size_t fread__(void *ptr, size_t size, size_t nitems, FILE *stream)
{
size_t nread = 0;
while(!feof(stream) && !ferror(stream) && (nread < nitems))
nread += fread((char*)ptr+nread*size, size, THMin(2147483648/size, nitems-nread), stream);
return nread;
}
void THPStorage_(writeFileRaw)(THStorage *self, io fd)
{
real *data;
int64_t size = self->size;
#ifndef THC_GENERIC_FILE
data = self->data;
#else
std::unique_ptr<char[]> cpu_data(new char[size * sizeof(real)]);
data = (real*)cpu_data.get();
THCudaCheck(cudaMemcpy(data, self->data, size * sizeof(real), cudaMemcpyDeviceToHost));
#endif
ssize_t result = doWrite(fd, &size, sizeof(int64_t));
if (result != sizeof(int64_t))
throw std::system_error(result, std::system_category());
// fast track for bytes and little endian
if (sizeof(real) == 1 || THP_nativeByteOrder() == THPByteOrder::THP_LITTLE_ENDIAN) {
char *bytes = (char *) data;
int64_t remaining = sizeof(real) * size;
while (remaining > 0) {
// we write and read in 1GB blocks to avoid bugs on some OSes
ssize_t result = doWrite(fd, bytes, THMin(remaining, 1073741824));
if (result < 0)
throw std::system_error(result, std::system_category());
bytes += result;
remaining -= result;
}
if (remaining != 0)
throw std::system_error(result, std::system_category());
} else {
int64_t buffer_size = std::min(size, (int64_t)5000);
std::unique_ptr<uint8_t[]> le_buffer(new uint8_t[buffer_size * sizeof(real)]);
for (int64_t i = 0; i < size; i += buffer_size) {
size_t to_convert = std::min(size - i, buffer_size);
if (sizeof(real) == 2) {
THP_encodeInt16Buffer((uint8_t*)le_buffer.get(),
(const int16_t*)data + i,
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
} else if (sizeof(real) == 4) {
THP_encodeInt32Buffer((uint8_t*)le_buffer.get(),
(const int32_t*)data + i,
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
} else if (sizeof(real) == 8) {
THP_encodeInt64Buffer((uint8_t*)le_buffer.get(),
(const int64_t*)data + i,
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
}
SYSCHECK(doWrite(fd, le_buffer.get(), to_convert * sizeof(real)));
}
}
}
THStorage * THPStorage_(readFileRaw)(io file, THStorage *_storage)
{
real *data;
int64_t size;
ssize_t result = doRead(file, &size, sizeof(int64_t));
if (result == 0)
throw std::runtime_error("unexpected EOF. The file might be corrupted.");
if (result != sizeof(int64_t))
throw std::system_error(result, std::system_category());
THStoragePtr storage;
if (_storage == nullptr) {
storage = THStorage_(newWithSize)(LIBRARY_STATE size);
} else {
THPUtils_assert(_storage->size == size,
"storage has wrong size: expected %ld got %ld",
size, _storage->size);
storage = _storage;
}
#ifndef THC_GENERIC_FILE
data = storage->data;
#else
std::unique_ptr<char[]> cpu_data(new char[size * sizeof(real)]);
data = (real*)cpu_data.get();
#endif
// fast track for bytes and little endian
if (sizeof(real) == 1 || THP_nativeByteOrder() == THPByteOrder::THP_LITTLE_ENDIAN) {
char *bytes = (char *) data;
int64_t remaining = sizeof(real) * storage->size;
while (remaining > 0) {
// we write and read in 1GB blocks to avoid bugs on some OSes
ssize_t result = doRead(file, bytes, THMin(remaining, 1073741824));
if (result == 0) // 0 means EOF, which is also an error
throw std::runtime_error("unexpected EOF. The file might be corrupted.");
if (result < 0)
throw std::system_error(result, std::system_category());
bytes += result;
remaining -= result;
}
if (remaining != 0)
throw std::system_error(result, std::system_category());
} else {
int64_t buffer_size = std::min(size, (int64_t)5000);
std::unique_ptr<uint8_t[]> le_buffer(new uint8_t[buffer_size * sizeof(real)]);
for (int64_t i = 0; i < size; i += buffer_size) {
size_t to_convert = std::min(size - i, buffer_size);
SYSCHECK(doRead(file, le_buffer.get(), sizeof(real) * to_convert));
if (sizeof(real) == 2) {
THP_decodeInt16Buffer((int16_t*)data + i,
le_buffer.get(),
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
} else if (sizeof(real) == 4) {
THP_decodeInt32Buffer((int32_t*)data + i,
le_buffer.get(),
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
} else if (sizeof(real) == 8) {
THP_decodeInt64Buffer((int64_t*)data + i,
le_buffer.get(),
THPByteOrder::THP_LITTLE_ENDIAN,
to_convert);
}
}
}
#ifdef THC_GENERIC_FILE
THCudaCheck(cudaMemcpy(storage->data, data, size * sizeof(real), cudaMemcpyHostToDevice));
#endif
return storage.release();
}
void THCStorage_resize(THCState *state, THCStorage *self, ptrdiff_t size)
{
THArgCheck(size >= 0, 2, "invalid size");
THAssert(self->allocator != NULL);
int device;
THCudaCheck(cudaGetDevice(&device));
if(!(self->flag & TH_STORAGE_RESIZABLE))
THError("Trying to resize storage that is not resizable");
size_t elementSize = at::elementSize(self->scalar_type);
if (self->allocator->realloc) {
void * data_ptr = self->data_ptr;
cudaError_t err = (*self->allocator->realloc)(
self->allocatorContext,
(void**)&(data_ptr),
self->size * elementSize,
size * elementSize, THCState_getCurrentStreamOnDevice(state, device));
if (err != cudaSuccess) {
THCudaCheck(err);
}
self->size = size;
self->device = device;
return;
}
if(size == 0)
{
if(self->flag & TH_STORAGE_FREEMEM) {
THCudaCheck(
(*self->allocator->free)(self->allocatorContext, self->data_ptr));
}
self->data_ptr = NULL;
self->size = 0;
self->device = device;
}
else
{
void *data = NULL;
cudaError_t err =
(*self->allocator->malloc)(self->allocatorContext,
(void**)&(data),
size * elementSize,
THCState_getCurrentStreamOnDevice(state, device));
THCudaCheck(err);
if (self->data_ptr) {
// Enable p2p access when the memcpy is across devices
THCState_getPeerToPeerAccess(state, device, self->device);
THCudaCheck(cudaMemcpyAsync(data,
self->data_ptr,
THMin(self->size, size) * elementSize,
cudaMemcpyDeviceToDevice,
THCState_getCurrentStream(state)));
if(self->flag & TH_STORAGE_FREEMEM) {
THCudaCheck(
(*self->allocator->free)(self->allocatorContext, self->data_ptr));
}
}
self->data_ptr = data;
self->size = size;
self->device = device;
}
}