Beispiel #1
0
int
APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
                         PyGpuArrayObject *om,
                         cudnnConvolutionDescriptor_t desc,
                         double alpha, double beta,
                         PyGpuArrayObject **output,
                         PARAMS_TYPE* params) {
  PyGpuContextObject *c = input->context;
  void *alpha_p;
  void *beta_p;
  float af = alpha, bf = beta;
  cudnnStatus_t err = CUDNN_STATUS_SUCCESS;

  if (PyGpuArray_DIMS(input)[1] != PyGpuArray_DIMS(kerns)[1] * params->num_groups) {
    PyErr_SetString(PyExc_ValueError,
		    "images and kernel must have the same stack size");
    return 1;
  }
  if ((PyGpuArray_DIMS(kerns)[0] % params->num_groups) != 0) {
    PyErr_SetString(PyExc_ValueError,
		    "Number of filters must be divisible by number of groups");
    return 1;
  }

  switch (input->ga.typecode) {
  case GA_DOUBLE:
    alpha_p = (void *)α
    beta_p = (void *)β
    break;
  case GA_FLOAT:
  case GA_HALF:
    alpha_p = (void *)⁡
    beta_p = (void *)&bf;
    break;
  default:
    PyErr_SetString(PyExc_TypeError, "Unsupported type in convolution");
    return 1;
  }

  if (params->inplace) {
    Py_XDECREF(*output);
    *output = om;
    Py_INCREF(*output);
  } else {
    if (theano_prep_output(output, PyGpuArray_NDIM(om), PyGpuArray_DIMS(om),
                           om->ga.typecode, GA_C_ORDER, c) != 0)
      return 1;
    if (beta != 0.0 && pygpu_move(*output, om))
      return 1;
  }

  if (PyGpuArray_DIMS(input)[0] == 0 || PyGpuArray_DIMS(kerns)[0] == 0 || PyGpuArray_DIMS(kerns)[1] == 0) {
    int err2 = GpuArray_memset(&(*output)->ga, 0);
    if (err2 != GA_NO_ERROR) {
        PyErr_Format(PyExc_RuntimeError,
                     "GpuDnnConv could not fill the output with zeros: %d", err2);
        return 1;
    }
    return 0;
  }

  if (c_set_tensor_for_conv(input, APPLY_SPECIFIC(input), params->num_groups) == -1)
    return 1;
  if (c_set_filter(kerns, APPLY_SPECIFIC(kerns), params->num_groups) == -1)
    return 1;
  if (c_set_tensor_for_conv(*output, APPLY_SPECIFIC(output), params->num_groups) == -1)
    return 1;
  size_t input_offset = PyGpuArray_STRIDE(input, 0) / params->num_groups;
  size_t kern_offset = PyGpuArray_STRIDE(kerns, 0) * PyGpuArray_DIM(kerns, 0) / params->num_groups;
  size_t output_offset = PyGpuArray_STRIDE(*output, 0) / params->num_groups;

  cudnnConvolutionFwdAlgo_t algo = params->conv_algo;
  #ifdef DEBUG
  char algorithm_name[128];
  #endif

  cuda_enter(c->ctx);

  if (params->choose_algo) {
    if (!params->choose_once) {
      reuse_algo = 1;
      for (unsigned int i = 0; i < PyGpuArray_NDIM(input); i++) {
        reuse_algo = (reuse_algo &&
                      PyGpuArray_DIM(input, i) == prev_img_dims[i]);
        reuse_algo = (reuse_algo &&
                      PyGpuArray_DIM(kerns, i) == prev_kern_dims[i]);
      }
    }

    if (!reuse_algo) {
      size_t free;

      int err2 = gpucontext_property(c->ctx, GA_CTX_PROP_LARGEST_MEMBLOCK, &free);
      if (err2 != GA_NO_ERROR) {
        PyErr_Format(PyExc_RuntimeError, "Error when trying to find the "
                     "memory information on the GPU");
        cuda_exit(c->ctx);
        return 1;
      }

