Skip to content

P
pytensor
提交 f39ba80a authored 作者: Pascal Lamblin's avatar Pascal Lamblin 提交者： GitHub
浏览文件
操作

Merge pull request #6161 from tfjgeorge/ger_inplace_param

uses inplace as op params for GpuGer, GpuGemm, GpuGemmBatch and GpuGemv
上级 2bbd2ed2 1b5d0fd0
隐藏空白字符变更
内嵌 并排
正在显示 包含 68 行增加 和 89 行删除
theano/gpuarray/blas.py
浏览文件 @ f39ba80a
......@@ -6,7 +6,8 @@ import theano
from theano import Apply, config, Op
from theano.compile import optdb
from theano.gof import LocalOptGroup
from theano.gof import LocalOptGroup, ParamsType
from theano.scalar import bool as bool_t
from theano.tensor.basic import as_tensor_variable
from theano.tensor.opt import in2out
......@@ -38,6 +39,7 @@ class GpuGemv(BlasOp):
Gemv on the GPU.
"""
params_type = ParamsType(inplace=bool_t)
__props__ = ('inplace',)
def __init__(self, inplace=False):
......@@ -68,9 +70,9 @@ class GpuGemv(BlasOp):
beta = beta.astype(expected)
return Apply(self, [y, alpha, A, x, beta], [y.type()])
def perform(self, node, inputs, out_storage):
def perform(self, node, inputs, out_storage, params):
y, alpha, A, x, beta = inputs
inplace = self.inplace
inplace = params.inplace
if inplace and y.strides[0] < 0:
inplace = False
if A.shape[1] == 0:
......@@ -82,27 +84,20 @@ class GpuGemv(BlasOp):
def c_code(self, node, name, inp, out, sub):
vars = dict(out=out[0], y=inp[0], alpha=inp[1], A=inp[2], x=inp[3],
beta=inp[4], fail=sub['fail'], name=name)
if self.inplace:
code = """
if (%(y)s->ga.strides[0] <= 0) {
%(out)s = theano_try_copy(%(out)s, %(y)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(y)s;
Py_INCREF(%(out)s);
}
""" % vars
else:
code = """
%(out)s = theano_try_copy(%(out)s, %(y)s);
if (%(out)s == NULL) {
%(fail)s
}
""" % vars
beta=inp[4], fail=sub['fail'], name=name,
params=sub['params'])
code = """
if (!%(params)s->inplace || %(y)s->ga.strides[0] <= 0) {
%(out)s = theano_try_copy(%(out)s, %(y)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(y)s;
Py_INCREF(%(out)s);
}
""" % vars
# in case of possible speed up using blas dot,
# temporary hack A to 1D for vector-vector dot
code += """
......@@ -156,6 +151,7 @@ class GpuGemm(BlasOp):
Gemm on the GPU.
"""
params_type = ParamsType(inplace=bool_t)
__props__ = ('inplace',)
_f16_ok = True
......@@ -195,9 +191,9 @@ class GpuGemm(BlasOp):
assert C.ndim == 2
return Apply(self, [C, alpha, A, B, beta], [C.type()])
def perform(self, node, inputs, outputs):
def perform(self, node, inputs, outputs, params):
C, alpha, A, B, beta = inputs
inplace = self.inplace
inplace = params.inplace
if inplace and not C.flags.forc:
inplace = False
outputs[0][0] = blas.gemm(alpha, A, B, beta, C,
......@@ -205,35 +201,26 @@ class GpuGemm(BlasOp):
def c_code(self, node, name, inp, out, sub):
vars = dict(out=out[0], C=inp[0], alpha=inp[1], A=inp[2], B=inp[3],
beta=inp[4], fail=sub['fail'], name=name)
if self.inplace:
code = """
if (!GpuArray_ISONESEGMENT(&%(C)s->ga)) {
%(out)s = theano_try_copy(%(out)s, %(C)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(C)s;
Py_INCREF(%(out)s);
}
""" % vars
else:
code = """
%(out)s = theano_try_copy(%(out)s, %(C)s);
if (%(out)s == NULL) {
%(fail)s
}
""" % vars
code += """
if (pygpu_blas_rgemm(cb_no_trans, cb_no_trans,
((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(A)s, %(B)s,
((dtype_%(beta)s *)PyArray_DATA(%(beta)s))[0],
%(out)s, 0) == -1) {
%(fail)s
}
