def get_native_compile_info(self, input_types, devtype):
assert devtype == "gpu"
d = dict(cdtype=core.np2c[input_types[0].dtype])
cuda_code=r"""
#include "cgt_cuda.h"
#include "lrn.cuh"
void launchker_$function(int num_img, int channels, int height, int width, int localsize, double alpha, double beta, %(cdtype)s* Xdata,
%(cdtype)s* topdata, %(cdtype)s* scalingdata, %(cdtype)s* topdiffdata, %(cdtype)s* bottomdiffdata) {
int nblocks, nthreads;
int size = num_img * width * height;
cgt_get_bt(size, nblocks, nthreads);
LRNComputeDiff<%(cdtype)s><<<nblocks, nthreads, 0>>>(size, (%(cdtype)s*)Xdata, (%(cdtype)s*)topdata,
(%(cdtype)s*)scalingdata, (%(cdtype)s*)topdiffdata, num_img, channels, height, width, localsize,
-beta, 2. * alpha * beta / localsize, (%(cdtype)s*)bottomdiffdata);
CUDA_CHECK_ERROR("CrossChannelLRNBackward");
}
"""%d
code = """
void launchker_$function(int num_img, int channels, int height, int width, int localsize, double alpha, double beta, %(cdtype)s* Xdata,
%(cdtype)s* topdata, %(cdtype)s* scaledata, %(cdtype)s* topdiffdata, %(cdtype)s* bottomdiffdata);
CGT_EXPORT_C void $function($closure* cldata, cgtArray** reads, cgtArray* bottom_diff) {
cgtArray *X=reads[0], *top=reads[1], *scaling=reads[2], *top_diff=reads[3];
int num_img = X->shape()[0],
channels = X->shape()[1],
height = X->shape()[2],
width = X->shape()[3];
launchker_$function(num_img, channels, height, width, cldata->localsize, cldata->alpha, cldata->beta, (%(cdtype)s*)X->data(),
(%(cdtype)s*)top->data(), (%(cdtype)s*)scaling->data(), (%(cdtype)s*)top_diff->data(), (%(cdtype)s*)bottom_diff->data());
}"""%d
return core.NativeCompileInfo(code, closure_triples = make_closure(self.info),
link_flags="-lcudart", gpu_deref_mask=(True,True,True,True),
extra_srcs=[core.SrcFile("cuda",cuda_code)])
def get_native_compile_info(self, input_types, devtype):
assert devtype == "gpu"
d = dict(cdtype=core.np2c[input_types[0].dtype])
cuda_code = r"""
#include "cgt_cuda.h"
#include "lrn.cuh"
void launchker_$function(int num_img, int channels, int height, int width, int localsize, double alpha, double beta, %(cdtype)s* Xdata, %(cdtype)s* topdata, %(cdtype)s* scaledata) {
int size = num_img * height * width;
int nblocks, nthreads;
cgt_get_bt(size, nblocks, nthreads);
LRNFillScale<%(cdtype)s><<<nblocks, nthreads, 0>>>(
size, Xdata, num_img, channels, height, width, localsize, alpha / localsize, scaledata);
CUDA_CHECK_ERROR("LRNFillScale");
size = num_img * channels * width * height;
cgt_get_bt(size, nblocks, nthreads);
LRNComputeOutput<%(cdtype)s><<<nblocks, nthreads, 0>>>(size, Xdata, scaledata, -beta, topdata);
CUDA_CHECK_ERROR("LRNComputeOutput");
}"""%d
code = r"""
extern void launchker_$function(int num_img, int channels, int height, int width, int localsize, double alpha, double beta, %(cdtype)s* Xdata, %(cdtype)s* topdata, %(cdtype)s* scaledata);
CGT_EXPORT_C void $function($closure* cldata, cgtArray** reads, cgtTuple* write) {
cgtArray* X = reads[0];
int num_img = X->shape()[0],
channels = X->shape()[1],
height = X->shape()[2],
width = X->shape()[3];
cgtArray* top = (cgtArray*)write->getitem(0);
cgtArray* scale = (cgtArray*)write->getitem(1);
launchker_$function(num_img, channels, height, width, cldata->localsize, cldata->alpha, cldata->beta, (%(cdtype)s*)X->data(), (%(cdtype)s*)top->data(), (%(cdtype)s*)scale->data());
}"""%d
return core.NativeCompileInfo(code, closure_triples = make_closure(self.info),
link_flags="-lcudart", gpu_deref_mask=(True,),
extra_srcs=[core.SrcFile("cuda",cuda_code)])
def get_native_compile_info(self, input_types, devtype):
code = r"""
CGT_EXPORT_C void $function(conv_closure* cl, cgtArray** reads, cgtTuple* write) {
max_pool<%(cdtype)s>(cl, reads[0], static_cast<cgtArray*>(write->getitem(0)), static_cast<cgtArray*>(write->getitem(1)));