      // Guess 4Mb if the info is not available
      if (free == 0) free = 4 * 1024 * 1024;

      if (params->choose_time) {
        int count;
        cudnnConvolutionFwdAlgoPerf_t choice;
        gpudata *tmpmem;

        tmpmem = gpudata_alloc(c->ctx, free, NULL, 0, NULL);
        if (tmpmem == NULL) {
          PyErr_SetString(PyExc_MemoryError, "Could not allocate working GPU memory");
          return -1;
        }
        // We don't sync the buffer as we don't care about the values.
        err = cudnnFindConvolutionForwardAlgorithmEx(
          params->handle, APPLY_SPECIFIC(input), PyGpuArray_DEV_DATA(input),
          APPLY_SPECIFIC(kerns), PyGpuArray_DEV_DATA(kerns),
          desc, APPLY_SPECIFIC(output), PyGpuArray_DEV_DATA(*output),
          1, &count, &choice, *(void **)tmpmem,
          free);
        gpudata_release(tmpmem);

        if (err != CUDNN_STATUS_SUCCESS) {
          PyErr_Format(PyExc_RuntimeError,
                       "error selecting convolution algo: %s",
                       cudnnGetErrorString(err));
          cuda_exit(c->ctx);
          return 1;
        }
        algo = choice.algo;

        #ifdef DEBUG
        if (count == 0) {
            PyErr_SetString(PyExc_RuntimeError, "No best-timed conv fwd algorithm found");
            return 1;
        } else if (choice.status != CUDNN_STATUS_SUCCESS) {
            PyErr_Format(PyExc_RuntimeError,
                         "error getting best-timed FWD algo: %s",
                         cudnnGetErrorString(choice.status));
            return 1;
        } // Else, count is necessarly 1 for current implementation.
        #endif

      } else {
        err = cudnnGetConvolutionForwardAlgorithm(
          params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(kerns),
          desc, APPLY_SPECIFIC(output),
          CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, free, &algo);
          if (err != CUDNN_STATUS_SUCCESS) {
            PyErr_Format(PyExc_RuntimeError,
                         "error selecting convolution algo: %s",
                         cudnnGetErrorString(err));
            cuda_exit(c->ctx);
            return 1;
          }
      }
      prev_algo = algo;
    } else {
      algo = prev_algo;
    }

    #ifdef DEBUG
    if (0 != theano_enum_to_string_cudnnConvolutionFwdAlgo_t(algo, algorithm_name))
        return 1;
    // NB: This is printed only when algorithm is chosen at runtime.
    if (reuse_algo)
        fprintf(stderr, "(reused %s)\n", algorithm_name);
    else
        fprintf(stderr, "(using %s)\n", algorithm_name);
    #endif

    if (params->choose_once) {
      reuse_algo = 1;
    } else {
      for (unsigned int i = 0; i < PyGpuArray_NDIM(input); i++) {
        prev_img_dims[i] = PyGpuArray_DIM(input, i);
        prev_kern_dims[i] = PyGpuArray_DIM(kerns, i);
      }
    }
  }

  /* Only these algos are supported for 3d conv with cuDNN >= V5.1. */
  if (PyGpuArray_NDIM(input) == 5 &&
      !(algo == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM ||
        algo == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM ||
        algo == CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING))
  {
    #ifdef DEBUG
    if (0 != theano_enum_to_string_cudnnConvolutionFwdAlgo_t(algo, algorithm_name))
        return 1;
    fprintf(stderr, "(%s unsupported for 3D: fallback to CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM)\n", algorithm_name);
    #endif
    algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
  }

  // Algo `small` does not work for a batch size > 2^16, with cuDNN >= V5.1.
  // Issue should be resolved for cuDNN > V6.0.
  if (cudnnGetVersion() < 6100 &&
      algo == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM &&
      PyGpuArray_DIM(input, 0) > 65536)
  {
    #ifdef DEBUG
    fprintf(stderr, "(CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM "
                    "will fail with batch size > 2^16, fallback to CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM)\n");
    #endif
    algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
  }

