void ContextProjectionBackward<DEVICE_TYPE_CPU>(const CpuMatrix& out_grad_mat,
CpuMatrix& in_grad_mat,
CpuMatrix& w_grad_mat,
const CpuIVector& seq_vec,
size_t context_length,
int context_start,
size_t begin_pad,
bool is_padding,
size_t total_pad) {
size_t input_dim = in_grad_mat ? in_grad_mat.getWidth()
: w_grad_mat ? w_grad_mat.getWidth() : 0;
const int* starts = seq_vec.getData();
size_t num_sequences = seq_vec.getSize() - 1;
for (size_t i = 0; i < num_sequences; ++i) {
for (size_t j = 0; j < context_length; ++j) {
int begin = starts[i] + context_start + j;
int end = starts[i + 1] + context_start + j;
int dst_begin = starts[i];
int dst_end = starts[i + 1];
if (begin < starts[i]) {
int64_t pad_size =
std::min(starts[i] - begin, starts[i + 1] - starts[i]);
if (is_padding && w_grad_mat) {
MatrixPtr mat = const_cast<CpuMatrix&>(out_grad_mat)
.subMatrix(starts[i], pad_size);
MatrixPtr sub = w_grad_mat.subMatrix(j, pad_size);
sub->addAtOffset(*mat, j * input_dim);
}
dst_begin = starts[i] + pad_size;
begin = starts[i];
}
if (end > starts[i + 1]) {
int64_t pad_size =
std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
if (is_padding && w_grad_mat) {
MatrixPtr mat = const_cast<CpuMatrix&>(out_grad_mat)
.subMatrix(starts[i + 1] - pad_size, pad_size);
MatrixPtr sub = w_grad_mat.subMatrix(
begin_pad + context_start + j - pad_size, pad_size);
sub->addAtOffset(*mat, j * input_dim);
}
dst_end = starts[i + 1] - pad_size;
end = starts[i + 1];
}
if (end <= begin) continue;
if (!in_grad_mat) continue;
MatrixPtr src = in_grad_mat.subMatrix(begin, end - begin);
MatrixPtr dst = const_cast<CpuMatrix&>(out_grad_mat)
.subMatrix(dst_begin, dst_end - dst_begin);
src->addAtOffset(*dst, j * input_dim);
}
}
}
void ContextProjectionForward<DEVICE_TYPE_CPU>(CpuMatrix& out_mat,
const CpuMatrix& input_mat,
const CpuMatrix& weight_mat,
const CpuIVector& seq_vec,
size_t context_length,
int context_start,
size_t begin_pad) {
const int* starts = seq_vec.getData();
const size_t num_sequences = seq_vec.getSize() - 1;
for (size_t i = 0; i < num_sequences; ++i) {
for (size_t j = 0; j < context_length; ++j) {
int begin = starts[i] + context_start + j;
int end = starts[i + 1] + context_start + j;
int dst_begin = starts[i];
int dst_end = starts[i + 1];
if (begin < starts[i]) {
int64_t pad_size =
std::min(starts[i] - begin, starts[i + 1] - starts[i]);
MatrixPtr mat = out_mat.subMatrix(starts[i], pad_size);
if (weight_mat) {
MatrixPtr sub =
const_cast<CpuMatrix&>(weight_mat).subMatrix(j, pad_size);
mat->addAtOffset(*sub, j * input_mat.getWidth());
}
dst_begin = starts[i] + pad_size;
begin = starts[i];
}
if (end > starts[i + 1]) {
int64_t pad_size =
std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
MatrixPtr mat = out_mat.subMatrix(starts[i + 1] - pad_size, pad_size);
if (weight_mat) {
MatrixPtr sub =
const_cast<CpuMatrix&>(weight_mat)
.subMatrix(begin_pad + context_start + j - pad_size,
pad_size);
mat->addAtOffset(*sub, j * input_mat.getWidth());
}
dst_end = starts[i + 1] - pad_size;
end = starts[i + 1];
}
if (end <= begin) continue;
MatrixPtr src =
const_cast<CpuMatrix&>(input_mat).subMatrix(begin, end - begin);
MatrixPtr dst = out_mat.subMatrix(dst_begin, dst_end - dst_begin);
dst->addAtOffset(*src, j * input_mat.getWidth());
}
}
}
void FunctionApi<DEVICE_TYPE_CPU>(CpuMatrix& output, const CpuMatrix& input) {
EXPECT_EQ(output.getHeight(), 100U);
EXPECT_EQ(output.getWidth(), 200U);
}