Fixed unsupported metal texture creation from a metal buffer on iOS simulator
This commit is contained in:
parent
8f084e6d88
commit
0493a6070d
|
@ -185,9 +185,34 @@ class SubRectExtractorMetal {
|
||||||
id<MTLCommandBuffer> command_buffer,
|
id<MTLCommandBuffer> command_buffer,
|
||||||
id<MTLBuffer> destination) {
|
id<MTLBuffer> destination) {
|
||||||
auto output_texture = MTLTextureWithBuffer(destination_size, destination);
|
auto output_texture = MTLTextureWithBuffer(destination_size, destination);
|
||||||
return InternalExecute(input_texture, sub_rect, flip_horizontally, alpha,
|
absl::Status status =
|
||||||
beta, destination_size, command_buffer,
|
InternalExecute(input_texture, sub_rect, flip_horizontally, alpha, beta,
|
||||||
output_texture);
|
destination_size, command_buffer, output_texture);
|
||||||
|
// On the simulator the `output_texture` cannot share it's underlying storage
|
||||||
|
// with the tensor's CPU Buffer. Hence the contents of the `output_texture`
|
||||||
|
// after sub rect extraction must be copied to the tensor's CPU buffer held by
|
||||||
|
// the `destination` buffer.
|
||||||
|
#if TARGET_IPHONE_SIMULATOR
|
||||||
|
NSUInteger output_bytes_per_row =
|
||||||
|
GetBytesPerRaw(output_format_, destination_size);
|
||||||
|
|
||||||
|
id<MTLBlitCommandEncoder> blitCommandEncoder =
|
||||||
|
command_buffer.blitCommandEncoder;
|
||||||
|
[blitCommandEncoder copyFromTexture:output_texture
|
||||||
|
sourceSlice:0
|
||||||
|
sourceLevel:0
|
||||||
|
sourceOrigin:MTLOriginMake(0, 0, 0)
|
||||||
|
sourceSize:MTLSizeMake(output_texture.width,
|
||||||
|
output_texture.height,
|
||||||
|
output_texture.depth)
|
||||||
|
toBuffer:destination
|
||||||
|
destinationOffset:0
|
||||||
|
destinationBytesPerRow:output_bytes_per_row
|
||||||
|
destinationBytesPerImage:0];
|
||||||
|
[blitCommandEncoder endEncoding];
|
||||||
|
#endif
|
||||||
|
|
||||||
|
return status;
|
||||||
}
|
}
|
||||||
|
|
||||||
private:
|
private:
|
||||||
|
@ -201,11 +226,21 @@ class SubRectExtractorMetal {
|
||||||
texture_desc.usage = MTLTextureUsageRenderTarget;
|
texture_desc.usage = MTLTextureUsageRenderTarget;
|
||||||
|
|
||||||
NSUInteger output_bytes_per_row = GetBytesPerRaw(output_format_, size);
|
NSUInteger output_bytes_per_row = GetBytesPerRaw(output_format_, size);
|
||||||
|
// Creating a no copy `MTLTexture` from an `MTLBuffer` with
|
||||||
|
// MTLStorageModeShared on the simultor results in the following error:
|
||||||
|
// "Linear textures from shared buffers is not supported on this device."
|
||||||
|
// To mitigate this crash, an empty `MTLTexture` is created using the
|
||||||
|
// `MTLDevice`. The solution invovles an extra copy for copying the
|
||||||
|
// underlying buffer of the texture back to the tensor's CPU Buffer.
|
||||||
|
id<MTLTexture> texture;
|
||||||
|
#if TARGET_IPHONE_SIMULATOR
|
||||||
|
texture = [buffer.device newTextureWithDescriptor:texture_desc];
|
||||||
|
#else
|
||||||
|
texture = [buffer newTextureWithDescriptor:texture_desc
|
||||||
|
offset:0
|
||||||
|
bytesPerRow:output_bytes_per_row];
|
||||||
|
#endif
|
||||||
|
|
||||||
id<MTLTexture> texture =
|
|
||||||
[buffer newTextureWithDescriptor:texture_desc
|
|
||||||
offset:0
|
|
||||||
bytesPerRow:output_bytes_per_row];
|
|
||||||
return texture;
|
return texture;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
Loading…
Reference in New Issue
Block a user