  // The FFT implementation does not support strides, 1x1 filters or inputs
  // with a spatial dimension larger than 1024. The tiled-FFT implementation
  // does not support strides.
  // If the chosen implementation is FFT or tiled-FFT, validate that it can
  // be used on the current data and default to a safe implementation if it
  // can't.
  // The following code is 2d-specific but it is fine as FFT and tiled-FFT are
  // defined only for 2d filters
  /* NB:
  TODO: These checkings seems outdated for FFT algorithms with cuDNN >= 5.1.
  New conditions apply and may depend on number of dimensions (2D or 3D)
  e.g. for FFT_TILING.
  TODO: More globally, how to handle CUDNN_STATUS_NOT_SUPPORTED with unsupported algorithms?
  */
  if ((algo == CUDNN_CONVOLUTION_FWD_ALGO_FFT ||
       algo == CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING) && PyGpuArray_NDIM(input) == 4) {

    // Extract the properties of the convolution descriptor
    int nd;
    int pad[2];
    int stride[2];
    int dilation[2];
    cudnnConvolutionMode_t mode;
    cudnnDataType_t data_type;
    err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                             dilation, &mode, &data_type);
    if (err != CUDNN_STATUS_SUCCESS) {
      PyErr_Format(PyExc_RuntimeError,
                   "error getting convolution properties: %s",
                   cudnnGetErrorString(err));
      cuda_exit(c->ctx);
      return 1;
    }

    if (algo == CUDNN_CONVOLUTION_FWD_ALGO_FFT) {
      if (stride[0] != 1 || stride[1] != 1 ||
          PyGpuArray_DIM(input, 2) > 1024 || PyGpuArray_DIM(input, 3) > 1024 ||
          (PyGpuArray_DIM(kerns, 2) == 1 && PyGpuArray_DIM(kerns, 3) == 1))
      {
        algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
      }
    } else {
      // algo == CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING
      if (stride[0] != 1 || stride[1] != 1) {
        algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
      }
    }
  }

  {
    size_t worksize;
    gpudata *workspace;
    err = cudnnGetConvolutionForwardWorkspaceSize(params->handle,
                                                  APPLY_SPECIFIC(input),
                                                  APPLY_SPECIFIC(kerns),
                                                  desc,
                                                  APPLY_SPECIFIC(output),
                                                  algo,
                                                  &worksize);

    if (err == CUDNN_STATUS_NOT_SUPPORTED) {
      // Fallback to none algo if not supported

      #ifdef DEBUG
      if (0 != theano_enum_to_string_cudnnConvolutionFwdAlgo_t(algo, algorithm_name))
        return 1;
      fprintf(stderr, "(%s error getting worksize: "
                      "fallback to CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM)\n", algorithm_name);
      #endif

      algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;

      err = cudnnGetConvolutionForwardWorkspaceSize(params->handle,
                                                    APPLY_SPECIFIC(input),
                                                    APPLY_SPECIFIC(kerns),
                                                    desc,
                                                    APPLY_SPECIFIC(output),
                                                    algo,
                                                    &worksize);
    }

    if (err != CUDNN_STATUS_SUCCESS) {
      PyErr_Format(PyExc_RuntimeError,
                   "error getting worksize: %s",
                   cudnnGetErrorString(err));
      cuda_exit(c->ctx);
      return 1;
    }

    /*
     * This is less than ideal since we need to free it after (which
     * introduces a synchronization point. But we don't have a module
     * to place a nice get_work_mem() function in.
     */
    if (worksize != 0) {
      workspace = gpudata_alloc(c->ctx, worksize, NULL, 0, NULL);
      if (workspace == NULL) {
        PyErr_SetString(PyExc_RuntimeError,
                        "Could not allocate working memory");
        cuda_exit(c->ctx);
        return 1;
      }
    }

    cuda_wait(input->ga.data, GPUARRAY_CUDA_WAIT_READ);
    cuda_wait(kerns->ga.data, GPUARRAY_CUDA_WAIT_READ);
    cuda_wait((*output)->ga.data, GPUARRAY_CUDA_WAIT_WRITE);

    for ( int g = 0; g < params->num_groups; g++) {
    err = cudnnConvolutionForward(
      params->handle,
      alpha_p,
      APPLY_SPECIFIC(input), ((char *)PyGpuArray_DEV_DATA(input)) + input_offset * g,
      APPLY_SPECIFIC(kerns), ((char *)PyGpuArray_DEV_DATA(kerns)) + kern_offset * g,
      desc, algo,
      worksize == 0 ? NULL : *(void **)workspace, worksize,
      beta_p,
      APPLY_SPECIFIC(output), ((char *)PyGpuArray_DEV_DATA(*output)) + output_offset * g);
    }

