void DataChannelMPI::receive(at::Tensor& data, rank_type src_rank) {
if (!data.is_contiguous())
throw std::logic_error("tensor to receive is not contiguous");
MPI_Recv(data.data_ptr(), data.numel(), mpi_datatype.at(data.type().scalarType()),
src_rank, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
}
void DataChannelMPI::send(at::Tensor& data, rank_type dst_rank) {
if (!data.is_contiguous())
throw std::logic_error("tensor to send is not contiguous");
MPI_Send(data.data_ptr(), data.numel(), mpi_datatype.at(data.type().scalarType()),
dst_rank, 0, MPI_COMM_WORLD);
}
at::Tensor sigmoid_add(at::Tensor x, at::Tensor y) {
AT_CHECK(x.type().is_cuda(), "x must be a CUDA tensor");
AT_CHECK(y.type().is_cuda(), "y must be a CUDA tensor");
auto output = at::zeros_like(x);
sigmoid_add_cuda(
x.data<float>(), y.data<float>(), output.data<float>(), output.numel());
return output;
}
rank_type DataChannelMPI::receive(at::Tensor& data) {
if (!data.is_contiguous())
throw std::logic_error("tensor to receive is not contiguous");
MPI_Status status;
MPI_Recv(data.data_ptr(), data.numel(), mpi_datatype.at(data.type().scalarType()),
MPI_ANY_SOURCE, 0, MPI_COMM_WORLD, &status);
return status.MPI_SOURCE;
}
void DataChannelMPI::allReduce(at::Tensor& data, THDReduceOp operation,
THDGroup group_id) {
const auto& comm = _groups.at(group_id).first;
if (comm == MPI_COMM_NULL)
return;
if (!data.is_contiguous())
throw std::runtime_error("all_reduce input has to be contiguous");
MPI_Allreduce(MPI_IN_PLACE, data.data_ptr(), data.numel(),
mpi_datatype.at(data.type().scalarType()), mpi_op.at(operation), comm);
}
void DataChannelMPI::broadcast(at::Tensor& data, rank_type src_rank,
THDGroup group_id) {
const auto& group_pair = _groups.at(group_id);
const auto& comm = group_pair.first;
if (comm == MPI_COMM_NULL)
return;
if (!data.is_contiguous())
throw std::runtime_error("broadcast input has to be contiguous");
rank_type group_src_rank = group_pair.second.mustGetGroupRank(src_rank);
MPI_Bcast(data.data_ptr(), data.numel(), mpi_datatype.at(data.type().scalarType()),
group_src_rank, comm);
}
void DataChannelMPI::reduce(at::Tensor& data, THDReduceOp operation,
rank_type dst_rank, THDGroup group_id) {
const auto& group_pair = _groups.at(group_id);
const auto& comm = group_pair.first;
if (comm == MPI_COMM_NULL)
return;
if (!data.is_contiguous())
throw std::runtime_error("reduce input has to be contiguous");
auto group_dst_rank = group_pair.second.mustGetGroupRank(dst_rank);
void *sendbuf = (_rank == dst_rank) ? MPI_IN_PLACE : data.data_ptr();
void *recvbuf = (_rank == dst_rank) ? data.data_ptr() : nullptr;
MPI_Reduce(sendbuf, recvbuf, data.numel(), mpi_datatype.at(data.type().scalarType()),
mpi_op.at(operation), group_dst_rank, comm);
}
void DataChannelMPI::allGather(std::vector<at::Tensor>& output,
at::Tensor& input, THDGroup group_id) {
const auto& group_pair = _groups.at(group_id);
const auto& comm = group_pair.first;
if (comm == MPI_COMM_NULL)
return;
if (output.size() != group_pair.second.size())
throw std::logic_error("allGather: number of output tensors and group size does not match");
for (auto out_tensor : output)
assertSameSizeAndType(out_tensor, input, "allGather");
auto recv_buffer = _newLikeFlat(output);
auto contig_input = input.contiguous();
MPI_Allgather(
contig_input.data_ptr(), contig_input.numel(), mpi_datatype.at(contig_input.type().scalarType()),
recv_buffer.data_ptr(), contig_input.numel(), mpi_datatype.at(recv_buffer.type().scalarType()),
