Skip to content

P
pytensor
提交 116f1fa6 authored 作者: Pierre Luc Carrier's avatar Pierre Luc Carrier
浏览文件
操作

Changed float to npy_%(dtype)s in ops GpuSoftmax and GpuSoftmaxWithBias.

上级 cc2cdc3f
隐藏空白字符变更
内嵌 并排
正在显示 包含 27 行增加 和 24 行删除
theano/sandbox/gpuarray/nnet.py
浏览文件 @ 116f1fa6
...@@ -471,6 +471,7 @@ class GpuSoftmax (Op): ...@@ -471,6 +471,7 @@ class GpuSoftmax (Op):
return NVCC_compiler return NVCC_compiler
def c_code(self, node, nodename, inp, out, sub): def c_code(self, node, nodename, inp, out, sub):
dtype = self.dtype
x, = inp x, = inp
z, = out z, = out
fail = sub['fail'] fail = sub['fail']
...@@ -503,7 +504,7 @@ class GpuSoftmax (Op): ...@@ -503,7 +504,7 @@ class GpuSoftmax (Op):
//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], 512); int n_threads = std::min(CudaNdarray_HOST_DIMS(%(x)s)[1], 512);
int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] * int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] *
2 * sizeof(float); 2 * sizeof(npy_%(dtype)s);
if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0) if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0)
{ {
...@@ -533,7 +534,7 @@ class GpuSoftmax (Op): ...@@ -533,7 +534,7 @@ class GpuSoftmax (Op):
<<< <<<
n_blocks, n_blocks,
n_threads, n_threads,
n_threads * sizeof(float) n_threads * sizeof(npy_%(dtype)s)
>>>( >>>(
CudaNdarray_HOST_DIMS(%(x)s)[0], CudaNdarray_HOST_DIMS(%(x)s)[0],
CudaNdarray_HOST_DIMS(%(x)s)[1], CudaNdarray_HOST_DIMS(%(x)s)[1],
...@@ -567,11 +568,11 @@ class GpuSoftmax (Op): ...@@ -567,11 +568,11 @@ class GpuSoftmax (Op):
def c_support_code_apply(self, node, nodename): def c_support_code_apply(self, node, nodename):
ret1 = nvcc_kernel("kSoftmax_%s" % nodename, ret1 = nvcc_kernel("kSoftmax_%s" % nodename,
params=['int M', 'int N', params=['int M', 'int N',
'const float * x', 'const int sx0', 'const int sx1', 'const npy_%(dtype)s * x', 'const int sx0', 'const int sx1',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'npy_%(dtype)s * sm', 'const int sm_s0', 'const int sm_s1'],
body=[ body=[
"extern __shared__ float buf[]", "extern __shared__ npy_%(dtype)s buf[]",
"float * buf2 = buf + N", "npy_%(dtype)s * buf2 = buf + N",
"for (int blockIDX = blockIdx.x; blockIDX < M;" "for (int blockIDX = blockIdx.x; blockIDX < M;"
" blockIDX += gridDim.x){", " blockIDX += gridDim.x){",
"for (int tx = threadIdx.x; tx< N; tx += blockDim.x){", "for (int tx = threadIdx.x; tx< N; tx += blockDim.x){",
...@@ -590,14 +591,14 @@ class GpuSoftmax (Op): ...@@ -590,14 +591,14 @@ class GpuSoftmax (Op):
]) ])
ret2 = nvcc_kernel("kSoftmax_fixed_shared%s" % nodename, ret2 = nvcc_kernel("kSoftmax_fixed_shared%s" % nodename,
params=['int M', 'int N', params=['int M', 'int N',
'const float * x', 'const int sx0', 'const int sx1', 'const npy_%(dtype)s * x', 'const int sx0', 'const int sx1',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'npy_%(dtype)s * sm', 'const int sm_s0', 'const int sm_s1'],
body=[ body=[
"extern __shared__ float buf[]", "extern __shared__ npy_%(dtype)s buf[]",
"for (int blockIDX = blockIdx.x; blockIDX < M;" "for (int blockIDX = blockIdx.x; blockIDX < M;"
" blockIDX += gridDim.x){", " blockIDX += gridDim.x){",
"const float *x_ptr = &x[blockIDX * sx0]", "const npy_%(dtype)s *x_ptr = &x[blockIDX * sx0]",
"float *sm_ptr = &sm[blockIDX * sm_s0]", "npy_%(dtype)s *sm_ptr = &sm[blockIDX * sm_s0]",
inline_softmax_fixed_shared('N', 'buf', 'x_ptr', 'sx1', inline_softmax_fixed_shared('N', 'buf', 'x_ptr', 'sx1',
'sm_ptr', 'sm_s1', 'sm_ptr', 'sm_s1',
'threadIdx.x', 'blockDim.x'), 'threadIdx.x', 'blockDim.x'),
...@@ -642,6 +643,7 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -642,6 +643,7 @@ class GpuSoftmaxWithBias (GpuOp):