    if (worksize != 0)
      gpudata_release(workspace);

    cuda_record(input->ga.data, GPUARRAY_CUDA_WAIT_READ);
    cuda_record(kerns->ga.data, GPUARRAY_CUDA_WAIT_READ);
    cuda_record((*output)->ga.data, GPUARRAY_CUDA_WAIT_WRITE);
  }
  cuda_exit(c->ctx);

  if (err != CUDNN_STATUS_SUCCESS) {
    PyErr_Format(PyExc_RuntimeError, "error doing operation: %s",
		 cudnnGetErrorString(err));
    return 1;
  }
  return 0;
}
Beispiel #2
0
// Theano op code
// Authors: Arjun Jain, Frederic Bastien, Jan Schluter
// Reference code: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/conv_layer.cu
//   and https://github.com/torch/cunn/blob/master/SpatialConvolutionMM.cu
// Adaptation for 3d
PyGpuArrayObject* corr3dMM(PyGpuArrayObject *const bottom,
                           PyGpuArrayObject *const weight,
                           PyGpuArrayObject *const top,
                           const size_t direction,
                           const size_t dH = 1,
                           const size_t dW = 1,
                           const size_t dD = 1,
                           const size_t dilH = 1,
                           const size_t dilW = 1,
                           const size_t dilD = 1,
                           const size_t padH = 0,
                           const size_t padW = 0,
                           const size_t padD = 0)
{
    if (PyGpuArray_NDIM(bottom) != 5)
    {
        PyErr_SetString(PyExc_ValueError, "GpuCorr3dMM requires bottom of 5D");
        return NULL;
    }
    if (!GpuArray_IS_C_CONTIGUOUS(&bottom->ga))
    {
        PyErr_Format(PyExc_ValueError,
                "GpuCorr3dMM requires bottom to be C-contiguous, "
                "but strides are: %ld %ld %ld %ld %ld\n",
                PyGpuArray_STRIDES(bottom)[0],
                PyGpuArray_STRIDES(bottom)[1],
                PyGpuArray_STRIDES(bottom)[2],
                PyGpuArray_STRIDES(bottom)[3],
                PyGpuArray_STRIDES(bottom)[4]);
        return NULL;
    }

    if (PyGpuArray_NDIM(weight) != 5)
    {
        PyErr_SetString(PyExc_ValueError, "GpuCorr3dMM requires weight of 5D");
        return NULL;
    }
    if (!GpuArray_IS_C_CONTIGUOUS(&weight->ga))
    {
        PyErr_Format(PyExc_ValueError,
                "GpuCorr3dMM requires weight to be C-contiguous, "
                "but strides are: %ld %ld %ld %ld %ld\n",
                PyGpuArray_STRIDES(weight)[0],
                PyGpuArray_STRIDES(weight)[1],
                PyGpuArray_STRIDES(weight)[2],
                PyGpuArray_STRIDES(weight)[3],
                PyGpuArray_STRIDES(weight)[4]);
        return NULL;
    }

    if (PyGpuArray_NDIM(top) != 5)
    {
        PyErr_SetString(PyExc_ValueError, "GpuCorr3dMM requires top of 5D");
        return NULL;
    }
    if (!GpuArray_IS_C_CONTIGUOUS(&top->ga))
    {
        PyErr_Format(PyExc_ValueError,
                "GpuCorr3dMM requires top to be C-contiguous, "
                "but strides are: %ld %ld %ld %ld %ld\n",
                PyGpuArray_STRIDES(top)[0],
                PyGpuArray_STRIDES(top)[1],
                PyGpuArray_STRIDES(top)[2],
                PyGpuArray_STRIDES(top)[3],
                PyGpuArray_STRIDES(top)[4]);
        return NULL;
    }

