AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность

Add support for multiple regions of interest (ROI) for batch JPEG decoding (#126)

Просмотр исходного кода
Этот коммит содержится в:
Aryan Salmanpour
2025-03-05 17:06:04 -05:00
коммит произвёл GitHub
родитель 863eafcdb6
Коммит 692d4bb72c
8 изменённых файлов: 43 добавлений и 46 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
src/rocjpeg_decoder.cpp
+12 -12
Просмотреть файл
@@ -216,7 +216,7 @@ RocJpegStatus RocJpegDecoder::DecodeBatched(RocJpegStreamHandle *jpeg_streams, i
jpeg_streams_params[j] = std::move(*jpeg_stream_params);
}
CHECK_ROCJPEG(jpeg_vaapi_decoder_.SubmitDecodeBatched(jpeg_streams_params.data() + i, current_batch_size, decode_params, current_surface_ids.data() + i));
CHECK_ROCJPEG(jpeg_vaapi_decoder_.SubmitDecodeBatched(jpeg_streams_params.data() + i, current_batch_size, &decode_params[i], current_surface_ids.data() + i));
for (int k = 0; k < current_batch_size; k++) {
HipInteropDeviceMem hip_interop_dev_mem = {};
@@ -231,8 +231,8 @@ RocJpegStatus RocJpegDecoder::DecodeBatched(RocJpegStreamHandle *jpeg_streams, i
bool is_roi_valid = false;
uint32_t roi_width;
uint32_t roi_height;
roi_width = decode_params->crop_rectangle.right - decode_params->crop_rectangle.left;
roi_height = decode_params->crop_rectangle.bottom - decode_params->crop_rectangle.top;
roi_width = decode_params[k + i].crop_rectangle.right - decode_params[k + i].crop_rectangle.left;
roi_height = decode_params[k + i].crop_rectangle.bottom - decode_params[k + i].crop_rectangle.top;
if (roi_width > 0 && roi_height > 0 && roi_width <= jpeg_stream_params->picture_parameter_buffer.picture_width && roi_height <= jpeg_stream_params->picture_parameter_buffer.picture_height) {
is_roi_valid = true;
@@ -247,38 +247,38 @@ RocJpegStatus RocJpegDecoder::DecodeBatched(RocJpegStreamHandle *jpeg_streams, i
is_roi_valid = false;
}
switch (decode_params->output_format) {
switch (decode_params[k + i].output_format) {
case ROCJPEG_OUTPUT_NATIVE:
// Copy the native decoded output buffers from interop memory directly to the destination buffers
CHECK_ROCJPEG(GetChromaHeight(hip_interop_dev_mem.surface_format, picture_height, chroma_height));
// Copy Luma (first channel) for any surface format
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, picture_height, 0, &destinations[k + i], decode_params, is_roi_valid));
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, picture_height, 0, &destinations[k + i], &decode_params[k + i], is_roi_valid));
if (hip_interop_dev_mem.surface_format == VA_FOURCC_NV12) {
// Copy the second channel (UV interleaved) for NV12
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 1, &destinations[k + i], decode_params, is_roi_valid));
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 1, &destinations[k + i], &decode_params[k + i], is_roi_valid));
} else if (hip_interop_dev_mem.surface_format == VA_FOURCC_444P ||
hip_interop_dev_mem.surface_format == VA_FOURCC_422V) {
// Copy the second and third channels for YUV444 and YUV440 (i.e., YUV422V)
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 1, &destinations[k + i], decode_params, is_roi_valid));
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 2, &destinations[k + i], decode_params, is_roi_valid));
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 1, &destinations[k + i], &decode_params[k + i], is_roi_valid));
CHECK_ROCJPEG(CopyChannel(hip_interop_dev_mem, chroma_height, 2, &destinations[k + i], &decode_params[k + i], is_roi_valid));
}
break;
case ROCJPEG_OUTPUT_YUV_PLANAR:
CHECK_ROCJPEG(GetChromaHeight(hip_interop_dev_mem.surface_format, picture_height, chroma_height));
CHECK_ROCJPEG(GetPlanarYUVOutputFormat(hip_interop_dev_mem, picture_width,
picture_height, chroma_height, &destinations[k + i], decode_params, is_roi_valid));
picture_height, chroma_height, &destinations[k + i], &decode_params[k + i], is_roi_valid));
break;
case ROCJPEG_OUTPUT_Y:
CHECK_ROCJPEG(GetYOutputFormat(hip_interop_dev_mem, picture_width,