}""" % dict(cdtype=core.np2c[input_types[0].dtype])
return core.NativeCompileInfo(code,
closure_triples=info2closure(self.info),
includes=["pooling.h"])
def get_native_compile_info(self, input_types, devtype):
code = r"""
CGT_EXPORT_C void $function(conv_closure* cl, cgtArray** reads, cgtArray* write) {
max_pool_pullback<%(cdtype)s>(reads[0], reads[1], reads[2], reads[3], write);
}""" % dict(cdtype=core.np2c[input_types[0].dtype])
return core.NativeCompileInfo(code,
closure_triples=info2closure(self.info),
includes=["pooling.h"])
def get_native_compile_info(self, input_types, devtype):
assert devtype == "gpu"
code = """
CGT_EXPORT_C void $setup(conv_closure* closure) {setup_cudnn(closure);}
CGT_EXPORT_C void $teardown(conv_closure* closure) {teardown_cudnn(closure);}
CGT_EXPORT_C void $function(conv_closure* closure, cgtArray** reads, cgtArray* write) {
if (!closure->handle) setup_cudnn(closure);
performConvBackwardData(closure, reads[1], reads[2], write);
}"""
return core.NativeCompileInfo(code,
closure_triples=make_closure(
self.ph, self.pw, self.sv, self.sh),
includes=["cudnn_support.h"],
link_flags="-lcudnn -lcudart")
def get_native_compile_info(self, _input_types, devtype):
assert devtype == "gpu"
code = """
CGT_EXPORT_C void $setup(pooling_closure* closure) {setup_cudnn(closure);}
CGT_EXPORT_C void $teardown(pooling_closure* closure) {teardown_cudnn(closure);}
CGT_EXPORT_C void $function(pooling_closure* closure, cgtArray** reads, cgtArray* write) {
if (!closure->handle) setup_cudnn(closure);
performPoolingForward(closure, reads[0], write);
}"""
return core.NativeCompileInfo(code,
closure_triples=poolinfo2closure(
self.info),
includes=["cudnn_support.h"],
link_flags="-lcudnn -lcudart")
def get_native_compile_info(self, input_types, devtype):
assert devtype == "cpu"
code = """
CGT_EXPORT_C void $function(void* cldata, cgtArray** reads, cgtTuple* write) {
float* x = static_cast<float*>(reads[0]->data());
float* y = static_cast<float*>(static_cast<cgtArray*>(write->getitem(0))->data());
float* z = static_cast<float*>(static_cast<cgtArray*>(write->getitem(1))->data());
for (int i=0; i < reads[0]->size(); ++i) {
y[i] = sinf(x[i]);
z[i] = cosf(x[i]);
}
}"""
return core.NativeCompileInfo(code,
includes=["math.h"],
link_flags="-lm")
def get_native_compile_info(self, input_types, devtype):
d = dict(cdtype=core.np2c[input_types[0].dtype])
d.update(self.info._asdict())
code = r"""
CGT_EXPORT_C void $function($closure* cl, cgtArray** reads, cgtArray* write) {
cgtArray* col = reads[0];
size_t batchsize = reads[1]->at<size_t>(0),
channels = reads[2]->at<size_t>(0),
height = reads[3]->at<size_t>(0),
width = reads[4]->at<size_t>(0);
for (int i=0; i < batchsize; ++i) {
col2im_cpu<%(cdtype)s, %(kernel_h)s,%(kernel_w)s,%(pad_h)s,%(pad_w)s,%(stride_h)s,%(stride_w)s>
((%(cdtype)s*)col->data() + col->stride(0)*i, channels, height, width,(%(cdtype)s*)write->data() + write->stride(0)*i);
}
}""" % d
return core.NativeCompileInfo(code,
includes=["im2col.h"],
closure_triples=info2closure(self.info))
def get_native_compile_info(self, input_types, devtype):
assert devtype == "cpu"
d = dict(cdtype=core.np2c[input_types[0].dtype])
d.update(self.info._asdict())
code = r"""
CGT_EXPORT_C void $function($closure* cl, cgtArray** reads, cgtArray* write) {
cgtArray* im = reads[0];
const size_t* imshape = im->shape();
int batchsize = imshape[0],
channels = imshape[1],
height = imshape[2],
width = imshape[3];
for (int i=0; i < batchsize; ++i) {
im2col_cpu<%(cdtype)s, %(kernel_h)s,%(kernel_w)s,%(pad_h)s,%(pad_w)s,%(stride_h)s,%(stride_w)s>
((%(cdtype)s*)im->data() + im->stride(0)*i, channels, height, width, (%(cdtype)s*)write->data() + write->stride(0)*i);
}
}""" % d
return core.NativeCompileInfo(code,
includes=["im2col.h"],
closure_triples=info2closure(self.info))
