe_resource = &avc_ctx->s4x_memv_data_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_ME_MV_DATA_SURFACE_INDEX);
+
+ /*memv input 16x*/
+ if(generic_state->b16xme_enabled)
+ {
+ gpe_resource = &avc_ctx->s16x_memv_data_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_16XME_MV_DATA_SURFACE_INDEX);
+ }
+ /* brc distortion output*/
+ gpe_resource = &avc_ctx->res_brc_dist_data_surface;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_ME_BRC_DISTORTION_INDEX);
+ /* memv distortion output*/
+ gpe_resource = &avc_ctx->s4x_memv_distortion_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_ME_DISTORTION_SURFACE_INDEX);
+ /*input current down scaled YUV surface*/
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_4x_surface_obj;
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX);
+ /*input ref scaled YUV surface*/
+ for(i = 0; i < slice_param->num_ref_idx_l0_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList0[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_4x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX+i*2 + 1);
+ }
+
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_4x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX);
+
+ for(i = 0; i < slice_param->num_ref_idx_l1_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList1[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_4x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX+i*2 + 1);
+ }
+ break;
+
+ }
+ case INTEL_ENC_HME_16x :
+ {
+ gpe_resource = &avc_ctx->s16x_memv_data_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_ME_MV_DATA_SURFACE_INDEX);
+
+ if(generic_state->b32xme_enabled)
+ {
+ gpe_resource = &avc_ctx->s32x_memv_data_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_32XME_MV_DATA_SURFACE_INDEX);
+ }
+
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_16x_surface_obj;
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX);
+
+ for(i = 0; i < slice_param->num_ref_idx_l0_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList0[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_16x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX+i*2 + 1);
+ }
+
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_16x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX);
+
+ for(i = 0; i < slice_param->num_ref_idx_l1_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList1[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_16x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX+i*2 + 1);
+ }
+ break;
+ }
+ case INTEL_ENC_HME_32x :
+ {
+ gpe_resource = &avc_ctx->s32x_memv_data_buffer;
+ gen9_add_buffer_2d_gpe_surface(ctx, gpe_context,
+ gpe_resource,
+ 1,
+ I965_SURFACEFORMAT_R8_UNORM,
+ GEN9_AVC_ME_MV_DATA_SURFACE_INDEX);
+
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_32x_surface_obj;
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX);
+
+ for(i = 0; i < slice_param->num_ref_idx_l0_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList0[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_32x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_FWD_REF_INDEX+i*2 + 1);
+ }
+
+ obj_surface = encode_state->reconstructed_object;
+ avc_priv_surface = obj_surface->private_data;
+ input_surface = avc_priv_surface->scaled_32x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX);
+
+ for(i = 0; i < slice_param->num_ref_idx_l1_active_minus1 + 1; i++)
+ {
+ surface_id = slice_param->RefPicList1[i].picture_id;
+ obj_surface = SURFACE(surface_id);
+ if (!obj_surface || !obj_surface->private_data)
+ break;
+ avc_priv_surface = obj_surface->private_data;
+
+ input_surface = avc_priv_surface->scaled_32x_surface_obj;
+
+ gen9_add_adv_gpe_surface(ctx, gpe_context,
+ input_surface,
+ GEN9_AVC_ME_CURR_FOR_BWD_REF_INDEX+i*2 + 1);
+ }
+ break;
+ }
+ default:
+ assert(0);
+
+ }
+}
+
+static VAStatus
+gen9_avc_kernel_me(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context,
+ int hme_type)
+{
+ struct encoder_vme_mfc_context * vme_context = (struct encoder_vme_mfc_context *)encoder_context->vme_context;
+ struct generic_encoder_context * generic_ctx = (struct generic_encoder_context * )vme_context->generic_enc_ctx;
+ 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 i965_gpe_context *gpe_context;
+ struct gpe_media_object_walker_parameter media_object_walker_param;
+ struct gpe_encoder_kernel_walker_parameter kernel_walker_param;
+ unsigned int downscaled_width_in_mb, downscaled_height_in_mb;
+ int media_function = 0;
+ int kernel_idx = 0;
+ struct me_param param ;
+ unsigned int scale_factor = 0;
+
+ switch(hme_type)
+ {
+ case INTEL_ENC_HME_4x :
+ {
+ media_function = INTEL_MEDIA_STATE_4X_ME;
+ scale_factor = 4;
+ break;
+ }
+ case INTEL_ENC_HME_16x :
+ {
+ media_function = INTEL_MEDIA_STATE_16X_ME;
+ scale_factor = 16;
+ break;
+ }
+ case INTEL_ENC_HME_32x :
+ {
+ media_function = INTEL_MEDIA_STATE_32X_ME;
+ scale_factor = 32;
+ break;
+ }
+ default:
+ assert(0);
+
+ }
+
+ downscaled_width_in_mb = ALIGN(generic_state->frame_width_in_pixel/scale_factor,16)/16;
+ downscaled_height_in_mb = ALIGN(generic_state->frame_height_in_pixel/scale_factor,16)/16;
+
+ /* I frame should not come here.*/
+ kernel_idx = (generic_state->frame_type == SLICE_TYPE_P)? GEN9_AVC_KERNEL_ME_P_IDX : GEN9_AVC_KERNEL_ME_B_IDX;
+ gpe_context = &(avc_ctx->context_me.gpe_contexts[kernel_idx]);
+
+ gen8_gpe_context_init(ctx, gpe_context);
+ gen9_gpe_reset_binding_table(ctx, gpe_context);
+
+ /*set curbe*/
+ memset(¶m,0,sizeof(param));
+ param.hme_type = hme_type;
+ generic_ctx->pfn_set_curbe_me(ctx,encode_state,gpe_context,encoder_context,¶m);
+
+ /*send surface*/
+ generic_ctx->pfn_send_me_surface(ctx,encode_state,gpe_context,encoder_context,¶m);
+
+ gen8_gpe_setup_interface_data(ctx, gpe_context);
+
+ memset(&kernel_walker_param, 0, sizeof(kernel_walker_param));
+ /* the scaling is based on 8x8 blk level */
+ kernel_walker_param.resolution_x = downscaled_width_in_mb ;
+ kernel_walker_param.resolution_y = downscaled_height_in_mb ;
+ kernel_walker_param.no_dependency = 1;
+
+ i965_init_media_object_walker_parameter(&kernel_walker_param, &media_object_walker_param);
+
+ gen9_avc_run_kernel_media_object_walker(ctx, encoder_context,
+ gpe_context,
+ media_function,
+ &media_object_walker_param);
+
+ return VA_STATUS_SUCCESS;
+}
--
2.11.0
More information about the Libva
mailing list