picture_height, &destinations[k + i], decode_params, is_roi_valid));
picture_height, &destinations[k + i], &decode_params[k + i], is_roi_valid));
break;
case ROCJPEG_OUTPUT_RGB:
CHECK_ROCJPEG(ColorConvertToRGB(hip_interop_dev_mem, picture_width,
picture_height, &destinations[k + i], decode_params, is_roi_valid));
picture_height, &destinations[k + i], &decode_params[k + i], is_roi_valid));
break;
case ROCJPEG_OUTPUT_RGB_PLANAR:
CHECK_ROCJPEG(ColorConvertToRGBPlanar(hip_interop_dev_mem, picture_width,
picture_height, &destinations[k + i], decode_params, is_roi_valid));
picture_height, &destinations[k + i], &decode_params[k + i], is_roi_valid));
break;
default:
break;
src/rocjpeg_vaapi_decoder.cpp
+8 -8
Просмотреть файл
@@ -711,11 +711,11 @@ RocJpegStatus RocJpegVappiDecoder::SubmitDecodeBatched(JpegStreamParameters *jpe
return ROCJPEG_STATUS_JPEG_NOT_SUPPORTED;
}
if ((decode_params->output_format == ROCJPEG_OUTPUT_RGB || decode_params->output_format == ROCJPEG_OUTPUT_RGB_PLANAR) && current_vcn_jpeg_spec_.can_convert_to_rgb && jpeg_streams_params[i].chroma_subsampling != CSS_440) {
if (decode_params->output_format == ROCJPEG_OUTPUT_RGB) {
if ((decode_params[i].output_format == ROCJPEG_OUTPUT_RGB || decode_params[i].output_format == ROCJPEG_OUTPUT_RGB_PLANAR) && current_vcn_jpeg_spec_.can_convert_to_rgb && jpeg_streams_params[i].chroma_subsampling != CSS_440) {
if (decode_params[i].output_format == ROCJPEG_OUTPUT_RGB) {
jpeg_stream_key.surface_format = VA_RT_FORMAT_RGB32;
jpeg_stream_key.pixel_format = VA_FOURCC_RGBA;
} else if (decode_params->output_format == ROCJPEG_OUTPUT_RGB_PLANAR) {
} else if (decode_params[i].output_format == ROCJPEG_OUTPUT_RGB_PLANAR) {
jpeg_stream_key.surface_format = VA_RT_FORMAT_RGBP;
jpeg_stream_key.pixel_format = VA_FOURCC_RGBP;
}
@@ -770,8 +770,6 @@ RocJpegStatus RocJpegVappiDecoder::SubmitDecodeBatched(JpegStreamParameters *jpe
uint32_t roi_width;
uint32_t roi_height;
roi_width = decode_params->crop_rectangle.right - decode_params->crop_rectangle.left;
roi_height = decode_params->crop_rectangle.bottom - decode_params->crop_rectangle.top;
// Iterate through all entries of jpeg_stream_groups.
// Check if there is a matching entry in the memory pool.
@@ -805,16 +803,18 @@ RocJpegStatus RocJpegVappiDecoder::SubmitDecodeBatched(JpegStreamParameters *jpe
for (int idx : indices) {
// if the HW JPEG decoder has a built-in ROI-decode capability then fill the requested crop rectangle to the picture parameter buffer
void* picture_parameter_buffer = &jpeg_streams_params[idx].picture_parameter_buffer;
roi_width = decode_params[idx].crop_rectangle.right - decode_params[idx].crop_rectangle.left;
roi_height = decode_params[idx].crop_rectangle.bottom - decode_params[idx].crop_rectangle.top;
if (current_vcn_jpeg_spec_.can_roi_decode && roi_width > 0 && roi_height > 0 &&
roi_width <= jpeg_streams_params[idx].picture_parameter_buffer.picture_width &&
roi_height <= jpeg_streams_params[idx].picture_parameter_buffer.picture_height) {
#if VA_CHECK_VERSION(1, 21, 0)
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.x = decode_params->crop_rectangle.left;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.y = decode_params->crop_rectangle.top;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.x = decode_params[idx].crop_rectangle.left;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.y = decode_params[idx].crop_rectangle.top;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.width = roi_width;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->crop_rectangle.height = roi_height;
#else
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->va_reserved[0] = decode_params->crop_rectangle.top << 16 | decode_params->crop_rectangle.left;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->va_reserved[0] = decode_params[idx].crop_rectangle.top << 16 | decode_params[idx].crop_rectangle.left;
reinterpret_cast<VAPictureParameterBufferJPEGBaseline*>(picture_parameter_buffer)->va_reserved[1] = roi_height << 16 | roi_width;
#endif
}