    // Extract some shape information for later and check shape consistency
    // bottom: (batchSize, nChannels, bottomHeight, bottomWidth, bottomDepth)
    const size_t batchSize = PyGpuArray_DIMS(bottom)[0];
    const size_t nChannels = PyGpuArray_DIMS(bottom)[1];
    const size_t bottomHeight = PyGpuArray_DIMS(bottom)[2];
    const size_t bottomWidth = PyGpuArray_DIMS(bottom)[3];
    const size_t bottomDepth = PyGpuArray_DIMS(bottom)[4];
    // weights: (nFilters, nChannels, rows, columns, slices)
    const size_t nFilters = PyGpuArray_DIMS(weight)[0];
    const size_t kH = PyGpuArray_DIMS(weight)[2];
    const size_t kW = PyGpuArray_DIMS(weight)[3];
    const size_t kD = PyGpuArray_DIMS(weight)[4];
    if (nChannels != PyGpuArray_DIMS(weight)[1]) {
        PyErr_SetString(PyExc_ValueError,
                "GpuCorr3dMM images and kernel must have the same stack size\n");
        return NULL;
    }
    // implicit dilated filter
    const size_t dil_kH = (kH - 1) * dilH + 1;
    const size_t dil_kW = (kW - 1) * dilW + 1;
    const size_t dil_kD = (kD - 1) * dilD + 1;
    // top: (batchSize, nFilters, topHeight, topWidth, topDepth)
    const size_t topHeight = (bottomHeight + 2*padH - dil_kH) / dH + 1;
    const size_t topWidth  = (bottomWidth + 2*padW - dil_kW) / dW + 1;
    const size_t topDepth  = (bottomDepth + 2*padD - dil_kD) / dD + 1;
    if (batchSize != PyGpuArray_DIMS(top)[0] ||
            nFilters != PyGpuArray_DIMS(top)[1] ||
            topHeight != PyGpuArray_DIMS(top)[2] ||
            topWidth != PyGpuArray_DIMS(top)[3] ||
            topDepth != PyGpuArray_DIMS(top)[4]) {
        PyErr_Format(PyExc_ValueError,
                "GpuCorr3dMM shape inconsistency:\n"
                "  bottom shape: %ld %ld %ld %ld %ld\n"
                "  weight shape: %ld %ld %ld %ld %ld\n"
                "  top shape: %ld %ld %ld %ld %ld (expected %ld %ld %ld %ld %ld)\n",
                batchSize, nChannels, bottomHeight, bottomWidth, bottomDepth,
                nFilters, nChannels, kH, kW, kD,
                PyGpuArray_DIMS(top)[0], PyGpuArray_DIMS(top)[1],
                PyGpuArray_DIMS(top)[2], PyGpuArray_DIMS(top)[3], PyGpuArray_DIMS(top)[4],
                batchSize, nFilters, topHeight, topWidth, topDepth);
        return NULL;
    }

    int err = gpublas_setup(bottom->context->ctx);
    if (err != GA_NO_ERROR) {
        PyErr_SetString(PyExc_RuntimeError, "Can't setup blas");
        return NULL;
    }

    // Get the max threads per blocks
    size_t max_threads_dim;
    err = gpucontext_property(bottom->context->ctx, GA_CTX_PROP_MAXLSIZE, &max_threads_dim);
    if (err != GA_NO_ERROR){
        PyErr_Format(PyExc_RuntimeError,
                     "Could not fetch max_threads_dim.");
        return NULL;
    }

    // Create temporary columns
    size_t col_dim[2];
    col_dim[0] = nChannels * kW * kH * kD;
    col_dim[1] = topHeight * topWidth * topDepth;
    PyGpuArrayObject* col = (PyGpuArrayObject*)pygpu_empty(2, col_dim,
                                                           bottom->ga.typecode,
                                                           GA_C_ORDER,
                                                           bottom->context,
                                                           Py_None);
    if (NULL == col)
    {
        PyErr_Format(PyExc_RuntimeError,
                "GpuCorr3dMM failed to allocate working memory of %ld x %ld\n",
                col_dim[0], col_dim[1]);
        return NULL;
    }

    // Define some useful variables
    const size_t bottom_stride = PyGpuArray_STRIDES(bottom)[0] / gpuarray_get_elsize(bottom->ga.typecode);
    const size_t top_stride = PyGpuArray_STRIDES(top)[0] / gpuarray_get_elsize(top->ga.typecode);
    const size_t K_ = col_dim[0];
    const size_t N_ = col_dim[1];
    const size_t M_ = nFilters;
    const DTYPE_INPUT_0 one = 1.0f;
    const DTYPE_INPUT_0 zero = 0.0f;

