提交 56b301a3 authored 作者: Frederic's avatar Frederic

Make GpuSoftmaxWithBias don't crash on GTX285 with some shapes.

上级 d0040637
...@@ -470,7 +470,7 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -470,7 +470,7 @@ class GpuSoftmaxWithBias (GpuOp):
def c_code_cache_version(self): def c_code_cache_version(self):
#return () #return ()
return (6,) + inline_softmax.code_version return (7,) + inline_softmax.code_version
def c_code(self, node, nodename, inp, out, sub): def c_code(self, node, nodename, inp, out, sub):
x, b = inp x, b = inp
...@@ -510,7 +510,7 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -510,7 +510,7 @@ class GpuSoftmaxWithBias (GpuOp):
{ {
int n_blocks = std::min(CudaNdarray_HOST_DIMS(%(x)s)[0],32*1024); int n_blocks = std::min(CudaNdarray_HOST_DIMS(%(x)s)[0],32*1024);
//TODO, detect the maximum number of thread per block. //TODO, detect the maximum number of thread per block.
int n_threads = std::min(CudaNdarray_HOST_DIMS(%(x)s)[1], 1024); int n_threads = std::min(CudaNdarray_HOST_DIMS(%(x)s)[1], 512);
int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] * 2 * sizeof(float); int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] * 2 * sizeof(float);
if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0) if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0)
{ {
......
...@@ -183,7 +183,9 @@ def test_softmax_with_bias(): ...@@ -183,7 +183,9 @@ def test_softmax_with_bias():
def cmp(n, m, catch=False): def cmp(n, m, catch=False):
"""Some old card won't accet the configuration arguments of """Some old card won't accet the configuration arguments of
this implementation.""" this implementation. For those cases set catch=True to skip
those errors.
"""
try: try:
#print "test_softmax",n,m #print "test_softmax",n,m
data = numpy.arange(n * m, dtype='float32').reshape(n, m) data = numpy.arange(n * m, dtype='float32').reshape(n, m)
...@@ -193,18 +195,22 @@ def test_softmax_with_bias(): ...@@ -193,18 +195,22 @@ def test_softmax_with_bias():
except RuntimeError, e: except RuntimeError, e:
if not catch: if not catch:
raise raise
assert (e.args[0] == # Different CUDA driver have different error message
'Cuda error: kSoftmaxWithBias_node_0: invalid configuration argument.\n' assert (e.args[0].startswith(
), e.args[0] 'Cuda error: kSoftmaxWithBias_node_0: invalid configuration argument.\n') or
e.args[0].startswith('Cuda error: kSoftmaxWithBias_node_0: invalid argument.\n'))
cmp(2, 5) cmp(2, 5)
#we need to test n>32*1024 to check that we make the block loop. #we need to test n>32*1024 to check that we make the block loop.
cmp(2 << 15, 5) cmp(2 << 15, 5)
cmp(4074, 400) cmp(4074, 400)
cmp(0, 10) cmp(0, 10)
cmp(4, 1000, True) cmp(784, 784)
cmp(4, 1024, True) cmp(4, 1000)
cmp(4, 2000, True) cmp(4, 1024)
cmp(4, 2024, True) cmp(4, 2000)
cmp(4, 2024)
#GTX285 don't have enought shared mem for this case.
cmp(4, 4074, True) cmp(4, 4074, True)
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论