comm
);
for (size_t i = 0; i < output.size(); ++i)
output[i].copy_(recv_buffer[i]);
}
std::tuple<Tensor, Tensor> fractional_max_pool2d_cpu(
const at::Tensor& input,
IntArrayRef pool_size,
IntArrayRef output_size,
const at::Tensor& randomSamples)
{
Tensor output = at::empty({0}, input.options());
Tensor indices = at::empty({0}, input.options().dtype(kLong));
fractional_max_pool2d_out_cpu_template(
input,
output,
output_size,
pool_size,
indices,
randomSamples);
return std::tuple<Tensor, Tensor>(output, indices);
}
std::tuple<at::Tensor,at::Tensor,at::Tensor> mkldnn_convolution_backward(
const at::Tensor& input, const at::Tensor& grad_output_t, const at::Tensor& weight,
IntList padding, IntList stride, IntList dilation, std::array<bool,3> output_mask)
{
Tensor grad_output = grad_output_t.contiguous();
Tensor grad_input, grad_weight, grad_bias;
if (output_mask[0]) {
grad_input = at::mkldnn_convolution_backward_input(
input.sizes(), grad_output, weight, padding, stride, dilation, output_mask[2]);
}
if (output_mask[1] || output_mask[2]) {
std::tie(grad_weight, grad_bias) = at::mkldnn_convolution_backward_weights(
weight.sizes(), grad_output, input, padding, stride, dilation, output_mask[2]);
}
return std::tuple<Tensor, Tensor, Tensor>{grad_input, grad_weight, grad_bias};
}
at::Tensor nms_cuda(const at::Tensor input,
float thresh)
{
AT_CHECK(input.ndimension() == 3,
"First argument should be a 3D Tensor, (batch_sz x n_boxes x 4)");
// AT_CHECK(scores.ndimens/ion() == 2,
// "Second argument should be a 2D Tensor, (batch_sz x n_boxes)");
// AT_CHECK(input.size(0) == scores.size(0),
// "First and second arguments must have equal-sized first dimensions");
// AT_CHECK(input.size(1) == scores.size(1),
// "First and second arguments must have equal-sized second dimensions");
AT_CHECK(input.size(2) == 4,
"First argument dimension 2 must have size 4, and should be of the form [x, y, w, h]");
AT_CHECK(input.is_contiguous(), "First argument must be a contiguous Tensor");
// AT_CHECK(scores.is_contiguous(), "Second argument must be a contiguous Tensor");
AT_CHECK(input.type().scalarType() == at::kFloat || input.type().scalarType() == at::kDouble,
"First argument must be Float or Double Tensor");
// AT_CHECK(scores.type().scalarType() == at::kFloat || scores.type().scalarType() == at::kDouble,
// "Second argument must be Float or Double Tensor");
AT_CHECK(input.is_contiguous(), "First argument must be a contiguous Tensor");
// AT_CHECK(scores.is_contiguous(), "Second argument must be a contiguous Tensor");
return non_max_suppression_cuda(input, thresh);
}
void DataChannelMPI::gather(std::vector<at::Tensor>& output,
at::Tensor& input, rank_type dst_rank,
THDGroup group_id) {
const auto& group_pair = _groups.at(group_id);
const auto& comm = group_pair.first;
if (comm == MPI_COMM_NULL)
return;
at::Tensor recv_buffer;
void *recvbuf = nullptr;
if (_rank != dst_rank) {
if (output.size() > 0)
throw std::logic_error("gather: number of input tensors should be 0 for non root");
} else {
if (output.size() != group_pair.second.size())
throw std::logic_error("gather: number of output tensors and group size does not match");
for (auto out_tensor : output)
assertSameSizeAndType(out_tensor, input, "gather");
recv_buffer = _newLikeFlat(output);
recvbuf = recv_buffer.data_ptr();
}
rank_type group_dst_rank = group_pair.second.mustGetGroupRank(dst_rank);
auto contig_input = input.contiguous();
MPI_Gather(
