Skip to content

P
pytensor
提交 b980a8ee authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #6317 from abergeron/gpuarray_07

Work to integrate libgpuarray 0.7 changes.
上级 814dc05a 8b9dc5f1
全部展开 隐藏空白字符变更
内嵌 并排
正在显示 包含 161 行增加 和 100 行删除
.jenkins/gpuarray-branch
浏览文件 @ b980a8ee
v0.6.9
\ No newline at end of file
v0.7.1
\ No newline at end of file
theano/gpuarray/__init__.py
浏览文件 @ b980a8ee
......@@ -46,23 +46,36 @@ def init_dev(dev, name=None, preallocate=None):
global pygpu_activated
if not config.cxx:
raise RuntimeError("The new gpu-backend need a c++ compiler.")
if (pygpu.version.major, pygpu.version.minor, pygpu.version.patch) < (0, 6, 1):
if (pygpu.version.major != 0 or pygpu.version.minor != 7 or
pygpu.version.patch < 0):
raise ValueError(
"Your installed version of pygpu is too old, please upgrade to 0.6.1 or later")
"Your installed version of pygpu(%s) is too old, please upgrade to 0.7.0 or later" %
pygpu.version.fullversion)
# This is for the C headers API, we need to match the exact version.
if pygpu.gpuarray.api_version()[0] != 1:
gpuarray_version_major_supported = 2
gpuarray_version_major_detected = pygpu.gpuarray.api_version()[0]
if gpuarray_version_major_detected != gpuarray_version_major_supported:
raise ValueError(
"Your installed libgpuarray is not in sync, please make sure to have the appropriate version")
"Your installed version oflibgpuarray is not in sync with the current Theano"
" version. The installed libgpuarray version support API version %d,"
" while current Theano support API version %d. Change the version of"
" libgpuarray or Theano to fix this problem.",
gpuarray_version_major_detected,
gpuarray_version_major_supported)
if dev not in init_dev.devmap:
args = dict()
if config.gpuarray.cache_path != '':
os.environ['GPUARRAY_CACHE_PATH'] = config.gpuarray.cache_path
args['kernel_cache_path'] = config.gpuarray.cache_path
if preallocate is None:
preallocate = config.gpuarray.preallocate
if preallocate < 0:
args['max_cache_size'] = 0
else:
args['initial_cache_size'] = preallocate
context = pygpu.init(
dev,
disable_alloc_cache=preallocate < 0,
single_stream=config.gpuarray.single_stream,
sched=config.gpuarray.sched)
sched=config.gpuarray.sched,
**args)
context.dev = dev
init_dev.devmap[dev] = context
reg_context(name, context)
......@@ -115,12 +128,12 @@ def init_dev(dev, name=None, preallocate=None):
# This will map the context name to the real context object.
if config.print_active_device:
try:
pcibusid = '(' + context.pcibusid + ')'
unique_id = '(' + context.unique_id + ')'
except pygpu.gpuarray.UnsupportedException:
pcibusid = ''
unique_id = ''
print("Mapped name %s to device %s: %s %s" %
(name, dev, context.devname, pcibusid),
(name, dev, context.devname, unique_id),
file=sys.stderr)
pygpu_activated = True
......@@ -207,5 +220,5 @@ else:
config.device.startswith('opencl') or
config.device.startswith('cuda') or
config.contexts != ''):
error("pygpu was configured but could not be imported or is too old (version 0.6 or higher required)",
error("pygpu was configured but could not be imported or is too old (version 0.7 or higher required)",
exc_info=True)
theano/gpuarray/basic_ops.py
浏览文件 @ b980a8ee
......@@ -158,7 +158,7 @@ class Kernel(object):
the `params` list consists of C typecodes
It can also have the key `cflags` which is a string of C flag
values like this `"GA_USE_DOUBLE|GA_USE_CLUDA"`.
values like this `"GA_USE_DOUBLE|GA_USE_SMALL"`.
