#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
#include "libavutil/avassert.h"
+#include "libavutil/imgutils.h"
#include "avfilter.h"
#include "dnn_interface.h"
#include "formats.h"
#include "internal.h"
+#include "libswscale/swscale.h"
typedef struct DnnProcessingContext {
const AVClass *class;
char *model_filename;
DNNBackendType backend_type;
- enum AVPixelFormat fmt;
char *model_inputname;
char *model_outputname;
// input & output of the model at execution time
DNNData input;
DNNData output;
+
+ struct SwsContext *sws_gray8_to_grayf32;
+ struct SwsContext *sws_grayf32_to_gray8;
+ struct SwsContext *sws_uv_scale;
+ int sws_uv_height;
} DnnProcessingContext;
#define OFFSET(x) offsetof(DnnProcessingContext, x)
{ "model", "path to model file", OFFSET(model_filename), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
{ "input", "input name of the model", OFFSET(model_inputname), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
{ "output", "output name of the model", OFFSET(model_outputname), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
- { "fmt", "AVPixelFormat of the frame", OFFSET(fmt), AV_OPT_TYPE_PIXEL_FMT, { .i64=AV_PIX_FMT_RGB24 }, AV_PIX_FMT_NONE, AV_PIX_FMT_NB - 1, FLAGS },
{ NULL }
};
static av_cold int init(AVFilterContext *context)
{
DnnProcessingContext *ctx = context->priv;
- int supported = 0;
- // as the first step, only rgb24 and bgr24 are supported
- const enum AVPixelFormat supported_pixel_fmts[] = {
- AV_PIX_FMT_RGB24,
- AV_PIX_FMT_BGR24,
- };
- for (int i = 0; i < sizeof(supported_pixel_fmts) / sizeof(enum AVPixelFormat); ++i) {
- if (supported_pixel_fmts[i] == ctx->fmt) {
- supported = 1;
- break;
- }
- }
- if (!supported) {
- av_log(context, AV_LOG_ERROR, "pixel fmt %s not supported yet\n",
- av_get_pix_fmt_name(ctx->fmt));
- return AVERROR(AVERROR_INVALIDDATA);
- }
if (!ctx->model_filename) {
av_log(ctx, AV_LOG_ERROR, "model file for network is not specified\n");
static int query_formats(AVFilterContext *context)
{
- AVFilterFormats *formats;
- DnnProcessingContext *ctx = context->priv;
- enum AVPixelFormat pixel_fmts[2];
- pixel_fmts[0] = ctx->fmt;
- pixel_fmts[1] = AV_PIX_FMT_NONE;
-
- formats = ff_make_format_list(pixel_fmts);
- return ff_set_common_formats(context, formats);
+ static const enum AVPixelFormat pix_fmts[] = {
+ AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
+ AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAYF32,
+ AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
+ AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
+ AV_PIX_FMT_NONE
+ };
+ AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
+ return ff_set_common_formats(context, fmts_list);
}
-static int config_input(AVFilterLink *inlink)
-{
- AVFilterContext *context = inlink->dst;
- DnnProcessingContext *ctx = context->priv;
- DNNReturnType result;
- DNNData dnn_data;