contig_input.data_ptr(), input.numel(), mpi_datatype.at(input.type().scalarType()),
recvbuf, input.numel(), mpi_datatype.at(input.type().scalarType()),
group_dst_rank, comm
);
// NOTE: this is a no-op in all processes except dst_rank
for (size_t i = 0; i < output.size(); ++i)
output[i].copy_(recv_buffer[i]);
}
Tensor fractional_max_pool2d_backward_cpu(
const at::Tensor& gradOutput_,
const at::Tensor& input,
IntArrayRef pool_size,
IntArrayRef output_size,
const at::Tensor& indices)
{
Tensor gradInput = at::empty({0}, input.options());
fractional_max_pool2d_backward_out_cpu_template(
input,
gradOutput_,
gradInput,
output_size,
pool_size,
indices);
return gradInput;
}
void DataChannelMPI::scatter(std::vector<at::Tensor>& input,
at::Tensor& output,
rank_type src_rank, THDGroup group_id) {
const auto& group_pair = _groups.at(group_id);
const auto& comm = group_pair.first;
if (comm == MPI_COMM_NULL)
return;
if (!output.is_contiguous())
throw std::runtime_error("scatter output has to be a contiguous tensor");
at::Tensor send_buffer;
void *sendbuf = nullptr;
if (_rank != src_rank) {
if (input.size() > 0)
throw std::logic_error("scatter: number of input tensors should be 0 for non root");
} else {
if (input.size() != group_pair.second.size())
throw std::logic_error("scatter: number of input tensors and group size does not match");
for (auto in_tensor : input)
assertSameSizeAndType(in_tensor, output, "scatter");
send_buffer = _newLikeFlat(input);
for (size_t i = 0; i < input.size(); ++i)
send_buffer[i].copy_(input[i]);
sendbuf = send_buffer.data_ptr();
}
rank_type group_src_rank = group_pair.second.mustGetGroupRank(src_rank);
MPI_Scatter(
sendbuf, output.numel(), mpi_datatype.at(output.type().scalarType()),
output.data_ptr(), output.numel(), mpi_datatype.at(output.type().scalarType()),
group_src_rank, comm
);
}
at::Tensor mkldnn_convolution(
const at::Tensor& input, const at::Tensor& weight, const at::Tensor& bias,
IntList padding, IntList stride, IntList dilation)
{
auto output = input.type().tensor(conv_output_size(
input.sizes(), weight.sizes(), padding, stride, dilation));
auto cpu_engine = CpuEngine::Instance().get_engine();
int32_t n = input.size(0);
int32_t ic = input.size(1);
int32_t ih = input.size(2);
int32_t iw = input.size(3);
int32_t oc = output.size(1);
int32_t oh = output.size(2);
int32_t ow = output.size(3);
int32_t kh = weight.size(2);
int32_t kw = weight.size(3);
int32_t sh = stride[0];
int32_t sw = stride[1];
int32_t ph = padding[0];
int32_t pw = padding[1];
auto data_t = memory::data_type::f32;
auto format_any = memory::format::any;
auto format_nchw = memory::format::nchw;
auto format_oihw = memory::format::oihw;
auto format_x = memory::format::x;
memory::dims input_tz = {n, ic, ih, iw};
memory::dims weight_tz = {oc, ic, kh, kw};
memory::dims bias_tz = {oc};
memory::dims output_tz = {n, oc, oh, ow};
memory::dims _stride = {sh, sw};
memory::dims _padding = {ph, pw};
auto input_md = memory::desc({input_tz}, data_t, format_any);
auto weight_md = memory::desc({weight_tz}, data_t, format_any);
auto bias_md = memory::desc({bias_tz}, data_t, format_any);
auto output_md = memory::desc({output_tz}, data_t, format_any);
std::shared_ptr<convolution_forward::desc> conv_forward_desc;
if (bias.defined()) {
conv_forward_desc.reset(new convolution_forward::desc(prop_kind::forward,
convolution_direct, input_md, weight_md, bias_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
} else {