return NVCC_compiler return NVCC_compiler
def c_code(self, node, nodename, inp, out, sub): def c_code(self, node, nodename, inp, out, sub):
dtype = self.dtype
x, b = inp x, b = inp
z, = out z, = out
fail = sub['fail'] fail = sub['fail']
...@@ -688,7 +690,7 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -688,7 +690,7 @@ class GpuSoftmaxWithBias (GpuOp):
//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], 512); int n_threads = std::min(CudaNdarray_HOST_DIMS(%(x)s)[1], 512);
int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] * int n_shared_bytes = CudaNdarray_HOST_DIMS(%(x)s)[1] *
2 * sizeof(float); 2 * sizeof(npy_%(dtype)s);
if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0) if (CudaNdarray_HOST_DIMS(%(x)s)[0] > 0)
{ {
if(n_shared_bytes < (32 * 1024 - 500)){ if(n_shared_bytes < (32 * 1024 - 500)){
...@@ -717,7 +719,7 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -717,7 +719,7 @@ class GpuSoftmaxWithBias (GpuOp):
<<< <<<
n_blocks, n_blocks,
n_threads, n_threads,
n_threads * sizeof(float) n_threads * sizeof(npy_%(dtype)s)
>>>( >>>(
CudaNdarray_HOST_DIMS(%(x)s)[0], CudaNdarray_HOST_DIMS(%(x)s)[0],
CudaNdarray_HOST_DIMS(%(x)s)[1], CudaNdarray_HOST_DIMS(%(x)s)[1],
...@@ -750,14 +752,15 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -750,14 +752,15 @@ class GpuSoftmaxWithBias (GpuOp):
""" % locals() """ % locals()
def c_support_code_apply(self, node, nodename): def c_support_code_apply(self, node, nodename):
dtype = self.dtype
ret1 = nvcc_kernel("kSoftmaxWithBias_%s" % nodename, ret1 = nvcc_kernel("kSoftmaxWithBias_%s" % nodename,
params=['int M', 'int N', params=['int M', 'int N',
'const float * x', 'const int sx0', 'const int sx1', 'const npy_%(dtype)s * x', 'const int sx0', 'const int sx1',
'const float * b', 'const int sb0', 'const npy_%(dtype)s * b', 'const int sb0',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'npy_%(dtype)s * sm', 'const int sm_s0', 'const int sm_s1'],
body=[ body=[
"extern __shared__ float buf[]", "extern __shared__ npy_%(dtype)s buf[]",
"float * buf2 = buf + N", "npy_%(dtype)s * buf2 = buf + N",
"for (int blockIDX = blockIdx.x; blockIDX < M;" "for (int blockIDX = blockIdx.x; blockIDX < M;"
" blockIDX += gridDim.x){", " blockIDX += gridDim.x){",
"for (int tx = threadIdx.x; tx< N; tx += blockDim.x){", "for (int tx = threadIdx.x; tx< N; tx += blockDim.x){",
...@@ -776,17 +779,17 @@ class GpuSoftmaxWithBias (GpuOp): ...@@ -776,17 +779,17 @@ class GpuSoftmaxWithBias (GpuOp):
]) ])
ret2 = nvcc_kernel("kSoftmaxWithBias_fixed_shared%s" % nodename, ret2 = nvcc_kernel("kSoftmaxWithBias_fixed_shared%s" % nodename,
params=['int M', 'int N', params=['int M', 'int N',
'const float * x', 'const npy_%(dtype)s * x',
'const int sx0', 'const int sx1', 'const int sx0', 'const int sx1',
'const float * b', 'const int sb0', 'const npy_%(dtype)s * b', 'const int sb0',
'float * sm', 'npy_%(dtype)s * sm',
'const int sm_s0', 'const int sm_s1'], 'const int sm_s0', 'const int sm_s1'],
body=[ body=[
"extern __shared__ float buf[]", "extern __shared__ npy_%(dtype)s buf[]",
"for (int blockIDX = blockIdx.x; blockIDX < M;" "for (int blockIDX = blockIdx.x; blockIDX < M;"
" blockIDX += gridDim.x){", " blockIDX += gridDim.x){",
"const float *x_ptr = &x[blockIDX * sx0]", "const npy_%(dtype)s *x_ptr = &x[blockIDX * sx0]",
"float *sm_ptr = &sm[blockIDX * sm_s0]", "npy_%(dtype)s *sm_ptr = &sm[blockIDX * sm_s0]",
inline_softmax_fixed_shared('N', 'buf', inline_softmax_fixed_shared('N', 'buf',
'x_ptr', 'sx1', 'x_ptr', 'sx1',
'sm_ptr', 'sm_s1', 'sm_ptr', 'sm_s1',
......
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