Skip to content

P
pytensor
提交 755d7cdf authored 作者: Chiheb Trabelsi's avatar Chiheb Trabelsi
浏览文件
操作

nnet.py has been modified in order to respect the flake8 style

nnet.py has been fixed nnet.py do not contain long lines.
上级 06adacd3
显示空白字符变更
内嵌 并排
正在显示 包含 32 行增加 和 35 行删除
theano/sandbox/cuda/nnet.py
浏览文件 @ 755d7cdf
...@@ -578,45 +578,46 @@ class GpuSoftmax(GpuOp): ...@@ -578,45 +578,46 @@ class GpuSoftmax(GpuOp):
""" % locals() """ % locals()
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 float * x',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'const int sx0',
body=[ 'const int sx1',
"extern __shared__ float buf[]", 'float * sm',
'const int sm_s0',
'const int sm_s1'],
body=["extern __shared__ float buf[]",
"float * buf2 = buf + N", "float * 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){",
"buf[tx] = x[blockIDX * sx0 + tx * sx1]", "buf[tx] = x[blockIDX * sx0 + tx * sx1]",
"buf2[tx] = buf[tx]", "buf2[tx] = buf[tx]", "}", "__syncthreads()",
"}", inline_softmax('N',
"__syncthreads()", 'buf',
inline_softmax('N', 'buf', 'buf2', 'buf2',
'threadIdx.x', 'blockDim.x'), 'threadIdx.x',
'blockDim.x'),
"for (int tx = threadIdx.x; tx< N; tx += blockDim.x){", "for (int tx = threadIdx.x; tx< N; tx += blockDim.x){",
# This set all value correctly # This set all value correctly
"sm[blockIDX * sm_s0 + tx * sm_s1] = buf[tx]", "sm[blockIDX * sm_s0 + tx * sm_s1] = buf[tx]", "}",
"}", "__syncthreads()", "}", ])
"__syncthreads()", 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 float * x', 'const int sx0', 'const int sx1',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'float * sm', 'const int sm_s0', 'const int sm_s1'],
body=[ body=["extern __shared__ float buf[]",
"extern __shared__ float 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 float *x_ptr = &x[blockIDX * sx0]",
"float *sm_ptr = &sm[blockIDX * sm_s0]", "float *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',
"__syncthreads()", 'blockDim.x'),
"}", "__syncthreads()", "}", ])
])
return ret1 + "\n" + ret2 return ret1 + "\n" + ret2
gpu_softmax = GpuSoftmax() gpu_softmax = GpuSoftmax()
...@@ -768,25 +769,20 @@ class GpuSoftmaxWithBias(GpuOp): ...@@ -768,25 +769,20 @@ class GpuSoftmaxWithBias(GpuOp):
'const float * x', 'const int sx0', 'const int sx1', 'const float * x', 'const int sx0', 'const int sx1',
'const float * b', 'const int sb0', 'const float * b', 'const int sb0',
'float * sm', 'const int sm_s0', 'const int sm_s1'], 'float * sm', 'const int sm_s0', 'const int sm_s1'],
body=[ body=["extern __shared__ float buf[]",
"extern __shared__ float buf[]",
"float * buf2 = buf + N", "float * 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){",
"buf[tx] = x[blockIDX * sx0 + tx * sx1]", "buf[tx] = x[blockIDX * sx0 + tx * sx1]",
"buf[tx] += b[tx * sb0]", "buf[tx] += b[tx * sb0]",
"buf2[tx] = buf[tx]", "buf2[tx] = buf[tx]", "}",
"}", "__syncthreads()", inline_softmax('N', 'buf', 'buf2',
"__syncthreads()", 'threadIdx.x',
inline_softmax('N', 'buf', 'buf2', 'blockDim.x'),
'threadIdx.x', 'blockDim.x'),
"for (int tx = threadIdx.x; tx< N; tx += blockDim.x){", "for (int tx = threadIdx.x; tx< N; tx += blockDim.x){",
"sm[blockIDX * sm_s0 + tx * sm_s1] = buf[tx]", "sm[blockIDX * sm_s0 + tx * sm_s1] = buf[tx]", "}",
"}", "__syncthreads()", "}", ])
"__syncthreads()",
"}",
])
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 float * x',
...@@ -802,7 +798,8 @@ class GpuSoftmaxWithBias(GpuOp): ...@@ -802,7 +798,8 @@ class GpuSoftmaxWithBias(GpuOp):
"float *sm_ptr = &sm[blockIDX * sm_s0]", "float *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',
'threadIdx.x', 'threadIdx.x',
'blockDim.x', 'blockDim.x',
'b', 'sb0'), 'b', 'sb0'),
......
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