Skip to content

P
pytensor
提交 c5d48a56 authored 作者: Pascal Lamblin's avatar Pascal Lamblin
浏览文件
操作

Work around inability to incref const * PyObjects

When linking with libpython 2.4 (at least), we cannot Py_INCREF / Py_XDECREF PyObject pointers that have been declared constant.
上级 35d3ea25
隐藏空白字符变更
内嵌 并排
正在显示 包含 42 行增加 和 56 行删除
theano/sandbox/cuda/cuda_ndarray.cu
浏览文件 @ c5d48a56
...@@ -2932,6 +2932,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -2932,6 +2932,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
// If matrix A or B has non-unit size and non-unit stride in both // If matrix A or B has non-unit size and non-unit stride in both
// dimensions, we can make a copy. // dimensions, we can make a copy.
CudaNdarray * A_new = NULL;
CudaNdarray * B_new = NULL;
if (((CudaNdarray_HOST_DIMS(A)[0] > 1) if (((CudaNdarray_HOST_DIMS(A)[0] > 1)
&& (CudaNdarray_HOST_STRIDES(A)[0] != 1) && (CudaNdarray_HOST_STRIDES(A)[0] != 1)
&& (CudaNdarray_HOST_DIMS(A)[1] > 1) && (CudaNdarray_HOST_DIMS(A)[1] > 1)
...@@ -2939,18 +2941,11 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -2939,18 +2941,11 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
|| (CudaNdarray_HOST_STRIDES(A)[0] < 0) || (CudaNdarray_HOST_STRIDES(A)[0] < 0)
|| (CudaNdarray_HOST_STRIDES(A)[1] < 0)) || (CudaNdarray_HOST_STRIDES(A)[1] < 0))
{ {
const CudaNdarray* A_new = (CudaNdarray*) CudaNdarray_Copy(A); A_new = (CudaNdarray*) CudaNdarray_Copy(A);
if (!A_new) if (!A_new)
return -1; return -1;
A = A_new; A = A_new;
} }
else
{
// In the case above, we will need to decref A_new at the end.
// To make things simpler, we incref A here, so we can always
// decref A.
Py_INCREF(A);
}
if (((CudaNdarray_HOST_DIMS(B)[0] > 1) if (((CudaNdarray_HOST_DIMS(B)[0] > 1)
&& (CudaNdarray_HOST_STRIDES(B)[0] != 1) && (CudaNdarray_HOST_STRIDES(B)[0] != 1)
...@@ -2959,21 +2954,15 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -2959,21 +2954,15 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
|| (CudaNdarray_HOST_STRIDES(B)[0] < 0) || (CudaNdarray_HOST_STRIDES(B)[0] < 0)
|| (CudaNdarray_HOST_STRIDES(B)[1] < 0)) || (CudaNdarray_HOST_STRIDES(B)[1] < 0))
{ {
const CudaNdarray* B_new = (CudaNdarray*) CudaNdarray_Copy(B); B_new = (CudaNdarray*) CudaNdarray_Copy(B);
if (!B_new) if (!B_new)
{ {
Py_XDECREF(A); // If A_new is NULL, meaning A was not copied nothing happens
Py_XDECREF(A_new);
return -1; return -1;
} }
B = B_new; B = B_new;
} }
else
{
// In the case above, we will need to decref B_new at the end.
// To make things simpler, we incref B here, so we can always
// decref B.
