could not get data from gpu when using metal

Question

I wrote a simple code as follows, to check whether GPU could do some computational works.

 id<MTLDevice> device = MTLCreateSystemDefaultDevice();
 NSLog(@"Device: %@", [device name]);

 id<MTLCommandQueue> commandQueue = [device newCommandQueue];

 NSError * ns_error = nil;
 id<MTLLibrary>defaultLibrary = [device newLibraryWithFile:@"/Users/i/tmp/tmp6/s.metallib" error:&ns_error];

 // Buffer for storing encoded commands that are sent to GPU
 id<MTLCommandBuffer> commandBuffer = [commandQueue commandBuffer];

 // Encoder for GPU commands
 id <MTLComputeCommandEncoder> computeCommandEncoder = [commandBuffer computeCommandEncoder];

 //set input and output data
 float tmpbuf[1000];
 float outbuf[1000];
 for( int i = 0; i < 1000; i++ )
 {
     tmpbuf[i] = i;
     outbuf[i] = 0;
 }

 int tmp_length = 100*sizeof(float);
 id<MTLBuffer> inVectorBuffer = [device newBufferWithBytes: tmpbuf length: tmp_length options: MTLResourceOptionCPUCacheModeDefault ];
 [computeCommandEncoder setBuffer: inVectorBuffer offset: 0 atIndex: 0 ];
 id<MTLBuffer> outVectorBuffer = [device newBufferWithBytes: outbuf length: tmp_length options: MTLResourceOptionCPUCacheModeDefault ];
 [computeCommandEncoder setBuffer: outVectorBuffer offset: 0 atIndex: 1 ];


 //get fuction
 id<MTLFunction> newfunc = [ defaultLibrary newFunctionWithName:@"sigmoid" ];

 //get pipelinestat
 id<MTLComputePipelineState> cpipeline = [device newComputePipelineStateWithFunction: newfunc error:&ns_error ];

 [computeCommandEncoder setComputePipelineState:cpipeline ];

 //
 MTLSize ts= {10, 10, 1};
 MTLSize numThreadgroups = {2, 5, 1};
 [computeCommandEncoder dispatchThreadgroups:numThreadgroups threadsPerThreadgroup:ts];
 [ computeCommandEncoder endEncoding ];
 [ commandBuffer commit];

 //get data computed by GPU
 NSData* outdata = [NSData dataWithBytesNoCopy:[outVectorBuffer contents ] length: tmp_length freeWhenDone:false ];
 float final_out[1000];
 [outdata getBytes:final_out length:tmp_length];

 //In my option, each value of final_out should be 0
 for( int i = 0; i < 1000; i++ )
 {
     printf("%.2f : %.2f\n", tmpbuf[i], final_out[i]);
 }   

The shader file, name s.shader, is as follows, which assign output with value 10.0

using namespace metal;
kernel void sigmoid(const device float *inVector [[ buffer(0) ]],
                device float *outVector [[ buffer(1) ]],
                uint id [[ thread_position_in_grid ]]) {
    // This calculates sigmoid for _one_ position (=id) in a vector per call on the GPU
    outVector[id] = 10.0;
}

In the above codes, I got data computed by GPU by variable final_out. In my option, each value of final_out should be 10.0, as presented in s.shader. However, all values of final_out is 0. Any problem in getting back data from GPU? Thanks.

Answer 1

Committing a command buffer simply tells the driver to start executing it. If you want read back the results of a GPU operation on the CPU, you either need to block the current thread with -waitUntilCompleted or add a block to be called when the command buffer completes with the -addCompletedHandler: method.

One other note: it looks like you're using buffers with a storage mode of Shared. If you were ever to use buffers with a storage mode of Managed, you'd also need to create a blit command encoder and call synchronizeResource: with the appropriate buffer(s), then wait for it to complete as described above, in order to copy back the results from the GPU.