Parameters
----------
......@@ -216,7 +216,7 @@ class Kernel(object):
else:
raise TypeError("can't get a dtype from %s" % (type(t),))
dtypes = [get_dtype(t) for t in types]
flags = dict(cluda=True)
flags = dict()
if any(d == np.float64 for d in dtypes):
flags['have_double'] = True
if any(d.itemsize < 4 for d in dtypes):
......@@ -231,8 +231,6 @@ class Kernel(object):
res = []
if self.flags.get('cflags', '') != '':
res.append(self.flags['cflags'])
if self.flags.get('cluda', False):
res.append('GA_USE_CLUDA')
if self.flags.get('have_double', False):
res.append('GA_USE_DOUBLE')
if self.flags.get('have_small', False):
......@@ -241,15 +239,16 @@ class Kernel(object):
res.append('GA_USE_COMPLEX')
if self.flags.get('have_half', False):
res.append('GA_USE_HALF')
return '|'.join(res)
res = '|'.join(res)
if not res:
return '0'
return res
def _get_py_flags(self):
res = dict(self.flags)
cflags = res.pop('cflags', '')
for fl in cflags.split('|'):
fl = fl.strip()
if fl == 'GA_USE_CLUDA':
res['cluda'] = True
if fl == 'GA_USE_DOUBLE':
res['have_double'] = True
if fl == 'GA_USE_SMALL':
......@@ -555,7 +554,7 @@ class CGpuKernelBase(COp, GpuKernelBase):
kflags = splt2[2].strip()
kcode = def_macros + '\n' + kcode + '\n' + undef_macros
res.append(Kernel(kcode, ktypes, kname,
flags=dict(cluda=True, cflags=kflags)))
flags=dict(cflags=kflags)))
n += 2
self._cached_kernels = res
return res
......@@ -703,39 +702,35 @@ class GpuFromHost(Op):
if (%(name)s_tmp == NULL)
%(fail)s
if (%(out)s != NULL && GpuArray_IS_C_CONTIGUOUS(&%(out)s->ga) &&
theano_size_check(%(out)s, PyArray_NDIM(%(name)s_tmp),
(size_t *)PyArray_DIMS(%(name)s_tmp),
get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)))) {
Py_BEGIN_ALLOW_THREADS
err = GpuArray_write(&%(out)s->ga, PyArray_DATA(%(name)s_tmp),
PyArray_NBYTES(%(name)s_tmp));
Py_END_ALLOW_THREADS
Py_DECREF(%(name)s_tmp);
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError, "Could not write data to gpu");
%(fail)s;
}
} else {
if (%(out)s == NULL || !GpuArray_IS_C_CONTIGUOUS(&%(out)s->ga) ||
!theano_size_check(%(out)s, PyArray_NDIM(%(name)s_tmp),
(size_t *)PyArray_DIMS(%(name)s_tmp),
get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)))) {
Py_XDECREF(%(out)s);
// This method will release the GIL when needed.
%(out)s = pygpu_fromhostdata(PyArray_DATA(%(name)s_tmp),
get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)),
PyArray_NDIM(%(name)s_tmp),
(size_t *)PyArray_DIMS(%(name)s_tmp),
(ssize_t *)PyArray_STRIDES(%(name)s_tmp),
%(ctx)s,
Py_None);
Py_DECREF(%(name)s_tmp);
%(out)s = pygpu_empty(PyArray_NDIM(%(name)s_tmp),
(size_t *)PyArray_DIMS(%(name)s_tmp),
get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)),
GA_C_ORDER, %(ctx)s, Py_None);
if (%(out)s == NULL) {
%(fail)s
Py_DECREF(%(name)s_tmp);
%(fail)s;
}
}
Py_BEGIN_ALLOW_THREADS
err = GpuArray_write(&%(out)s->ga, PyArray_DATA(%(name)s_tmp),
PyArray_NBYTES(%(name)s_tmp));
Py_END_ALLOW_THREADS
Py_DECREF(%(name)s_tmp);
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError, "Could not write data to gpu");
%(fail)s;
}
""" % {'name': name, 'inp': inputs[0], 'ctx': sub['params'],
'out': outputs[0], 'fail': sub['fail']}
def c_code_cache_version(self):
return (9,)
return (10,)
class GpuToGpu(Op):
......@@ -1619,7 +1614,8 @@ class GpuEye(GpuKernelBase, Op):
for i in xrange(3)]
def gpu_kernels(self, node, name):
code = """
code = """#include "cluda.h"
KERNEL void eye(GLOBAL_MEM %(ctype)s *a, ga_size a_off,
ga_size n, ga_size m, ga_ssize k) {
a = (GLOBAL_MEM %(ctype)s *)(((GLOBAL_MEM char *)a) + a_off);
......
theano/gpuarray/c_code/corr3d_gemm.c
浏览文件 @ b980a8ee
#section kernels
#kernel dilated_im3d2col_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size :
// TODO check kernel flags
#include "cluda.h"
// This uses a lot of code from Caffe (http://caffe.berkeleyvision.org/);
// sources are clearly marked. Below we reproduce the original license of
// the Caffe software.