conv_forward_desc.reset(new convolution_forward::desc(prop_kind::forward,
convolution_direct, input_md, weight_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
}
std::shared_ptr<convolution_forward::primitive_desc> conv_forward_pd;
conv_forward_pd.reset(new convolution_forward::primitive_desc(
*conv_forward_desc, cpu_engine));
auto input_usr_memory = memory({{{input_tz}, data_t, format_nchw}, cpu_engine},
input.data_ptr());
auto weight_usr_memory = memory({{{weight_tz}, data_t, format_oihw}, cpu_engine},
weight.data_ptr());
auto output_usr_memory = memory({{{output_tz}, data_t, format_nchw}, cpu_engine},
output.data_ptr());
std::vector<primitive> net;
auto input_pd = conv_forward_pd->src_primitive_desc();
auto input_memory = input_usr_memory;
if (input_usr_memory.get_primitive_desc() != memory::primitive_desc(input_pd)) {
input_memory = memory(input_pd);
net.push_back(reorder(input_usr_memory, input_memory));
}
auto weight_pd = conv_forward_pd->weights_primitive_desc();
auto weight_memory = weight_usr_memory;
if (weight_usr_memory.get_primitive_desc() != memory::primitive_desc(weight_pd)) {
weight_memory = memory(weight_pd);
net.push_back(reorder(weight_usr_memory, weight_memory));
}
auto output_pd = conv_forward_pd->dst_primitive_desc();
auto output_memory = output_usr_memory;
if (output_usr_memory.get_primitive_desc() != memory::primitive_desc(output_pd)) {
output_memory = memory(output_pd);
}
std::shared_ptr<convolution_forward> conv_forward;
std::shared_ptr<memory> bias_usr_memory;
if (bias.defined()) {
bias_usr_memory.reset(new memory({{{bias_tz}, data_t, format_x}, cpu_engine},
bias.data_ptr()));
conv_forward.reset(new convolution_forward(*conv_forward_pd, input_memory,
weight_memory, *bias_usr_memory, output_memory));
} else {
conv_forward.reset(new convolution_forward(*conv_forward_pd, input_memory,
weight_memory, output_memory));
}
net.push_back(*conv_forward);
if (output_memory != output_usr_memory) {
net.push_back(reorder(output_memory, output_usr_memory));
}
Stream::Instance().get_stream().submit(net);
return output;
}
std::tuple<at::Tensor, at::Tensor> mkldnn_convolution_backward_weights(
IntList weight_size, const at::Tensor& grad_output, const at::Tensor& input,
IntList padding, IntList stride, IntList dilation, bool bias_defined)
{
auto grad_weight = grad_output.type().tensor(weight_size);
Tensor grad_bias;
if (bias_defined) {
grad_bias = grad_output.type().tensor({grad_output.size(1)});
}
auto cpu_engine = CpuEngine::Instance().get_engine();
int32_t n = input.size(0);
int32_t ic = input.size(1);
int32_t ih = input.size(2);
int32_t iw = input.size(3);
int32_t oc = grad_output.size(1);
int32_t oh = grad_output.size(2);
int32_t ow = grad_output.size(3);
int32_t kh = grad_weight.size(2);
int32_t kw = grad_weight.size(3);
int32_t sh = stride[0];
int32_t sw = stride[1];
int32_t ph = padding[0];
int32_t pw = padding[1];
auto data_t = memory::data_type::f32;
auto format_any = memory::format::any;
auto format_nchw = memory::format::nchw;
auto format_oihw = memory::format::oihw;
auto format_x = memory::format::x;
memory::dims input_tz = {n, ic, ih, iw};
memory::dims weight_tz = {oc, ic, kh, kw};
memory::dims bias_tz = {oc};
memory::dims output_tz = {n, oc, oh, ow};
memory::dims _stride = {sh, sw};
memory::dims _padding = {ph, pw};
memory::desc input_md({input_tz}, data_t, format_any);
memory::desc weight_md({weight_tz}, data_t, format_any);
memory::desc bias_md({bias_tz}, data_t, format_any);