+#define LOG_FORMAT_CHANNEL_MISMATCH() \
+ av_log(ctx, AV_LOG_ERROR, \
+ "the frame's format %s does not match " \
+ "the model input channel %d\n", \
+ av_get_pix_fmt_name(fmt), \
+ model_input->channels);
- result = ctx->model->get_input(ctx->model->model, &dnn_data, ctx->model_inputname);
- if (result != DNN_SUCCESS) {
- av_log(ctx, AV_LOG_ERROR, "could not get input from the model\n");
- return AVERROR(EIO);
- }
+static int check_modelinput_inlink(const DNNData *model_input, const AVFilterLink *inlink)
+{
+ AVFilterContext *ctx = inlink->dst;
+ enum AVPixelFormat fmt = inlink->format;
// the design is to add explicit scale filter before this filter
- if (dnn_data.height != -1 && dnn_data.height != inlink->h) {
+ if (model_input->height != -1 && model_input->height != inlink->h) {
av_log(ctx, AV_LOG_ERROR, "the model requires frame height %d but got %d\n",
- dnn_data.height, inlink->h);
+ model_input->height, inlink->h);
return AVERROR(EIO);
}
- if (dnn_data.width != -1 && dnn_data.width != inlink->w) {
+ if (model_input->width != -1 && model_input->width != inlink->w) {
av_log(ctx, AV_LOG_ERROR, "the model requires frame width %d but got %d\n",
- dnn_data.width, inlink->w);
+ model_input->width, inlink->w);
return AVERROR(EIO);
}
- if (dnn_data.channels != 3) {
- av_log(ctx, AV_LOG_ERROR, "the model requires input channels %d\n",
- dnn_data.channels);
+ switch (fmt) {
+ case AV_PIX_FMT_RGB24:
+ case AV_PIX_FMT_BGR24:
+ if (model_input->channels != 3) {
+ LOG_FORMAT_CHANNEL_MISMATCH();
+ return AVERROR(EIO);
+ }
+ if (model_input->dt != DNN_FLOAT && model_input->dt != DNN_UINT8) {
+ av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type as float32 and uint8.\n");
+ return AVERROR(EIO);
+ }
+ return 0;
+ case AV_PIX_FMT_GRAY8:
+ if (model_input->channels != 1) {
+ LOG_FORMAT_CHANNEL_MISMATCH();
+ return AVERROR(EIO);
+ }
+ if (model_input->dt != DNN_UINT8) {
+ av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type uint8.\n");
+ return AVERROR(EIO);
+ }
+ return 0;
+ case AV_PIX_FMT_GRAYF32:
+ case AV_PIX_FMT_YUV420P:
+ case AV_PIX_FMT_YUV422P:
+ case AV_PIX_FMT_YUV444P:
+ case AV_PIX_FMT_YUV410P:
+ case AV_PIX_FMT_YUV411P:
+ if (model_input->channels != 1) {
+ LOG_FORMAT_CHANNEL_MISMATCH();
+ return AVERROR(EIO);
+ }
+ if (model_input->dt != DNN_FLOAT) {
+ av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type float32.\n");
+ return AVERROR(EIO);
+ }
+ return 0;
+ default:
+ av_log(ctx, AV_LOG_ERROR, "%s not supported.\n", av_get_pix_fmt_name(fmt));
return AVERROR(EIO);
}
- if (dnn_data.dt != DNN_FLOAT && dnn_data.dt != DNN_UINT8) {
- av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type as float32 and uint8.\n");
+
+ return 0;
+}
+
+static int config_input(AVFilterLink *inlink)
+{
+ AVFilterContext *context = inlink->dst;
+ DnnProcessingContext *ctx = context->priv;
+ DNNReturnType result;
+ DNNData model_input;