    PyGpuArrayObject *output;
    if (direction == 0) {  // forward pass
        output = top;
        // valid correlation: im3d2col, then gemm
        // Iterate over batch
        for (size_t n = 0; n < batchSize; n++) {
            // First, im3d2col
            err = im3d2col(max_threads_dim,
                           bottom->ga.data, n * bottom_stride, nChannels, bottomHeight,
                           bottomWidth, bottomDepth, kH, kW, kD, dilH, dilW, dilD,
                           padH, padW, padD, dH, dW, dD, col->ga.data);
            if (err != GA_NO_ERROR) {
                Py_DECREF(col);
                return NULL;
            }
            // Second, gemm
            err = gpublas_sgemm(cb_fortran, cb_no_trans, cb_no_trans,
                                N_, M_, K_, one,
                                col->ga.data, 0, N_,
                                weight->ga.data, 0, K_,
                                zero,
                                top->ga.data, n * top_stride, N_);
            if (err != GA_NO_ERROR) {
                PyErr_Format(PyExc_RuntimeError,
                        "GpuCorr3dMM encountered an error running sgemm.\n");
                Py_DECREF(col);
                return NULL;
            }
        }
    }
    else if (direction == 1) {  // backprop wrt. weights
        output = weight;
        // valid convolution: im3col, then gemm
        // Iterate over batch
        for (size_t n = 0; n < batchSize; n++) {
            // First, im3d2col
            err = im3d2col(max_threads_dim,
                           bottom->ga.data, n * bottom_stride, nChannels, bottomHeight,
                           bottomWidth, bottomDepth, kH, kW, kD, dilH, dilW, dilD,
                           padH, padW, padD, dH, dW, dD, col->ga.data);
            if (err != GA_NO_ERROR) {
                Py_DECREF(col);
                return NULL;
            }
            // Second, gemm
            // Note that we accumulate into weight. We do so by setting beta = 0
            // for the first iteration and beta = 1 for subsequent ones. (This
            // is faster than setting weight to all zeros before the loop.)
            err = gpublas_sgemm(cb_fortran, cb_trans, cb_no_trans,
                                K_, M_, N_, one,
                                col->ga.data, 0, N_,
                                top->ga.data, n * top_stride, N_,
                                (n == 0) ? zero : one,
                                weight->ga.data, 0, K_);
            if (err != GA_NO_ERROR) {
                PyErr_Format(PyExc_RuntimeError,
                        "GpuCorr3dMM encountered an error running sgemm.\n");
                Py_DECREF(col);
                return NULL;
            }
        }
    }
    else if (direction == 2) {  // backprop wrt. inputs
        output = bottom;
        // full convolution: gemm, then col2im3d
        // Iterate over batch
        for (size_t n = 0; n < batchSize; n++) {
            // gemm into columns
            err = gpublas_sgemm(cb_fortran, cb_no_trans, cb_trans,
                                N_, K_, M_, one,
                                top->ga.data, n * top_stride, N_,
                                weight->ga.data, 0, K_,
                                zero,
                                col->ga.data, 0, N_);
            if (err != GA_NO_ERROR) {
                PyErr_Format(PyExc_RuntimeError,
                        "GpuCorr3dMM encountered an error running sgemm.\n");
                Py_DECREF(col);
                return NULL;
            }
            // col2im3d back to the data
            err = col2im3d(max_threads_dim,
                           col->ga.data, nChannels,
                           bottomHeight, bottomWidth, bottomDepth,
                           kH, kW, kD, dilH, dilW, dilD, padH, padW, padD,
                           dH, dW, dD, bottom->ga.data, n * bottom_stride);
            if (err != GA_NO_ERROR) {
                Py_DECREF(col);
                return NULL;
            }
        }
    }
    // Free temporary columns
    Py_DECREF(col);

    // Note that we don't change the refcount of the output matrix here. Output
    // (re)allocation and refcounting is done in BaseGpuCorr3dMM.c_code_helper();
    // in here output is just aliased to one of bottom, weights, or top.
    return output;
}