Question

我想在循环中携带多个内核调用。我尝试过的代码如下：

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

   id<MTLCommandQueue> commandQueue = [device newCommandQueue];

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

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

   //set input and output data
   float tmpbuf[2][1000];
   float outbuf[2][1000];
   float final_out[2][1000];
   for( int i = 0; i < 1000; i++ )
   {
      tmpbuf[0][i] = i;
      outbuf[0][i] = 0;
      tmpbuf[1][i] = 10*i;
      outbuf[1][i] = 0;
    }

    int tmp_length = 1000*sizeof(float);
    //get pipelinestat
    id<MTLComputePipelineState> cpipeline[2];
    cpipeline[0]    = [device newComputePipelineStateWithFunction: newfunc error:&ns_error ];
    cpipeline[1]    = [device newComputePipelineStateWithFunction: newfunc error:&ns_error ];
    id<MTLBuffer> inVectorBuffer[2];
    id<MTLBuffer> outVectorBuffer[2];
    id <MTLComputeCommandEncoder> computeCommandEncoder[2];
    computeCommandEncoder[0] = [commandBuffer computeCommandEncoder];
    computeCommandEncoder[1] = [commandBuffer computeCommandEncoder];

    MTLSize ts= {10, 1, 1}; 
    MTLSize numThreadgroups = {100, 1, 1};

    for( int k = 0; k < 2; k++ )
    {
         inVectorBuffer[k]  = [device newBufferWithBytes: tmpbuf[k] length: tmp_length options: MTLResourceOptionCPUCacheModeDefault ];
         [computeCommandEncoder[k] setBuffer: inVectorBuffer[k] offset: 0 atIndex: 0 ];
         outVectorBuffer[k] = [device newBufferWithBytes: outbuf[k] length: tmp_length options: MTLResourceOptionCPUCacheModeDefault ];
         [computeCommandEncoder[k] setBuffer: outVectorBuffer[k] offset: 0 atIndex: 1 ];


         [computeCommandEncoder[k] setComputePipelineState:cpipeline[k] ];
         [computeCommandEncoder[k] dispatchThreadgroups:numThreadgroups threadsPerThreadgroup:ts];
         [computeCommandEncoder[k] endEncoding ];

    }

    [ commandBuffer commit];
    [ commandBuffer waitUntilCompleted];enter code here

它无法正常工作。当它运行时，重新编译以下信息

/Library/Caches/com.apple.xbs/Sources/GPUDriversIntel/GPUDriversIntel-10.14.58/Common/MTLRenderer/Intel/ivb/MTLIGAccelCommandBuffer.m:240: failed assertion `Already have uncommitted encoder'
 Abort trap: 6

有人能指出这个问题。提前谢谢。

Answer 1

一次只能激活一个计算命令编码器（每个命令缓冲区）。不要在循环外创建多个编码器，而是考虑在循环中创建命令encoder 循环，并在下一次循环迭代之前在其上结束编码。

for循环中的多个内核调用

1 个答案: