[FFmpeg-devel,V2,5/7] libavfilter/dnn: avoid memcpy for tensorflow dnn output

use TF_Tensor's cpu address to avoid extra memcpy.

Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
---
 libavfilter/dnn_backend_tf.c | 36 ++++++++++++------------------------
 libavfilter/vf_sr.c          |  3 ---
 2 files changed, 12 insertions(+), 27 deletions(-)

Em qua, 24 de abr de 2019 às 23:14, Guo, Yejun <yejun.guo@intel.com> escreveu:
>
> use TF_Tensor's cpu address to avoid extra memcpy.
>
> Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> ---
>  libavfilter/dnn_backend_tf.c | 36 ++++++++++++------------------------
>  libavfilter/vf_sr.c          |  3 ---
>  2 files changed, 12 insertions(+), 27 deletions(-)
>
> diff --git a/libavfilter/dnn_backend_tf.c b/libavfilter/dnn_backend_tf.c
> index 7bee45c..be8401e 100644
> --- a/libavfilter/dnn_backend_tf.c
> +++ b/libavfilter/dnn_backend_tf.c
> @@ -35,6 +35,7 @@ typedef struct TFModel{
>      TF_Status *status;
>      TF_Output input, output;
>      TF_Tensor *input_tensor;
> +    TF_Tensor *output_tensor;
>  } TFModel;
>
>  static void free_buffer(void *data, size_t length)
> @@ -460,13 +461,11 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename)
>          return NULL;
>      }
>
> -    tf_model = av_malloc(sizeof(TFModel));
> +    tf_model = av_mallocz(sizeof(TFModel));
>      if (!tf_model){
>          av_freep(&model);
>          return NULL;
>      }
> -    tf_model->session = NULL;
> -    tf_model->input_tensor = NULL;
>
>      if (load_tf_model(tf_model, model_filename) != DNN_SUCCESS){
>          if (load_native_model(tf_model, model_filename) != DNN_SUCCESS){
> @@ -488,36 +487,22 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename)
>  DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *output)
>  {
>      TFModel *tf_model = (TFModel *)model->model;
> -    TF_Tensor *output_tensor;
> -    uint64_t count;
> -    uint64_t old_count = output->height * output->width * output->channels * sizeof(float);
> +    if (tf_model->output_tensor)
> +        TF_DeleteTensor(tf_model->output_tensor);
>
>      TF_SessionRun(tf_model->session, NULL,
>                    &tf_model->input, &tf_model->input_tensor, 1,
> -                  &tf_model->output, &output_tensor, 1,
> +                  &tf_model->output, &tf_model->output_tensor, 1,
>                    NULL, 0, NULL, tf_model->status);
>
>      if (TF_GetCode(tf_model->status) != TF_OK){
>          return DNN_ERROR;
>      }
>
> -    output->height = TF_Dim(output_tensor, 1);
> -    output->width = TF_Dim(output_tensor, 2);
> -    output->channels = TF_Dim(output_tensor, 3);
> -    count = output->height * output->width * output->channels * sizeof(float);
> -    if (output->data) {
> -        if (count > old_count) {
> -            av_freep(&output->data);
> -        }
> -    }
> -    if (!output->data) {
> -        output->data = av_malloc(count);
> -        if (!output->data){
> -            return DNN_ERROR;
> -        }
> -    }
> -    memcpy(output->data, TF_TensorData(output_tensor), count);
> -    TF_DeleteTensor(output_tensor);
> +    output->height = TF_Dim(tf_model->output_tensor, 1);
> +    output->width = TF_Dim(tf_model->output_tensor, 2);
> +    output->channels = TF_Dim(tf_model->output_tensor, 3);
> +    output->data = TF_TensorData(tf_model->output_tensor);
>
>      return DNN_SUCCESS;
>  }
> @@ -541,6 +526,9 @@ void ff_dnn_free_model_tf(DNNModel **model)
>          if (tf_model->input_tensor){
>              TF_DeleteTensor(tf_model->input_tensor);
>          }
> +        if (tf_model->output_tensor){
> +            TF_DeleteTensor(tf_model->output_tensor);
> +        }
>          av_freep(&tf_model);
>          av_freep(model);
>      }
> diff --git a/libavfilter/vf_sr.c b/libavfilter/vf_sr.c
> index 7c92730..53bd8ea 100644
> --- a/libavfilter/vf_sr.c
> +++ b/libavfilter/vf_sr.c
> @@ -277,9 +277,6 @@ static av_cold void uninit(AVFilterContext *context)
>      int i;
>      SRContext *sr_context = context->priv;
>
> -    if (sr_context->backend_type == DNN_TF)
> -        av_freep(&sr_context->output.data);
> -
>      if (sr_context->dnn_module){
>          (sr_context->dnn_module->free_model)(&sr_context->model);
>          av_freep(&sr_context->dnn_module);
> --
> 2.7.4
>

LGTM.

> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".