Py_INCREF(B);
}
// If matrix C has non-unit size and non-unit stride in both // If matrix C has non-unit size and non-unit stride in both
// dimensions, or negative strides, we can't operate. We cannot copy // dimensions, or negative strides, we can't operate. We cannot copy
...@@ -2992,8 +2981,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -2992,8 +2981,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
CudaNdarray_HOST_STRIDES(C)[1], CudaNdarray_HOST_STRIDES(C)[1],
CudaNdarray_HOST_DIMS(C)[0], CudaNdarray_HOST_DIMS(C)[0],
CudaNdarray_HOST_DIMS(C)[1]); CudaNdarray_HOST_DIMS(C)[1]);
Py_XDECREF(A); Py_XDECREF(A_new);
Py_XDECREF(B); Py_XDECREF(B_new);
return -1; return -1;
} }
...@@ -3057,8 +3046,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3057,8 +3046,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
cublasSgemm(T0, T1, D0, D1, D2, a, x, sx, y, sy, b, z, sz); \ cublasSgemm(T0, T1, D0, D1, D2, a, x, sx, y, sy, b, z, sz); \
} else { \ } else { \
PyErr_SetString(PyExc_AssertionError, "negative stride to sGemm");\ PyErr_SetString(PyExc_AssertionError, "negative stride to sGemm");\
Py_XDECREF(A);\ Py_XDECREF(A_new);\
Py_XDECREF(B);\ Py_XDECREF(B_new);\
return -1; \ return -1; \
} }
...@@ -3076,8 +3065,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3076,8 +3065,8 @@ int CudaNdarray_gemm(float alpha, const CudaNdarray * A, const CudaNdarray * B,
return -1; return -1;
}; };
CNDA_THREAD_SYNC; CNDA_THREAD_SYNC;
Py_XDECREF(A); Py_XDECREF(A_new);
Py_XDECREF(B); Py_XDECREF(B_new);
cublasStatus err = cublasGetError(); cublasStatus err = cublasGetError();
if (CUBLAS_STATUS_SUCCESS != err) if (CUBLAS_STATUS_SUCCESS != err)
...@@ -3117,6 +3106,8 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3117,6 +3106,8 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B,
// If matrix A has non-unit size and non-unit stride in both // If matrix A has non-unit size and non-unit stride in both
// dimensions, or negative strides, we cannot operate, but we can // dimensions, or negative strides, we cannot operate, but we can
// make a copy. // make a copy.
CudaNdarray * A_new = NULL;
CudaNdarray * B_new = NULL;
if (((CudaNdarray_HOST_DIMS(A)[0] > 1) if (((CudaNdarray_HOST_DIMS(A)[0] > 1)
&& (CudaNdarray_HOST_STRIDES(A)[0] != 1) && (CudaNdarray_HOST_STRIDES(A)[0] != 1)
&& (CudaNdarray_HOST_DIMS(A)[1] > 1) && (CudaNdarray_HOST_DIMS(A)[1] > 1)
...@@ -3124,33 +3115,25 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3124,33 +3115,25 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B,
|| (CudaNdarray_HOST_STRIDES(A)[0] < 0) || (CudaNdarray_HOST_STRIDES(A)[0] < 0)
|| (CudaNdarray_HOST_STRIDES(A)[1] < 0)) || (CudaNdarray_HOST_STRIDES(A)[1] < 0))
{ {
const CudaNdarray* A_new = (CudaNdarray*) CudaNdarray_Copy(A); A_new = (CudaNdarray*) CudaNdarray_Copy(A);
if (!A_new) if (!A_new)
return -1; return -1;
A = A_new; A = A_new;
} }
else
{
// Incref A, so we can decref it at the end in all cases
Py_INCREF(A);
}
// If vector B as a negative stride, we also have to make a copy. // If vector B as a negative stride, we also have to make a copy.
