Restrictions

The concepts of 8K and non-8K in this section are described as follows:

Non-8K: Width x Height is in the range of (10 x 6, 4096 x 4096].
8K: The width or height is in the range of (4096, 8192].

Resolution Restrictions

Input resolution range:
10 x 6 to 8192 x 8192

The APIs for implementing VPC functions have different resolution requirements. For details, see the API descriptions in VPC. When the input image format is YUV440SP or YUV440P, the maximum width of the image is 4096.

Output resolution range:

Model	Resolution Range
Atlas 200/300/500 Inference Product Atlas Training Series Product	10 x 6 to 4096 x 4096

Model

Resolution Range

Atlas 200/300/500 Inference Product

Atlas Training Series Product

10 x 6 to 4096 x 4096

Restrictions on Image Formats, Width and Height Alignment, and Buffers

During image processing by VPC, acldvppMalloc and acldvppFree are called to allocate and free the input and output buffers on the device. The buffer lifetime is managed by the user.

For details about the restrictions on the image formats and width and height alignment, see Table 1 and Table 2. For details about the input image formats supported in the 8K scenario, see Table 3.

To implement VPC functions with related APIs:

For details about the definition of the image format, see acldvppPixelFormat. For details about the concepts such as width stride, height stride, and 8K, see Terminology.
The minimum and maximum values of the width stride and height stride vary with different versions:
- Atlas 200/300/500 Inference Product : The width stride ranges from 32 to 16384 (4096 x 4). For a 4096-pixel ARGB image, one pixel occupies four bytes, and a line of pixels occupies 4096 x 4. The height stride ranges from 6 to 16384.
- Atlas Training Series Product : The width stride ranges from 32 to 16384 (4096 x 4). For a 4096-pixel ARGB image, one pixel occupies four bytes, and a line of pixels occupies 4096 x 4. The height stride ranges from 6 to 16384.

**Table 1** Restrictions on the input image format, width and height alignment, and buffer size
Model	Input Format	Input Width and Height	Input Width Stride, Height Stride, and Buffer Size Atlas 200/300/500 Inference Product Atlas Training Series Product
General format, supported by all models	YUV400 8-bit	No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride x 3/2
	YUV420SP NV12 8-bit	Width: Must be a multiple of 2. Height: Must be a multiple of 2.	Width stride: Round up the width to the nearest multiple of 16. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride x 3/2
	YUV420SP NV21 8-bit
	YUV422SP 8-bit	Width: Must be a multiple of 2. Height: No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride x 2
	YVU422SP 8-bit
	YUV444SP 8-bit	No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride x 3
	YVU444SP 8-bit	No alignment requirement.
	YUV422 Packed YUYV 8-bit	Width: Must be a multiple of 2. Height: No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16 and multiply the result by 2. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride
	YUV422 Packed UYVY 8-bit
	YUV422 Packed YVYU 8-bit
	YUV422 Packed VYUY 8-bit
	YUV444 Packed 8-bit	No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16 and multiply the result by 3. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride
	RGB888
	BGR888
	ARGB8888	No alignment requirement.	Width stride: Round up the width to the nearest multiple of 16 and multiply the result by 4. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride
	ABGR8888
	RGBA8888
	BGRA8888

**Table 2** Restrictions on the output image format, width and height alignment, and buffer size
Version	Output Format	Output Width and Height	Output Width Stride, Height Stride, and Buffer Size Atlas 200/300/500 Inference Product Atlas Training Series Product
General format, supported by all versions	YUV420SP NV12 8-bit	Width: Must be a multiple of 2. Height: Must be a multiple of 2.	Width stride: Round up the width to the nearest multiple of 16. Height stride: Round up the height to the nearest multiple of 2. Buffer size (in bytes) = Width stride x Height stride x 3/2
General format, supported by all versions	YUV420SP NV21 8-bit

In the 8K scenario, the input image formats supported by each model are different, as listed in the following table.

**Table 3** Input image formats supported by each model in the 8K scenario
Model	Supported Input Image Formats
Atlas 200/300/500 Inference Product Atlas Training Series Product	YUV420SP NV12 8-bit, YUV420SP NV21 8-bit