......@@ -87,6 +88,8 @@ KERNEL void dilated_im3d2col_kernel(const ga_size n,
}
#kernel im3d2col_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
KERNEL void im3d2col_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_im,
const ga_size offset_im,
......@@ -139,6 +142,8 @@ KERNEL void im3d2col_kernel(const ga_size n,
// GPU kernel for the case of dilation
#kernel dilated_col2im3d_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size, size :
#include "cluda.h"
KERNEL void dilated_col2im3d_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_col,
const ga_size offset_col,
......@@ -207,6 +212,7 @@ KERNEL void dilated_col2im3d_kernel(const ga_size n,
}
#kernel col2im3d_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size, size :
#include "cluda.h"
KERNEL void col2im3d_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_col,
......
theano/gpuarray/c_code/corr_gemm.c
浏览文件 @ b980a8ee
#section kernels
#kernel dilated_im2col_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// TODO check kernel flags
// This uses a lot of code from Caffe (http://caffe.berkeleyvision.org/);
// sources are clearly marked. Below we reproduce the original license of
......@@ -77,6 +78,7 @@ KERNEL void dilated_im2col_kernel(const ga_size n,
}
#kernel im2col_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
KERNEL void im2col_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_im,
......@@ -122,6 +124,8 @@ KERNEL void im2col_kernel(const ga_size n,
// GPU kernel for the case of dilation
#kernel dilated_col2im_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, size, size, *, size, size :
#include "cluda.h"
KERNEL void dilated_col2im_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_col, const ga_size offset_col,
const ga_size height, const ga_size width, const ga_size channels,
......@@ -172,6 +176,7 @@ KERNEL void dilated_col2im_kernel(const ga_size n,
}
#kernel col2im_kernel : size, *, size, size, size, size, size, size, size, size, size, size, size, size, *, size, size :
#include "cluda.h"
KERNEL void col2im_kernel(const ga_size n,
GLOBAL_MEM const DTYPE_INPUT_0 * data_col, const ga_size offset_col,
......
theano/gpuarray/c_code/dnn_fwd.c
浏览文件 @ b980a8ee
......@@ -199,7 +199,7 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
if (!reuse_algo) {
char pci_id[16];
gpucontext_property(c->ctx, GA_CTX_PROP_PCIBUSID, pci_id);
gpucontext_property(c->ctx, GA_CTX_PROP_UNIQUE_ID, pci_id);
// check out cache
hashkey = dnn_conv_shape(APPLY_SPECIFIC(input), input, APPLY_SPECIFIC(kerns), kerns, desc, *output, groups);
if (hashkey.empty()) {
......
theano/gpuarray/c_code/dnn_gi.c
浏览文件 @ b980a8ee
......@@ -168,7 +168,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
if (!reuse_algo) {
char pci_id[16];
gpucontext_property(c->ctx, GA_CTX_PROP_PCIBUSID, pci_id);
gpucontext_property(c->ctx, GA_CTX_PROP_UNIQUE_ID, pci_id);
// check out cache
hashkey = dnn_conv_shape(APPLY_SPECIFIC(input), *input, APPLY_SPECIFIC(kerns), kerns, desc, output, groups);
if (hashkey.empty()) {
......
theano/gpuarray/c_code/dnn_gw.c
浏览文件 @ b980a8ee
......@@ -155,7 +155,7 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
if (!reuse_algo) {
char pci_id[16];
gpucontext_property(c->ctx, GA_CTX_PROP_PCIBUSID, pci_id);
gpucontext_property(c->ctx, GA_CTX_PROP_UNIQUE_ID, pci_id);
// check out cache
hashkey = dnn_conv_shape(APPLY_SPECIFIC(input), input, APPLY_SPECIFIC(kerns), *kerns, desc, output, groups);
if (hashkey.empty()) {
......