if (CudaNdarray_HOST_STRIDES(B)[0] < 0) if (CudaNdarray_HOST_STRIDES(B)[0] < 0)
{ {
const CudaNdarray* B_new = (CudaNdarray*) CudaNdarray_Copy(B); B_new = (CudaNdarray*) CudaNdarray_Copy(B);
if (!B_new) if (!B_new)
{ {
Py_XDECREF(A); // If A was not copied, A_new is NULL, and Py_XDECREF does not
// do anything
Py_XDECREF(A_new);
return -1; return -1;
} }
B = B_new; B = B_new;
} }
else
{
// Incref B, so we can decref it at the end in all cases
Py_INCREF(B);
}
// cudablas does not handle negative strides as expected // cudablas does not handle negative strides as expected
if ( (CudaNdarray_HOST_STRIDES(A)[0] < 0) if ( (CudaNdarray_HOST_STRIDES(A)[0] < 0)
...@@ -3159,8 +3142,8 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3159,8 +3142,8 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B,
PyErr_Format(PyExc_ValueError, "illegal strides in args to gemv (%i,%i)", PyErr_Format(PyExc_ValueError, "illegal strides in args to gemv (%i,%i)",
CudaNdarray_HOST_STRIDES(A)[0], CudaNdarray_HOST_STRIDES(A)[0],
CudaNdarray_HOST_STRIDES(A)[1]); CudaNdarray_HOST_STRIDES(A)[1]);
Py_XDECREF(A); Py_XDECREF(A_new);
Py_XDECREF(B); Py_XDECREF(B_new);
return -1; return -1;
} }
...@@ -3215,15 +3198,15 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B, ...@@ -3215,15 +3198,15 @@ int CudaNdarray_sgemv(float alpha, const CudaNdarray * A, const CudaNdarray * B,
CudaNdarray_HOST_DIMS(A)[1], CudaNdarray_HOST_DIMS(A)[1],
CudaNdarray_HOST_DIMS(B)[0], CudaNdarray_HOST_DIMS(B)[0],
CudaNdarray_HOST_DIMS(C)[0]); CudaNdarray_HOST_DIMS(C)[0]);
Py_XDECREF(A); Py_XDECREF(A_new);
Py_XDECREF(B); Py_XDECREF(B_new);
return -1; return -1;
} }
} }
CNDA_THREAD_SYNC; CNDA_THREAD_SYNC;
Py_XDECREF(A); Py_XDECREF(A_new);
Py_XDECREF(B); Py_XDECREF(B_new);
cublasStatus err = cublasGetError(); cublasStatus err = cublasGetError();
if (CUBLAS_STATUS_SUCCESS != err) if (CUBLAS_STATUS_SUCCESS != err)
...@@ -3253,7 +3236,7 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y, ...@@ -3253,7 +3236,7 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y,
} }
int x_strides = CudaNdarray_HOST_STRIDES(x)[0]; int x_strides = CudaNdarray_HOST_STRIDES(x)[0];
const CudaNdarray * x_ = x; CudaNdarray * x_new = NULL;
if(x_strides == 0){ if(x_strides == 0){
if(CudaNdarray_HOST_DIMS(x)[0] != 1){ if(CudaNdarray_HOST_DIMS(x)[0] != 1){
PyErr_Format(PyExc_RuntimeError, PyErr_Format(PyExc_RuntimeError,
...@@ -3264,24 +3247,27 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y, ...@@ -3264,24 +3247,27 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y,
} }
x_strides = 1; x_strides = 1;
} else if(x_strides < 0){ } else if(x_strides < 0){
x_ = (CudaNdarray*)CudaNdarray_Copy(x); x_new = (CudaNdarray*) CudaNdarray_Copy(x);
x_strides = CudaNdarray_HOST_STRIDES(x_)[0]; x = x_new;
x_strides = CudaNdarray_HOST_STRIDES(x)[0];
} }
int y_strides = CudaNdarray_HOST_STRIDES(y)[0]; int y_strides = CudaNdarray_HOST_STRIDES(y)[0];
const CudaNdarray * y_ = y; CudaNdarray * y_new = NULL;
if(y_strides == 0){ if(y_strides == 0){
if(CudaNdarray_HOST_DIMS(y)[0] != 1){ if(CudaNdarray_HOST_DIMS(y)[0] != 1){
PyErr_Format(PyExc_RuntimeError, PyErr_Format(PyExc_RuntimeError,
"CudaNdarray_sger: Invalid input y (should not happen)." "CudaNdarray_sger: Invalid input y (should not happen)."
