Skip to content

P
pytensor
提交 b9243185 authored 作者: João Victor Tozatti Risso's avatar João Victor Tozatti Risso
浏览文件
操作

Clean up CTC GPU wrapper

* Add GPU context initialization (cuda_enter) and finalization (cuda_exit) * Add proper synchronization on CUDA stream (cuda_wait) before and after performing operations * Reorder C variables at the start of the function to improve readability * Add ext_cuda initialization to utilize the synchronization and context functions in the GPU Signed-off-by: 's avatarJoão Victor Tozatti Risso <joaovictor.risso@gmail.com>
上级 4e67dc7f
隐藏空白字符变更
内嵌 并排
正在显示 包含 56 行增加 和 49 行删除
theano/gpuarray/c_code/ctc_wrapper.c
浏览文件 @ b9243185
#section init_code
setup_ext_cuda();
#section support_code
typedef struct ctc_context {
......@@ -42,7 +46,7 @@ int ctc_check_result(ctcStatus_t retcode, const char * msg)
const char * ctc_msg = ctcGetStatusString( retcode );
PyErr_Format( PyExc_RuntimeError,
"%s CTC error: %s",
"GpuConnectionistTemporalClassification: %s CTC error: %s",
msg,
ctc_msg );
return 1;
......@@ -112,9 +116,25 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
{
ctc_context_t ctc_object;
ctc_context_t * context = &ctc_object;
size_t gpu_workspace_size;
int ctc_error = 0;
const size_t num_activations = PyGpuArray_DIMS( in_activations )[0];
const size_t minibatch_size = PyGpuArray_DIMS( in_activations )[1];
const size_t alphabet_size = PyGpuArray_DIMS( in_activations )[2];
const size_t cost_size = minibatch_size;
const size_t grad_dims[3] = { num_activations, minibatch_size, alphabet_size };
float * costs = NULL,
* activations = NULL,
* gradients = NULL;
cuda_enter( gpu_context->ctx );
ctc_context_init( context, gpu_context );
float * activations = NULL;
switch (in_activations->ga.typecode)
{
case GA_FLOAT:
......@@ -122,7 +142,8 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
break;
default:
ctc_context_destroy( context );
PyErr_SetString(PyExc_TypeError, "Unsupported type for activations!");
PyErr_SetString( PyExc_TypeError,
"GpuConnectionistTemporalClassification: Unsupported type for activations." );
return 1;
}
......@@ -134,7 +155,7 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
ctc_context_destroy( context );
PyErr_Format( PyExc_MemoryError,
"Could not allocate storage for input lengths" );
"GpuConnectionistTemporalClassification: Could not allocate memory for input lengths." );
return 1;
}
......@@ -147,19 +168,12 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
ctc_context_destroy( context );
PyErr_Format( PyExc_MemoryError,
"Could not allocate storage for labels and their lengths" );
"GpuConnectionistTemporalClassification: Could not allocate memory for labels and their lengths." );
return 1;
}
const size_t minibatch_size = PyGpuArray_DIMS( in_activations )[1];
const size_t alphabet_size = PyGpuArray_DIMS( in_activations )[2];
float * costs = NULL;
const size_t cost_size = minibatch_size;
if (NULL == *out_costs || // symbolic variable has no real backing
PyGpuArray_NDIM( *out_costs ) != 1 ||
PyGpuArray_DIMS( *out_costs )[0] != cost_size)
if ( theano_prep_output( out_costs, 1, &cost_size, in_activations->ga.typecode,
GA_C_ORDER, gpu_context ) != 0 )
{
Py_XDECREF( *out_costs );
......@@ -171,8 +185,10 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
// Destroy previous CTC context before returning exception
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
PyErr_Format( PyExc_MemoryError,
"Could not allocate storage for CTC costs");
"GpuConnectionistTemporalClassification: Could not allocate memory for CTC costs." );
return 1;
}
}
......@@ -181,26 +197,12 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
GpuArray_memset( &((*out_costs)->ga), 0 );
}