beta=inp[4], fail=sub['fail'], name=name,
params=sub['params'])
code = """
if (!%(params)s->inplace || !GpuArray_ISONESEGMENT(&%(C)s->ga)) {
%(out)s = theano_try_copy(%(out)s, %(C)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(C)s;
Py_INCREF(%(out)s);
}
if (pygpu_blas_rgemm(cb_no_trans, cb_no_trans,
((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(A)s, %(B)s,
((dtype_%(beta)s *)PyArray_DATA(%(beta)s))[0],
%(out)s, 0) == -1) {
%(fail)s
}
""" % vars
if config.gpuarray.sync:
code += """
......@@ -242,7 +229,7 @@ class GpuGemm(BlasOp):
return code
def c_code_cache_version(self):
return (5,)
return (6,)
gpugemm_no_inplace = GpuGemm(inplace=False)
gpugemm_inplace = GpuGemm(inplace=True)
......@@ -253,6 +240,7 @@ class GpuGer(BlasOp):
Ger on the GPU.
"""
params_type = ParamsType(inplace=bool_t)
__props__ = ('inplace',)
def __init__(self, inplace=False):
......@@ -278,9 +266,9 @@ class GpuGer(BlasOp):
assert y.ndim == 1
return Apply(self, [A, alpha, x, y], [A.type()])
def perform(self, node, inp, out):
def perform(self, node, inp, out, params):
A, alpha, x, y = inp
inplace = self.inplace
inplace = params.inplace
if inplace and not A.flags.forc:
inplace = False
out[0][0] = blas.ger(alpha, x, y, A,
......@@ -288,33 +276,23 @@ class GpuGer(BlasOp):
def c_code(self, node, name, inp, out, sub):
vars = dict(out=out[0], A=inp[0], alpha=inp[1], x=inp[2], y=inp[3],
fail=sub['fail'], name=name)
if self.inplace:
code = """
if (!GpuArray_ISONESEGMENT(&%(A)s->ga)) {
%(out)s = theano_try_copy(%(out)s, %(A)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(A)s;
Py_INCREF(%(out)s);
}
""" % vars
else:
code = """
%(out)s = theano_try_copy(%(out)s, %(A)s);
if (%(out)s == NULL) {
%(fail)s
}
""" % vars
code += """
if (pygpu_blas_rger(((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(x)s, %(y)s, %(out)s, 0) == -1) {
%(fail)s
}
""" % vars
fail=sub['fail'], name=name, params=sub['params'])
code = """
if (!%(params)s->inplace || !GpuArray_ISONESEGMENT(&%(A)s->ga)) {
%(out)s = theano_try_copy(%(out)s, %(A)s);
if (%(out)s == NULL) {
%(fail)s
}
} else {
Py_XDECREF(%(out)s);
%(out)s = %(A)s;
Py_INCREF(%(out)s);
}
if (pygpu_blas_rger(((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(x)s, %(y)s, %(out)s, 0) == -1) {
%(fail)s
}
""" % vars
if config.gpuarray.sync:
code += """
GpuArray_sync(&%(out)s->ga);
......@@ -322,7 +300,7 @@ class GpuGer(BlasOp):
return code
def c_code_cache_version(self):
return (3,)
return (4,)
gpuger_no_inplace = GpuGer(inplace=False)
......@@ -396,6 +374,7 @@ gpu_dot22 = GpuDot22()
class GpuGemmBatch(BlasOp):
params_type = ParamsType(inplace=bool_t)
__props__ = ('inplace',)
_f16_ok = True
......@@ -432,11 +411,11 @@ class GpuGemmBatch(BlasOp):
def c_code(self, node, name, inp, out, sub):
vars = dict(out=out[0], C=inp[0], alpha=inp[1], A=inp[2], B=inp[3],
beta=inp[4], inplace=int(self.inplace),
beta=inp[4], params=sub['params'],
fail=sub['fail'], name=name)
code = """
int err;
if (%(inplace)s){
if (%(params)s->inplace){
if (!GpuArray_ISONESEGMENT(&%(C)s->ga)) {
%(out)s = theano_try_copy(%(out)s, %(C)s);
if (%(out)s == NULL) {
......@@ -472,7 +451,7 @@ class GpuGemmBatch(BlasOp):
return code
def c_code_cache_version(self):
return (2,)
return (3,)
gpugemmbatch_no_inplace = GpuGemmBatch(inplace=False)
gpugemmbatch_inplace = GpuGemmBatch(inplace=True)
......
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