theano/gpuarray/c_code/magma_cholesky.c
浏览文件 @ b980a8ee
#section kernels
#kernel tril_kernel : size, size, size, *:
#include "cluda.h"
KERNEL void tril_kernel(const ga_size nthreads, const ga_size ncols,
const ga_size a_off, GLOBAL_MEM DTYPE_INPUT_0 *a) {
......@@ -17,6 +18,7 @@ KERNEL void tril_kernel(const ga_size nthreads, const ga_size ncols,
}
#kernel triu_kernel : size, size, size, *:
#include "cluda.h"
KERNEL void triu_kernel(const ga_size nthreads, const ga_size ncols,
const ga_size a_off, GLOBAL_MEM DTYPE_INPUT_0 *a) {
......
theano/gpuarray/c_code/magma_qr.c
浏览文件 @ b980a8ee
#section kernels
#kernel triu_kernel : size, size, size, *:
#include "cluda.h"
KERNEL void triu_kernel(const ga_size nthreads, const ga_size ncols,
const ga_size a_off, GLOBAL_MEM DTYPE_INPUT_0 *a) {
......
theano/gpuarray/c_code/pool.c
浏览文件 @ b980a8ee
#section kernels
#kernel max_pool2d_kernel : size, size, size, size, size, size, size, *, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (borrowed from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool2d_kernel(const ga_size nthreads,
......@@ -44,6 +45,7 @@ KERNEL void max_pool2d_kernel(const ga_size nthreads,
}
#kernel max_pool3d_kernel : size, size, size, size, size, size, size, size, size, *, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool3d_kernel(const ga_size nthreads,
......@@ -95,6 +97,7 @@ KERNEL void max_pool3d_kernel(const ga_size nthreads,
}
#kernel ave_pool2d_kernel : size, size, size, size, size, size, size, *, size, size, size, size, size, size, size, bool, bool, *, size:
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void ave_pool2d_kernel(const ga_size nthreads,
......@@ -150,6 +153,7 @@ KERNEL void ave_pool2d_kernel(const ga_size nthreads,
}
#kernel ave_pool3d_kernel : size, size, size, size, size, size, size, size, size, *, size, size, size, size, size, size, size, size, size, size, bool, bool, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void ave_pool3d_kernel(const ga_size nthreads,
......
theano/gpuarray/c_code/pool_ave_grad.c
浏览文件 @ b980a8ee
#section kernels
#kernel ave_pool2d_grad_kernel : size, size, size, size, size, size, size, *, size, *, size, size, size, size, size, size, size, bool, bool, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void ave_pool2d_grad_kernel(const ga_size nthreads,
......@@ -50,6 +51,7 @@ KERNEL void ave_pool2d_grad_kernel(const ga_size nthreads,
}
#kernel ave_pool3d_grad_kernel : size, size, size, size, size, size, size, size, size, *, size, *, size, size, size, size, size, size, size, size, size, size, bool, bool, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void ave_pool3d_grad_kernel(const ga_size nthreads,
......
theano/gpuarray/c_code/pool_grad_grad.c
浏览文件 @ b980a8ee
#section kernels
#kernel max_pool2d_grad_grad_kernel : size, size, size, size, size, size, size, *, size, *, size, *, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
KERNEL void max_pool2d_grad_grad_kernel(const ga_size nthreads,
const ga_size num, const ga_size channels, const ga_size pooled_height,
......@@ -47,6 +48,7 @@ KERNEL void max_pool2d_grad_grad_kernel(const ga_size nthreads,
}
#kernel max_pool3d_grad_grad_kernel : size, size, size, size, size, size, size, size, size, *, size, *, size, *, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
KERNEL void max_pool3d_grad_grad_kernel(const ga_size nthreads,
const ga_size num, const ga_size channels, const ga_size pooled_depth,
......
theano/gpuarray/c_code/pool_max_grad.c
浏览文件 @ b980a8ee
#section kernels
#kernel max_pool2d_grad_kernel : size, size, size, size, size, size, size, *, size, *, size, *, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (borrowed from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool2d_grad_kernel(const ga_size nthreads,
......@@ -43,6 +44,7 @@ KERNEL void max_pool2d_grad_kernel(const ga_size nthreads,
}
#kernel max_pool3d_grad_kernel : size, size, size, size, size, size, size, size, size, *, size, *, size, *, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool3d_grad_kernel(const ga_size nthreads,
......