+ int check;
+
+ result = ctx->model->get_input(ctx->model->model, &model_input, ctx->model_inputname);
+ if (result != DNN_SUCCESS) {
+ av_log(ctx, AV_LOG_ERROR, "could not get input from the model\n");
return AVERROR(EIO);
}
+ check = check_modelinput_inlink(&model_input, inlink);
+ if (check != 0) {
+ return check;
+ }
+
ctx->input.width = inlink->w;
ctx->input.height = inlink->h;
- ctx->input.channels = dnn_data.channels;
- ctx->input.dt = dnn_data.dt;
+ ctx->input.channels = model_input.channels;
+ ctx->input.dt = model_input.dt;
result = (ctx->model->set_input_output)(ctx->model->model,
&ctx->input, ctx->model_inputname,
return 0;
}
+static int prepare_sws_context(AVFilterLink *outlink)
+{
+ AVFilterContext *context = outlink->src;
+ DnnProcessingContext *ctx = context->priv;
+ AVFilterLink *inlink = context->inputs[0];
+ enum AVPixelFormat fmt = inlink->format;
+ DNNDataType input_dt = ctx->input.dt;
+ DNNDataType output_dt = ctx->output.dt;
+
+ switch (fmt) {
+ case AV_PIX_FMT_RGB24:
+ case AV_PIX_FMT_BGR24:
+ if (input_dt == DNN_FLOAT) {
+ ctx->sws_gray8_to_grayf32 = sws_getContext(inlink->w * 3,
+ inlink->h,
+ AV_PIX_FMT_GRAY8,
+ inlink->w * 3,
+ inlink->h,
+ AV_PIX_FMT_GRAYF32,
+ 0, NULL, NULL, NULL);
+ }
+ if (output_dt == DNN_FLOAT) {
+ ctx->sws_grayf32_to_gray8 = sws_getContext(outlink->w * 3,
+ outlink->h,
+ AV_PIX_FMT_GRAYF32,
+ outlink->w * 3,
+ outlink->h,
+ AV_PIX_FMT_GRAY8,
+ 0, NULL, NULL, NULL);
+ }
+ return 0;
+ case AV_PIX_FMT_YUV420P:
+ case AV_PIX_FMT_YUV422P:
+ case AV_PIX_FMT_YUV444P:
+ case AV_PIX_FMT_YUV410P:
+ case AV_PIX_FMT_YUV411P:
+ av_assert0(input_dt == DNN_FLOAT);
+ av_assert0(output_dt == DNN_FLOAT);
+ ctx->sws_gray8_to_grayf32 = sws_getContext(inlink->w,
+ inlink->h,
+ AV_PIX_FMT_GRAY8,
+ inlink->w,
+ inlink->h,
+ AV_PIX_FMT_GRAYF32,
+ 0, NULL, NULL, NULL);
+ ctx->sws_grayf32_to_gray8 = sws_getContext(outlink->w,
+ outlink->h,
+ AV_PIX_FMT_GRAYF32,
+ outlink->w,
+ outlink->h,
+ AV_PIX_FMT_GRAY8,
+ 0, NULL, NULL, NULL);
+
+ if (inlink->w != outlink->w || inlink->h != outlink->h) {
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
+ int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
+ int sws_src_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
+ int sws_dst_h = AV_CEIL_RSHIFT(outlink->h, desc->log2_chroma_h);
+ int sws_dst_w = AV_CEIL_RSHIFT(outlink->w, desc->log2_chroma_w);
+ ctx->sws_uv_scale = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
+ sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
+ SWS_BICUBIC, NULL, NULL, NULL);
+ ctx->sws_uv_height = sws_src_h;
+ }
+ return 0;
+ default:
+ //do nothing
+ break;
+ }
+
+ return 0;
+}
+
static int config_output(AVFilterLink *outlink)
{
AVFilterContext *context = outlink->src;
outlink->w = ctx->output.width;
outlink->h = ctx->output.height;
+ prepare_sws_context(outlink);
+
return 0;
}