memory::desc output_md({output_tz}, data_t, format_any);
// need to re-create conv_forward_pd to feed conv_backward_weight_pd
std::shared_ptr<convolution_forward::desc> conv_forward_desc;
if (bias_defined) {
conv_forward_desc.reset(new convolution_forward::desc(prop_kind::forward,
convolution_direct, input_md, weight_md, bias_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
} else {
conv_forward_desc.reset(new convolution_forward::desc(prop_kind::forward,
convolution_direct, input_md, weight_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
}
std::shared_ptr<convolution_forward::primitive_desc> conv_forward_pd;
conv_forward_pd.reset(new convolution_forward::primitive_desc(
*conv_forward_desc, cpu_engine));
std::shared_ptr<convolution_backward_weights::desc> conv_backward_weight_desc;
if (bias_defined) {
conv_backward_weight_desc.reset(new convolution_backward_weights::desc(
convolution_direct, input_md, weight_md, bias_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
} else {
conv_backward_weight_desc.reset(new convolution_backward_weights::desc(
convolution_direct, input_md, weight_md, output_md,
_stride, _padding, _padding, padding_kind::zero));
}
std::shared_ptr<convolution_backward_weights::primitive_desc> conv_backward_weight_pd;
conv_backward_weight_pd.reset(new convolution_backward_weights::primitive_desc(
*conv_backward_weight_desc, cpu_engine, *conv_forward_pd));
auto input_usr_memory = memory({{{input_tz}, data_t, format_nchw}, cpu_engine},
input.data_ptr());
auto grad_output_usr_memory = memory({{{output_tz}, data_t, format_nchw}, cpu_engine},
grad_output.data_ptr());
auto grad_weight_usr_memory = memory({{{weight_tz}, data_t, format_oihw}, cpu_engine},
grad_weight.data_ptr());
std::shared_ptr<memory> grad_bias_memory;
std::vector<primitive> net;
auto input_pd = conv_backward_weight_pd->src_primitive_desc();
auto input_memory = input_usr_memory;
if (input_usr_memory.get_primitive_desc() != memory::primitive_desc(input_pd)) {
input_memory = memory(input_pd);
net.push_back(reorder(input_usr_memory, input_memory));
}
auto grad_output_pd = conv_backward_weight_pd->diff_dst_primitive_desc();
auto grad_output_memory = grad_output_usr_memory;
if (grad_output_usr_memory.get_primitive_desc() != memory::primitive_desc(grad_output_pd)) {
grad_output_memory = memory(grad_output_pd);
net.push_back(reorder(grad_output_usr_memory, grad_output_memory));
}
auto grad_weight_pd = conv_backward_weight_pd->diff_weights_primitive_desc();
auto grad_weight_memory = grad_weight_usr_memory;
if (grad_weight_usr_memory.get_primitive_desc() != memory::primitive_desc(grad_weight_pd)) {
grad_weight_memory = memory(grad_weight_pd);
}
std::shared_ptr<convolution_backward_weights> conv_backward_weight;
if (bias_defined) {
grad_bias_memory.reset(new memory({{{bias_tz}, data_t, format_x}, cpu_engine},
grad_bias.data_ptr()));
conv_backward_weight.reset(new convolution_backward_weights(*conv_backward_weight_pd,
input_memory, grad_output_memory, grad_weight_memory, *grad_bias_memory));
} else {
conv_backward_weight.reset(new convolution_backward_weights(*conv_backward_weight_pd,
input_memory, grad_output_memory, grad_weight_memory));
}
net.push_back(*conv_backward_weight);
if (grad_weight_memory != grad_weight_usr_memory) {
net.push_back(reorder(grad_weight_memory, grad_weight_usr_memory));
}
Stream::Instance().get_stream().submit(net);
return std::tuple<at::Tensor, at::Tensor>{grad_weight, grad_bias};
}