switch ( (*out_costs)->ga.typecode )
{
case GA_FLOAT:
costs = (float *) PyGpuArray_DEV_DATA( *out_costs );
break;
default:
ctc_context_destroy( context );
PyErr_SetString(PyExc_TypeError, "Unsupported type for costs!");
return 1;
}
float * gradients = NULL;
costs = (float *) PyGpuArray_DEV_DATA( *out_costs );
if ( NULL != out_gradients ) // if gradient computation is not disabled
{
if ( NULL == *out_gradients ||
PyGpuArray_NDIM( *out_gradients ) != 3 ||
PyGpuArray_DIMS( *out_gradients )[0] != PyGpuArray_DIMS( in_activations )[0] ||
PyGpuArray_DIMS( *out_gradients )[1] != PyGpuArray_DIMS( in_activations )[1] ||
PyGpuArray_DIMS( *out_gradients )[2] != PyGpuArray_DIMS( in_activations )[2] )
if ( theano_prep_output( out_gradients, 3, grad_dims, in_activations->ga.typecode,
GA_C_ORDER, gpu_context ) != 0 )
{
Py_XDECREF( *out_gradients );
......@@ -212,8 +214,10 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
{
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
PyErr_Format( PyExc_MemoryError,
"Could not allocate storage for CTC gradients!" );
"GpuConnectionistTemporalClassification: Could not allocate memory for CTC gradients." );
return 1;
}
}
......@@ -222,21 +226,9 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
GpuArray_memset( &((*out_gradients)->ga), 0 );
}
switch ( (*out_gradients)->ga.typecode )
{
case GA_FLOAT:
gradients = (float *) PyGpuArray_DEV_DATA( *out_gradients );
break;
default:
ctc_context_destroy( context );
PyErr_SetString(PyExc_TypeError, "Unsupported type for gradients!");
return 1;
}
gradients = (float *) PyGpuArray_DEV_DATA( *out_gradients );
}
size_t gpu_workspace_size;
int ctc_error = 0;
ctc_error = ctc_check_result( get_workspace_size( context->label_lengths,
context->input_lengths, alphabet_size, minibatch_size, context->options,
&gpu_workspace_size ),
......@@ -247,6 +239,8 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
// Destroy previous CTC context before returning exception
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
return 1;
}
......@@ -256,11 +250,15 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
{
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
PyErr_Format( PyExc_MemoryError,
"Failed to allocate memory for CTC workspace!" );
"GpuConnectionistTemporalClassification: Failed to allocate memory for CTC workspace." );
return 1;
}
cuda_wait( in_activations->ga.data, GPUARRAY_CUDA_WAIT_READ );
ctc_error = ctc_check_result( compute_ctc_loss( activations, gradients,
context->flat_labels, context->label_lengths, context->input_lengths,
alphabet_size, minibatch_size, costs, *(void **)context->workspace,
......@@ -269,10 +267,18 @@ int APPLY_SPECIFIC(ctc_cost_gpu)(PyGpuArrayObject * in_activations,
if ( ctc_error ) // Exception is set by ctc_check_result, return error here
{
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
return 1;
}
cuda_wait( (*out_costs)->ga.data, GPUARRAY_CUDA_WAIT_WRITE );
if ( out_gradients != NULL )
cuda_wait( (*out_gradients)->ga.data, GPUARRAY_CUDA_WAIT_WRITE );
ctc_context_destroy( context );
cuda_exit( gpu_context->ctx );
return 0;
}
......
theano/gpuarray/ctc.py
浏览文件 @ b9243185
......@@ -77,8 +77,9 @@ class GpuConnectionistTemporalClassification(gof.COp):
return dirs
def c_headers(self):
return ['ctc.h', 'numpy_compat.h', 'gpuarray_helper.h', 'gpuarray/types.h',
'gpuarray_api.h', 'gpuarray/array.h', 'gpuarray/util.h', 'gpuarray/extension.h']
return ['ctc.h', 'numpy_compat.h', 'gpuarray/ext_cuda.h',
'gpuarray_helper.h', 'gpuarray/types.h', 'gpuarray_api.h',
'gpuarray/array.h', 'gpuarray/util.h', 'gpuarray/extension.h']
def get_params(self, node):
return node.inputs[0].type.context
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论