Restrictions on Image Cropping and Pasting

Product Version	Restrictions on Image Cropping and Pasting
Atlas 200/300/500 Inference Product	The crop ROI must not be larger than the input image. The paste ROI must not be larger than the output image. A maximum of 256 paste ROIs are supported. The left offset of the paste ROI relative to the output image is 16-pixel aligned. The paste ROI can be directly aligned on the leftmost side of the output image, that is, with zero left offset relative to the output image. Zero also meets the 16-pixel alignment requirement. It is recommended that the paste ROI width be rounded up to the nearest multiple of 16. Otherwise, invalid extra data will be padded to meet the alignment requirement. As shown in Figure 4 and Figure 5, the green region indicates invalid data. For resizing with the aspect ratio preserved, the left offset for pasting must be a multiple of 16. Therefore, the acldvppVpcCropResizePasteAsync call may result in paste ROI deviation from the center of the output image. In this case, you can add the acldvppVpcBatchCropResizeMakeBorderAsync call to implementing the cropping, resizing, and padding functions together. The restrictions on the odd and even numbers of the crop and paste ROIs are as follows: Only the output image is in YUV420SP format. The left offset and top offset of the paste ROI must be even numbers, and the right offset and bottom offset must be odd numbers. For image resizing, the aspect ratio ranges of the paste and crop ROIs are [1/32, 16]. In inference scenarios, when the VPC output is used as the input of model inference, width x height of the valid paste ROI must be a multiple of 16 x 2. If this requirement is not met, the paste ROI may contain invalid data for padding. In this case, you need to use AIPP to crop the valid ROI and remove invalid data to ensure inference accuracy.
Atlas Training Series Product	The crop ROI must not be larger than the input image. The paste ROI must not be larger than the output image. A maximum of 256 paste ROIs are supported. The left offset of the paste ROI relative to the output image is 16-pixel aligned. The paste ROI can be directly aligned on the leftmost side of the output image, that is, with zero left offset relative to the output image. Zero also meets the 16-pixel alignment requirement. It is recommended that the paste ROI width be rounded up to the nearest multiple of 16. Otherwise, invalid extra data will be padded to meet the alignment requirement. As shown in Figure 4 and Figure 5, the green region indicates invalid data. For resizing with the aspect ratio preserved, the left offset for pasting must be a multiple of 16. Therefore, the acldvppVpcCropResizePasteAsync call may result in paste ROI deviation from the center of the output image. In this case, you can add the acldvppVpcBatchCropResizeMakeBorderAsync call to implementing the cropping, resizing, and padding functions together. The restrictions on the odd and even numbers of the crop and paste ROIs are as follows: Only the output image is in YUV420SP format. The left offset and top offset of the paste ROI must be even numbers, and the right offset and bottom offset must be odd numbers. For image resizing, the aspect ratio ranges of the paste and crop ROIs are [1/32, 16]. In inference scenarios, when the VPC output is used as the input of model inference, width x height of the valid paste ROI must be a multiple of 16 x 2. If this requirement is not met, the paste ROI may contain invalid data for padding. In this case, you need to use AIPP to crop the valid ROI and remove invalid data to ensure inference accuracy.

Product Version

Restrictions on Image Cropping and Pasting

Atlas 200/300/500 Inference Product

The crop ROI must not be larger than the input image.
The paste ROI must not be larger than the output image. A maximum of 256 paste ROIs are supported.
The left offset of the paste ROI relative to the output image is 16-pixel aligned.
- The paste ROI can be directly aligned on the leftmost side of the output image, that is, with zero left offset relative to the output image. Zero also meets the 16-pixel alignment requirement.
- It is recommended that the paste ROI width be rounded up to the nearest multiple of 16. Otherwise, invalid extra data will be padded to meet the alignment requirement. As shown in Figure 4 and Figure 5, the green region indicates invalid data.
- For resizing with the aspect ratio preserved, the left offset for pasting must be a multiple of 16. Therefore, the acldvppVpcCropResizePasteAsync call may result in paste ROI deviation from the center of the output image. In this case, you can add the acldvppVpcBatchCropResizeMakeBorderAsync call to implementing the cropping, resizing, and padding functions together.
The restrictions on the odd and even numbers of the crop and paste ROIs are as follows:
Only the output image is in YUV420SP format. The left offset and top offset of the paste ROI must be even numbers, and the right offset and bottom offset must be odd numbers.
For image resizing, the aspect ratio ranges of the paste and crop ROIs are [1/32, 16].
In inference scenarios, when the VPC output is used as the input of model inference, width x height of the valid paste ROI must be a multiple of 16 x 2. If this requirement is not met, the paste ROI may contain invalid data for padding. In this case, you need to use AIPP to crop the valid ROI and remove invalid data to ensure inference accuracy.

Atlas Training Series Product

The crop ROI must not be larger than the input image.
The paste ROI must not be larger than the output image. A maximum of 256 paste ROIs are supported.
The left offset of the paste ROI relative to the output image is 16-pixel aligned.
- The paste ROI can be directly aligned on the leftmost side of the output image, that is, with zero left offset relative to the output image. Zero also meets the 16-pixel alignment requirement.
- It is recommended that the paste ROI width be rounded up to the nearest multiple of 16. Otherwise, invalid extra data will be padded to meet the alignment requirement. As shown in Figure 4 and Figure 5, the green region indicates invalid data.
- For resizing with the aspect ratio preserved, the left offset for pasting must be a multiple of 16. Therefore, the acldvppVpcCropResizePasteAsync call may result in paste ROI deviation from the center of the output image. In this case, you can add the acldvppVpcBatchCropResizeMakeBorderAsync call to implementing the cropping, resizing, and padding functions together.
The restrictions on the odd and even numbers of the crop and paste ROIs are as follows:
Only the output image is in YUV420SP format. The left offset and top offset of the paste ROI must be even numbers, and the right offset and bottom offset must be odd numbers.
For image resizing, the aspect ratio ranges of the paste and crop ROIs are [1/32, 16].
In inference scenarios, when the VPC output is used as the input of model inference, width x height of the valid paste ROI must be a multiple of 16 x 2. If this requirement is not met, the paste ROI may contain invalid data for padding. In this case, you need to use AIPP to crop the valid ROI and remove invalid data to ensure inference accuracy.

Restrictions on Downsampling of YUV Images

VPC converts the input image format to YUV444 or RGB for internal processing based on the input or output image format. YUV444 or RGB output format has no restriction on the odd and even numbers of the width and height, while the YUV420SP or YUV422SP output format, which performs downsampling processing, has such restrictions due to the data format. Therefore, abnormal data may exist on the output image boarders.

Figure 1 Example of abnormal images

The root cause is that during the computation, when the output image is at an odd start point, the image is correct for the YUV444 format. However, when the YUV420SP format is downsampled, the odd and even lines share the same UV. In this case, the Y of the start line and the UV of the previous line form a new pixel and abnormal data is generated.

Parent topic: VPC