-static int copy_from_frame_to_dnn(DNNData *dnn_data, const AVFrame *in)
+static int copy_from_frame_to_dnn(DnnProcessingContext *ctx, const AVFrame *frame)
{
- // extend this function to support more formats
- av_assert0(in->format == AV_PIX_FMT_RGB24 || in->format == AV_PIX_FMT_BGR24);
-
- if (dnn_data->dt == DNN_FLOAT) {
- float *dnn_input = dnn_data->data;
- for (int i = 0; i < in->height; i++) {
- for(int j = 0; j < in->width * 3; j++) {
- int k = i * in->linesize[0] + j;
- int t = i * in->width * 3 + j;
- dnn_input[t] = in->data[0][k] / 255.0f;
- }
+ int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
+ DNNData *dnn_input = &ctx->input;
+
+ switch (frame->format) {
+ case AV_PIX_FMT_RGB24:
+ case AV_PIX_FMT_BGR24:
+ if (dnn_input->dt == DNN_FLOAT) {
+ sws_scale(ctx->sws_gray8_to_grayf32, (const uint8_t **)frame->data, frame->linesize,
+ 0, frame->height, (uint8_t * const*)(&dnn_input->data),
+ (const int [4]){frame->linesize[0] * sizeof(float), 0, 0, 0});
+ } else {
+ av_assert0(dnn_input->dt == DNN_UINT8);
+ av_image_copy_plane(dnn_input->data, bytewidth,
+ frame->data[0], frame->linesize[0],
+ bytewidth, frame->height);
}
- } else {
- uint8_t *dnn_input = dnn_data->data;
- av_assert0(dnn_data->dt == DNN_UINT8);
- for (int i = 0; i < in->height; i++) {
- for(int j = 0; j < in->width * 3; j++) {
- int k = i * in->linesize[0] + j;
- int t = i * in->width * 3 + j;
- dnn_input[t] = in->data[0][k];
- }
+ return 0;
+ case AV_PIX_FMT_GRAY8:
+ case AV_PIX_FMT_GRAYF32:
+ av_image_copy_plane(dnn_input->data, bytewidth,
+ frame->data[0], frame->linesize[0],
+ bytewidth, frame->height);
+ return 0;
+ case AV_PIX_FMT_YUV420P:
+ case AV_PIX_FMT_YUV422P:
+ case AV_PIX_FMT_YUV444P:
+ case AV_PIX_FMT_YUV410P:
+ case AV_PIX_FMT_YUV411P:
+ sws_scale(ctx->sws_gray8_to_grayf32, (const uint8_t **)frame->data, frame->linesize,
+ 0, frame->height, (uint8_t * const*)(&dnn_input->data),
+ (const int [4]){frame->width * sizeof(float), 0, 0, 0});
+ return 0;
+ default:
+ return AVERROR(EIO);
+ }
+
+ return 0;
+}
+
+static int copy_from_dnn_to_frame(DnnProcessingContext *ctx, AVFrame *frame)
+{
+ int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
+ DNNData *dnn_output = &ctx->output;
+
+ switch (frame->format) {
+ case AV_PIX_FMT_RGB24:
+ case AV_PIX_FMT_BGR24:
+ if (dnn_output->dt == DNN_FLOAT) {
+ sws_scale(ctx->sws_grayf32_to_gray8, (const uint8_t *[4]){(const uint8_t *)dnn_output->data, 0, 0, 0},
+ (const int[4]){frame->linesize[0] * sizeof(float), 0, 0, 0},
+ 0, frame->height, (uint8_t * const*)frame->data, frame->linesize);
+
+ } else {
+ av_assert0(dnn_output->dt == DNN_UINT8);
+ av_image_copy_plane(frame->data[0], frame->linesize[0],
+ dnn_output->data, bytewidth,
+ bytewidth, frame->height);
}
+ return 0;
+ case AV_PIX_FMT_GRAY8:
+ // it is possible that data type of dnn output is float32,
+ // need to add support for such case when needed.
