Here's my scenario:
In an MTKView
(and a single command buffer therein), I have a MTLTexture
that is the result of a MTLComputeEncoder
operation.
My draw
method on the MTKView
is called as a high priority instance of a dispatch_get_global_queue
. At the tail end of drawRect
in that MTKView
, I "scoop" that texture and set it as a property on a different MTKView
. I then dispatch a low priority thread to call that view's draw
method to copy a region of the original texture into the second MTKView
, along with an eventual presentDrawable
call -- all on a different command buffer. Meanwhile, the original MTKView
ends the draw
method with its own call to presentDrawable
.
I should mention that both these MTKViews
are configured with isPaused=YES
and enableSetNeedsDisplay=NO
.
This is all working reasonably well, but I occasionally get a crash with either a "[CAMetalLayerDrawable texture] should not be called after presenting the drawable
" warning or the occasional crash in some unidentifiable thread:
QuartzCore`-[CAMetalDrawable present]:
0x7fff881cf471 <+0>: pushq %rbp
0x7fff881cf472 <+1>: movq %rsp, %rbp
0x7fff881cf475 <+4>: movq %rdi, %rax
0x7fff881cf478 <+7>: movq 0x15f25ce1(%rip), %rcx ; CAMetalDrawable._priv
0x7fff881cf47f <+14>: movq (%rax,%rcx), %rcx
0x7fff881cf483 <+18>: movq 0x20(%rcx), %rdi <---- crashes here
0x7fff881cf487 <+22>: xorps %xmm0, %xmm0
0x7fff881cf48a <+25>: movq %rax, %rsi
0x7fff881cf48d <+28>: popq %rbp
0x7fff881cf48e <+29>: jmp 0x7fff881cf493 ; layer_private_present(_CAMetalLayerPrivate*, CAMetalDrawable*, double, unsigned int)
If I remove/comment out the ancillary "scoop" and redraw of the original texture, there is no crash.
Is there a more acceptable way to accomplish this? Possibly using an MTLBlitCommandEncoder
to explicitly copy the contents of the supplied texture?
A little more clarity with the code is a wise idea, possibly:
This is the drawRect
method of the primary MTKView
, which is driven by a dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_HIGH, 0)
:
- (void)drawRect:(NSRect)dirtyRect {
dispatch_semaphore_wait(inflightSemaphore, DISPATCH_TIME_FOREVER);
[super drawRect:dirtyRect];
…
id<MTLTexture> theTexture = …
[renderedQueue addObject:theTexture];
// populate secondary views
for (BoardMonitorMTKView *v in self.downstreamOutputs)
[v enqueueTexture:(hasValidContent) ? [renderedQueue firstObject] : nil];
__block dispatch_semaphore_t fSemaphore = inflightSemaphore;
__weak __block NSMutableArray *rQueue = renderedQueue;
[commandBuffer addCompletedHandler:^(id<MTLCommandBuffer> buffer) {
if ([rQueue count]==2)
[rQueue removeObjectAtIndex:0];
dispatch_semaphore_signal(fSemaphore);
}];
[commandBuffer presentDrawable:self.currentDrawable];
[commandBuffer commit];
}
In the secondary view:
-(void)enqueueTexture:(id<MTLTexture>)inputTexture {
dispatch_semaphore_wait(inflightSemaphore2, DISPATCH_TIME_FOREVER);
if (inputTexture) {
[textureQueue addObject:inputTexture];
if ([textureQueue count]>kTextureQueueCapacity)
[textureQueue removeObject:[textureQueue firstObject]];
}
dispatch_semaphore_signal(inflightSemaphore2);
dispatch_queue_t aQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_LOW, 0);
dispatch_async(aQueue, ^{
[self draw];
});
}
- (void)drawRect:(NSRect)dirtyRect {
dispatch_semaphore_wait(inflightSemaphore2, DISPATCH_TIME_FOREVER);
[super drawRect:dirtyRect];
…
id<MTLTexture>inputTexture = [textureQueue firstObject];
MTLRenderPassDescriptor *renderPassDescriptor = self.currentRenderPassDescriptor;
id<MTLRenderCommandEncoder> encoder = [commandBuffer renderCommandEncoderWithDescriptor:renderPassDescriptor];
MTLViewport viewPort = {0.0, 0.0, (double)self.drawableSize.width, (double)self.drawableSize.height, -1.0, 1.0};
[encoder setViewport:viewPort];
[encoder setRenderPipelineState:metalVertexPipelineState];
// draw main content
NSUInteger vSize = _vertexInfo.metalVertexCount*sizeof(AAPLVertex);
id<MTLBuffer> mBuff = [self.device newBufferWithBytes:_vertexInfo.metalVertices
length:vSize
options:MTLResourceStorageModeShared];
[encoder setVertexBuffer:mBuff offset:0 atIndex:0];
[encoder setVertexBytes:&_viewportSize length:sizeof(_viewportSize) atIndex:1];
[encoder setFragmentTexture:inputTexture atIndex:0];
[encoder drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:_vertexInfo.metalVertexCount];
[encoder endEncoding];
__block dispatch_semaphore_t fSemaphore = inflightSemaphore2;
[commandBuffer addCompletedHandler:^(id<MTLCommandBuffer> buffer) {
[textureQueue removeObject:[textureQueue firstObject]];
dispatch_semaphore_signal(fSemaphore);
}];
[commandBuffer presentDrawable:self.currentDrawable];
[commandBuffer commit];
}
I'm not sure you're aware that it's both acceptable and common to create your own texture and render to that. You don't have to render to the texture of a drawable of a view or layer.
Create your own texture(s) to render to and then, for just the present step, render from your texture to a drawable's texture.
Mind you, depending on exactly what you're doing, you may want a pool of three or so textures that you rotate through. The issue you need to be concerned with is whether Metal is still reading from a texture, so you don't write to it before it's done being read.