theano/gpuarray/c_code/pool_max_rop.c
浏览文件 @ b980a8ee
#section kernels
#kernel max_pool2d_rop_kernel : size, size, size, size, size, size, size, *, size, *, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (borrowed from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool2d_rop_kernel(const ga_size nthreads,
......@@ -50,6 +51,7 @@ KERNEL void max_pool2d_rop_kernel(const ga_size nthreads,
}
#kernel max_pool3d_rop_kernel : size, size, size, size, size, size, size, size, size, *, size, *, size, size, size, size, size, size, size, size, size, size, *, size :
#include "cluda.h"
// (adopted from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/pooling_layer.cu)
KERNEL void max_pool3d_rop_kernel(const ga_size nthreads,
......
theano/gpuarray/elemwise.py
浏览文件 @ b980a8ee
......@@ -1743,7 +1743,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_ccontig"
k_var = "kernel_reduce_ccontig_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0,
const %(in_type)s *A, const ga_size offset_A,
......@@ -1781,7 +1782,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_1"
k_var = "kernel_reduce_1_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0,
const %(in_type)s *A, const ga_size offset_A,
......@@ -1821,7 +1823,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_11"
k_var = "kernel_reduce_11_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0, const ga_size d1,
const %(in_type)s *A, const ga_size offset_A,
......@@ -1909,7 +1912,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
load_in + "(A[i3 * sA3 + i2 * sA2 + i1 * sA1 + i0 * sA0])",
{}, True)
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s{
%(init)s
for (int i0 = blockIdx.x; i0 < d0; i0 += gridDim.x){
......@@ -1943,7 +1947,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_010"
k_var = "kernel_reduce_010_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0, const ga_size d1, const ga_size d2,
const %(in_type)s *A, const ga_size offset_A,
......@@ -1989,7 +1994,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_010_AD"
k_var = "kernel_reduce_010_AD_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size A, const ga_size B, const ga_size C, const ga_size D,
const %(in_type)s *X, const ga_size offset_X,
......@@ -2053,7 +2059,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[i0 * sA0 + 0 * sA1 + i2 * sA2])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2088,7 +2095,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_110"
k_var = "kernel_reduce_110_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0, const ga_size d1, const ga_size d2,
const %(in_type)s *A, const ga_size offset_A,
......@@ -2133,7 +2141,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[i1 * sA1 + i2 * sA2])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2163,7 +2172,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[0])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2195,7 +2205,7 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_001"
k_var = "kernel_reduce_001_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0, const ga_size d1, const ga_size d2,
const %(in_type)s *A, const ga_size offset_A,
......@@ -2244,7 +2254,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[i0 * sA0 + i1 * sA1])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2280,7 +2291,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[i0 * sA0 + i2 * sA2])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2314,7 +2326,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
{}, True)
reduce_init = self._assign_init(load_in + "(A[0])")
sio = StringIO()
print("""
print("""#include "cluda.h"
%(decl)s
{
%(init)s
......@@ -2345,7 +2358,8 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
kname = "kernel_reduce_1011"
k_var = "kernel_reduce_1011_" + nodename
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size d0, const ga_size d1, const ga_size d2, const ga_size d3,
const %(in_type)s *A, const ga_size offset_A,
......@@ -2502,15 +2516,15 @@ class GpuCAReduceCPY(GpuKernelBase, HideC, CAReduceDtype):
def gpu_kernels(self, node, name):
if not any(getattr(self, 'redux', [node.inputs[0].ndim != 0])):
# Some OpenCL compilers do not accept no-arguments kernels
src = "KERNEL void reduk(GLOBAL_MEM float *a) {}"
# Some OpenCL compilers do not accept no-arguments empty kernels
src = "#include \"cluda.h\"\nKERNEL void reduk(GLOBAL_MEM float *a) { a[0] = 0; }"
params = ['float32']
else:
k = self.get_kernel_cache(node)
_, src, _, _ = k._get_basic_kernel(k.init_local_size,
node.inputs[0].ndim)
nd = node.inputs[0].ndim
params = ['uint32', gpuarray.GpuArray]
params = ['uint32', gpuarray.GpuArray, 'uint32']
params.extend('uint32' for _ in range(nd))
params.append(gpuarray.GpuArray)
params.append('uint32')
......@@ -2617,9 +2631,10 @@ class GpuCAReduceCPY(GpuKernelBase, HideC, CAReduceDtype):
code += """
args[0] = &n;
args[1] = tmp->ga.data;
args[2] = &tmp->ga.offset;
""" % dict(output=output)
p = 2
p = 3
for i in range(node.inputs[0].ndim):
code += """
proxy_dim[%(i)s] = %(input)s->ga.dimensions[%(i)s];
......@@ -2677,7 +2692,7 @@ class GpuCAReduceCPY(GpuKernelBase, HideC, CAReduceDtype):
return code
def c_code_cache_version_apply(self, node):
return (3, self.kernel_version(node))
return (4, self.kernel_version(node))
def generate_kernel(self, node, odtype, redux):
if isinstance(self.scalar_op, scalar.basic.Add):
......