+ av_assert0(dnn_output->dt == DNN_UINT8);
+ av_image_copy_plane(frame->data[0], frame->linesize[0],
+ dnn_output->data, bytewidth,
+ bytewidth, frame->height);
+ return 0;
+ case AV_PIX_FMT_GRAYF32:
+ av_assert0(dnn_output->dt == DNN_FLOAT);
+ av_image_copy_plane(frame->data[0], frame->linesize[0],
+ dnn_output->data, bytewidth,
+ bytewidth, frame->height);
+ return 0;
+ case AV_PIX_FMT_YUV420P:
+ case AV_PIX_FMT_YUV422P:
+ case AV_PIX_FMT_YUV444P:
+ case AV_PIX_FMT_YUV410P:
+ case AV_PIX_FMT_YUV411P:
+ sws_scale(ctx->sws_grayf32_to_gray8, (const uint8_t *[4]){(const uint8_t *)dnn_output->data, 0, 0, 0},
+ (const int[4]){frame->width * sizeof(float), 0, 0, 0},
+ 0, frame->height, (uint8_t * const*)frame->data, frame->linesize);
+ return 0;
+ default:
+ return AVERROR(EIO);
}
return 0;
}
-static int copy_from_dnn_to_frame(AVFrame *out, const DNNData *dnn_data)
+static av_always_inline int isPlanarYUV(enum AVPixelFormat pix_fmt)
+{
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
+ av_assert0(desc);
+ return !(desc->flags & AV_PIX_FMT_FLAG_RGB) && desc->nb_components == 3;
+}
+
+static int copy_uv_planes(DnnProcessingContext *ctx, AVFrame *out, const AVFrame *in)
{
- // extend this function to support more formats
- av_assert0(out->format == AV_PIX_FMT_RGB24 || out->format == AV_PIX_FMT_BGR24);
-
- if (dnn_data->dt == DNN_FLOAT) {
- float *dnn_output = dnn_data->data;
- for (int i = 0; i < out->height; i++) {
- for(int j = 0; j < out->width * 3; j++) {
- int k = i * out->linesize[0] + j;
- int t = i * out->width * 3 + j;
- out->data[0][k] = av_clip_uintp2((int)(dnn_output[t] * 255.0f), 8);
- }
+ if (!ctx->sws_uv_scale) {
+ av_assert0(in->height == out->height && in->width == out->width);
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(in->format);
+ int uv_height = AV_CEIL_RSHIFT(in->height, desc->log2_chroma_h);
+ for (int i = 1; i < 3; ++i) {
+ int bytewidth = av_image_get_linesize(in->format, in->width, i);
+ av_image_copy_plane(out->data[i], out->linesize[i],
+ in->data[i], in->linesize[i],
+ bytewidth, uv_height);
}
} else {
- uint8_t *dnn_output = dnn_data->data;
- av_assert0(dnn_data->dt == DNN_UINT8);
- for (int i = 0; i < out->height; i++) {
- for(int j = 0; j < out->width * 3; j++) {
- int k = i * out->linesize[0] + j;
- int t = i * out->width * 3 + j;
- out->data[0][k] = dnn_output[t];
- }
- }
+ sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
+ 0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
+ sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 2), in->linesize + 2,
+ 0, ctx->sws_uv_height, out->data + 2, out->linesize + 2);
}
return 0;
DNNReturnType dnn_result;
AVFrame *out;
- copy_from_frame_to_dnn(&ctx->input, in);
+ copy_from_frame_to_dnn(ctx, in);
dnn_result = (ctx->dnn_module->execute_model)(ctx->model, &ctx->output, 1);
if (dnn_result != DNN_SUCCESS){
av_frame_free(&in);
return AVERROR(EIO);
}
- av_assert0(ctx->output.channels == 3);
out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
if (!out) {
}
av_frame_copy_props(out, in);
- copy_from_dnn_to_frame(out, &ctx->output);
+ copy_from_dnn_to_frame(ctx, out);
+
+ if (isPlanarYUV(in->format))
+ copy_uv_planes(ctx, out, in);
+
av_frame_free(&in);
return ff_filter_frame(outlink, out);
}
{
DnnProcessingContext *context = ctx->priv;
+ sws_freeContext(context->sws_gray8_to_grayf32);
+ sws_freeContext(context->sws_grayf32_to_gray8);
+ sws_freeContext(context->sws_uv_scale);
+
if (context->dnn_module)
(context->dnn_module->free_model)(&context->model);
projects / ffmpeg / blobdiff