" We received a CudaNdarray vector with a stride of 0" " We received a CudaNdarray vector with a stride of 0"
" that has more than 1 elements!"); " that has more than 1 elements!");
Py_XDECREF(x_new);
return -1; return -1;
} }
y_strides = 1; y_strides = 1;
} else if(y_strides < 0){ } else if(y_strides < 0){
y_ = (CudaNdarray*)CudaNdarray_Copy(y); y_new = (CudaNdarray*) CudaNdarray_Copy(y);
y_strides = CudaNdarray_HOST_STRIDES(y_)[0]; y = y_new;
y_strides = CudaNdarray_HOST_STRIDES(y)[0];
} }
// Create appropriate strides if A is a row or column vector // Create appropriate strides if A is a row or column vector
...@@ -3297,8 +3283,8 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y, ...@@ -3297,8 +3283,8 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y,
&& (CudaNdarray_HOST_STRIDES(A)[1] > 0))) && (CudaNdarray_HOST_STRIDES(A)[1] > 0)))
{ {
cublasSger(CudaNdarray_HOST_DIMS(x)[0], CudaNdarray_HOST_DIMS(y)[0], alpha, cublasSger(CudaNdarray_HOST_DIMS(x)[0], CudaNdarray_HOST_DIMS(y)[0], alpha,
CudaNdarray_DEV_DATA(x_), x_strides, CudaNdarray_DEV_DATA(x), x_strides,
CudaNdarray_DEV_DATA(y_), y_strides, CudaNdarray_DEV_DATA(y), y_strides,
CudaNdarray_DEV_DATA(A), sa_1); CudaNdarray_DEV_DATA(A), sa_1);
} }
// Since Sger expects A in col-major, we invert x and y to fake this. // Since Sger expects A in col-major, we invert x and y to fake this.
...@@ -3307,8 +3293,8 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y, ...@@ -3307,8 +3293,8 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y,
&& (CudaNdarray_HOST_STRIDES(A)[0] > 0))) && (CudaNdarray_HOST_STRIDES(A)[0] > 0)))
{ {
cublasSger(CudaNdarray_HOST_DIMS(y)[0], CudaNdarray_HOST_DIMS(x)[0], alpha, cublasSger(CudaNdarray_HOST_DIMS(y)[0], CudaNdarray_HOST_DIMS(x)[0], alpha,
CudaNdarray_DEV_DATA(y_), y_strides, CudaNdarray_DEV_DATA(y), y_strides,
CudaNdarray_DEV_DATA(x_), x_strides, CudaNdarray_DEV_DATA(x), x_strides,
CudaNdarray_DEV_DATA(A), sa_0); CudaNdarray_DEV_DATA(A), sa_0);
} }
// A has to be either c- or f-contiguous, with no negative strides // A has to be either c- or f-contiguous, with no negative strides
...@@ -3316,14 +3302,14 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y, ...@@ -3316,14 +3302,14 @@ int CudaNdarray_sger(float alpha, const CudaNdarray * x, const CudaNdarray * y,
{ {
PyErr_SetString(PyExc_NotImplementedError, PyErr_SetString(PyExc_NotImplementedError,
"non-contiguous A, or negative strides, in sger"); "non-contiguous A, or negative strides, in sger");
Py_XDECREF(x_new);
Py_XDECREF(y_new);
return -1; return -1;
} }
} }
CNDA_THREAD_SYNC; CNDA_THREAD_SYNC;
if(x_ != x) Py_XDECREF(x_new);
Py_DECREF(x_); Py_XDECREF(y_new);
if(y_ != y)
Py_DECREF(y_);
cublasStatus err = cublasGetError(); cublasStatus err = cublasGetError();
if (CUBLAS_STATUS_SUCCESS != err) if (CUBLAS_STATUS_SUCCESS != err)
......
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