theano/gpuarray/extra_ops.py
浏览文件 @ b980a8ee
......@@ -74,7 +74,8 @@ class GpuCumOp(GpuKernelBase, Op):
k_var = "k_cumadd_" + nodename
dtype_x = node.inputs[0].dtype
flags = Kernel.get_flags(dtype_x)
code = """
code = """#include "cluda.h"
KERNEL void %(kname)s(float* input, ga_size input_offset,
float* output, ga_size output_offset,
ga_ssize inputStrides_x, ga_ssize inputStrides_y, ga_ssize inputStrides_z,
......@@ -112,7 +113,8 @@ class GpuCumOp(GpuKernelBase, Op):
gpuarray.SSIZE, gpuarray.SSIZE, gpuarray.SSIZE,
gpuarray.SSIZE, gpuarray.SSIZE, gpuarray.SSIZE,
'int32', 'int32', gpuarray.GpuArray, gpuarray.SIZE]
code = """
code = """#include "cluda.h"
// helper functions
WITHIN_KERNEL
void k_reductionPhase(float* partialCumOp) {
......@@ -213,7 +215,8 @@ class GpuCumOp(GpuKernelBase, Op):
# k_finalCumOp
kname = "k_finalCumOp"
k_var = "k_finalCumOp_" + nodename
code = """
code = """#include "cluda.h"
KERNEL void k_finalCumOp(float* output, ga_size output_offset,
float* blockSum, ga_size blockSum_offset,
size_t nbElementsPerCumOp,
......
theano/gpuarray/fp16_help.py
浏览文件 @ b980a8ee
......@@ -22,7 +22,7 @@ def load_w(dtype):
"""
if dtype == 'float16':
return '__half2float'
return 'ga_half2float'
else:
return ''
......@@ -37,6 +37,6 @@ def write_w(dtype):
"""
if dtype == 'float16':
return '__float2half_rn'
return 'ga_float2half'
else:
return ''
theano/gpuarray/kernel_codegen.py
浏览文件 @ b980a8ee
......@@ -34,7 +34,9 @@ def nvcc_kernel(name, params, body):
else:
yield b
bodystr = ';\n'.join(flatbody())
return """KERNEL void %(name)s (%(paramstr)s)
return """#include "cluda.h"
KERNEL void %(name)s (%(paramstr)s)
{
%(bodystr)s;
}
......
theano/gpuarray/multinomial.py
浏览文件 @ b980a8ee
......@@ -66,7 +66,8 @@ class GPUAMultinomialFromUniform(GpuKernelBase, Op):
work_ctype = pygpu.gpuarray.dtype_to_ctype(work_dtype(node.inputs[0].dtype))
write_out_ctype = write_w(node.outputs[0].dtype)
load_in_ctype = load_w(node.inputs[0].dtype)
code = """
code = """#include "cluda.h"
KERNEL void k_multi_warp_multinomial(
const ga_size nb_multi,
const ga_size nb_outcomes,
......@@ -276,7 +277,8 @@ class GPUAChoiceFromUniform(GpuKernelBase, Op):
def gpu_kernels(self, node, name):
replace = int(self.replace)
code = """
code = """#include "cluda.h"
KERNEL void k_multi_warp_multinomial_wor(
const ga_size nb_multi,
const ga_size nb_outcomes,
......
theano/gpuarray/neighbours.py
浏览文件 @ b980a8ee
......@@ -61,7 +61,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
kernels = []
kname = "k_multi_warp_less"
k_var = "k_multi_warp_less_" + nodename
code = """
code = """#include "cluda.h"
// a version that uses less registers but doesn't work in all cases.
%(mode_constants)s
KERNEL void %(kname)s(
......@@ -163,7 +164,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
kname = "k_multi_warp"
k_var = "k_multi_warp_" + nodename
code = """
code = """#include "cluda.h"
%(mode_constants)s
KERNEL void %(kname)s(
const ga_int mode,
......@@ -500,7 +502,7 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
size_t threads_per_block[3] = {d, c, 1};
//get the max threads per blocks
size_t max_threads_dim;
int err = gpucontext_property(%(params)s->context->ctx, GA_CTX_PROP_MAXLSIZE, &max_threads_dim);
int err = gpucontext_property(%(params)s->context->ctx, GA_CTX_PROP_MAXLSIZE0, &max_threads_dim);
if (err != GA_NO_ERROR){
PyErr_SetString(PyExc_RuntimeError, "Could not fetch max_threads_dims");
%(fail)s;
......
theano/gpuarray/nnet.py
浏览文件 @ b980a8ee
......@@ -75,7 +75,8 @@ class GpuCrossentropySoftmaxArgmax1HotWithBias(GpuKernelBase, Op):
gpuarray.GpuArray, gpuarray.SIZE, gpuarray.SSIZE
]
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(const ga_size M, const ga_size N,
GLOBAL_MEM const %(type_x)s* x_data, const ga_size offset_x, const ga_ssize xs0, const ga_ssize xs1,
GLOBAL_MEM const %(type_b)s* b, const ga_size offset_b, const ga_ssize bs0,
......@@ -393,7 +394,8 @@ class GpuCrossentropySoftmax1HotWithBiasDx(GpuKernelBase, Op):
gpuarray.GpuArray, gpuarray.SIZE, gpuarray.SSIZE, gpuarray.SSIZE,
]
sio = StringIO()
print("""
print("""#include "cluda.h"
KERNEL void %(kname)s(
const ga_size N, const ga_size K,
GLOBAL_MEM const %(type_dnll)s* dnll, const ga_size offset_dnll, const ga_ssize dnll_s0,
......@@ -495,7 +497,7 @@ class GpuSoftmax(GpuKernelBase, Op):
{
size_t n_blocks[3] = {std::min(PyGpuArray_DIMS(%(x)s)[0], (size_t)(32 * 1024)), 1, 1};
//TODO, detect the maximum number of thread per block.
size_t threads_per_block[3] = {std::min(PyGpuArray_DIMS(%(x)s)[1], (size_t)256), 1, 1}; // TODO: Read GA_CTX_PROP_MAXLSIZE
size_t threads_per_block[3] = {std::min(PyGpuArray_DIMS(%(x)s)[1], (size_t)256), 1, 1}; // TODO: Read GA_CTX_PROP_MAXLSIZE0
size_t shmem_sz = PyGpuArray_DIMS(%(x)s)[1] *
2 * sizeof(npy_%(work_x)s);
ssize_t stride_X0 = PyGpuArray_STRIDES(%(x)s)[0] / %(itemsize_x)s;
......@@ -557,7 +559,8 @@ class GpuSoftmax(GpuKernelBase, Op):
kernels = []
kname = "kSoftmax"
k_var = "kSoftmax_" + nodename
code = """
code = """#include "cluda.h"
KERNEL void %(kname)s (const ga_size M, const ga_size N,
GLOBAL_MEM const %(type_x)s * x, const ga_size offset_x, const ga_ssize sx0, const ga_ssize sx1,
GLOBAL_MEM %(type_sm)s * sm, const ga_size offset_sm, const ga_ssize sm_s0, const ga_ssize sm_s1 GA_DECL_SHARED_PARAM(%(type_acc)s, buf))
......@@ -630,7 +633,8 @@ class GpuSoftmax(GpuKernelBase, Op):
flags=flags, objvar=k_var))
kname = "kSoftmax_fixed_shared"
k_var = "kSoftmax_fixed_shared" + nodename
code = """
code = """#include "cluda.h"
KERNEL void %(kname)s (const ga_size M, const ga_size N,
GLOBAL_MEM const %(type_x)s * x, const ga_size offset_x, const ga_ssize sx0, const ga_ssize sx1,
GLOBAL_MEM %(type_sm)s * sm, const ga_size offset_sm, const ga_ssize sm_s0, const ga_ssize sm_s1 GA_DECL_SHARED_PARAM(%(type_acc)s, buf))
......@@ -788,7 +792,7 @@ class GpuSoftmaxWithBias(GpuKernelBase, Op):
{
size_t n_blocks[3] = {std::min(PyGpuArray_DIMS(%(x)s)[0], (size_t)(32*1024)), 1, 1};
//TODO, detect the maximum number of thread per block.
size_t threads_per_block[3] = {std::min(PyGpuArray_DIMS(%(x)s)[1], (size_t)256), 1, 1}; // TODO: Read GA_CTX_PROP_MAXLSIZE
size_t threads_per_block[3] = {std::min(PyGpuArray_DIMS(%(x)s)[1], (size_t)256), 1, 1}; // TODO: Read GA_CTX_PROP_MAXLSIZE0
size_t shmem_sz = PyGpuArray_DIMS(%(x)s)[1] *
2 * sizeof(npy_%(work_x)s);
ssize_t stride_X0 = PyGpuArray_STRIDES(%(x)s)[0] / %(itemsize_x)s;
......@@ -854,7 +858,8 @@ class GpuSoftmaxWithBias(GpuKernelBase, Op):
kernels = []
kname = "kSoftmaxWithBias"
k_var = "kSoftmaxWithBias_" + nodename
code = """
code = """#include "cluda.h"
KERNEL void %(kname)s (const ga_size M, const ga_size N,
GLOBAL_MEM const %(type_x)s * x, const ga_size offset_x, const ga_ssize sx0, const ga_ssize sx1,
GLOBAL_MEM const %(type_b)s * b, const ga_size offset_b, const ga_ssize sb0,
......@@ -930,7 +935,8 @@ class GpuSoftmaxWithBias(GpuKernelBase, Op):
flags=flags, objvar=k_var))
kname = "kSoftmaxWithBias_fixed_shared"
k_var = "kSoftmaxWithBias_fixed_shared" + nodename
code = """
code = """#include "cluda.h"
KERNEL void %(kname)s (const ga_size M, const ga_size N,
GLOBAL_MEM const %(type_x)s * x, const ga_size offset_x, const ga_ssize sx0, const ga_ssize sx1,
GLOBAL_MEM const %(type_b)s * b, const ga_size offset_b, const ga_ssize sb0,
......
theano/gpuarray/opt.py
浏览文件 @ b980a8ee
......@@ -1110,17 +1110,11 @@ def local_gpua_advanced_boolean_subtensor(op, context_name, inputs, outputs):
@op_lifter([tensor.AdvancedIncSubtensor1])
@register_opt2([tensor.AdvancedIncSubtensor1], 'fast_compile')
def local_gpua_advanced_incsubtensor1(op, context_name, inputs, outputs):
context = get_context(context_name)
# This is disabled on non-cuda contexts
if context.kind != b'cuda':
return None
x, y, ilist = inputs
set_instead_of_inc = op.set_instead_of_inc
compute_capability = int(context.bin_id[-2])
if (compute_capability >= 2 and x.ndim == 1 and y.ndim == 0 and
if (x.ndim == 1 and y.ndim == 0 and
config.deterministic == 'default'):
x = x.dimshuffle(0, 'x')
y = y.dimshuffle('x', 'x')
......@@ -1128,7 +1122,7 @@ def local_gpua_advanced_incsubtensor1(op, context_name, inputs, outputs):
set_instead_of_inc=set_instead_of_inc)(x, y, ilist)
ret = GpuDimShuffle(ret.type.broadcastable, [0])(ret)
return ret
elif (compute_capability < 2 or x.ndim != 2 or y.ndim != 2 or
elif (x.ndim != 2 or y.ndim != 2 or
config.deterministic == 'more'):
return GpuAdvancedIncSubtensor1(
set_instead_of_inc=set_instead_of_inc)
......
theano/gpuarray/rng_mrg.py
浏览文件 @ b980a8ee
......@@ -80,7 +80,8 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base):
else:
raise ValueError('Unsupported data type for output',
self.output_type.dtype)
code = """
code = """#include "cluda.h"
KERNEL void mrg_uniform(
GLOBAL_MEM %(otype)s *sample_data,
ga_size sample_offset,
......
theano/gpuarray/tests/c_code/tstgpueye.c
浏览文件 @ b980a8ee
#section kernels
#kernel eye : *, size, size, size :
#include <cluda.h>
/* The eye name will be used to generate supporting objects. The only
you probably need to care about is the kernel object which will be
named 'k_' + <the name above> (k_eye in this case). This name also
......
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