haojin2 commented on a change in pull request #15169: Softmax with length
URL: https://github.com/apache/incubator-mxnet/pull/15169#discussion_r302931319
##########
File path: src/operator/nn/softmax-inl.h
##########
@@ -464,35 +755,56 @@ void SoftmaxGradCompute(const nnvm::NodeAttrs& attrs,
mxnet::TShape shape = AxisShapeCompact(inputs[0].shape_, &axis, true);
int out_idx = softmax_has_dtype_override(attrs) ? 2 : 1;
+ out_idx = softmax_use_length(attrs) ? 3 : out_idx;
bool safe_acc = dmlc::GetEnv("MXNET_SAFE_ACCUMULATION", false);
MXNET_REAL_ACC_TYPE_SWITCH(inputs[0].type_flag_, OType, AType, {
MSHADOW_REAL_TYPE_SWITCH(outputs[0].type_flag_, DType, {
MXNET_ASSIGN_REQ_SWITCH(req[0], Req, {
- if (safe_acc) {
- if (shape.ndim() == 2) {
- SoftmaxGrad<OP1, OP2, Req, negate, AType>(
- ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
- inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
- shape.get<2>(), axis, static_cast<DType>(temperature));
+ if (!softmax_use_length(attrs)) {
+ if (safe_acc) {
+ if (shape.ndim() == 2) {
+ SoftmaxGrad<OP1, OP2, Req, negate, AType>(
+ ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
+ inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
+ shape.get<2>(), axis, static_cast<DType>(temperature));
+ } else {
+ SoftmaxGrad<OP1, OP2, Req, negate, AType>(
+ ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
+ inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
+ shape.get<3>(), axis, static_cast<DType>(temperature));
+ }
} else {
- SoftmaxGrad<OP1, OP2, Req, negate, AType>(
- ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
- inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
- shape.get<3>(), axis, static_cast<DType>(temperature));
+ if (shape.ndim() == 2) {
+ SoftmaxGrad<OP1, OP2, Req, negate, DType>(
+ ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
+ inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
+ shape.get<2>(), axis, static_cast<DType>(temperature));
+ } else {
+ SoftmaxGrad<OP1, OP2, Req, negate, DType>(
+ ctx.get_stream<xpu>(), inputs[out_idx].dptr<OType>(),
+ inputs[0].dptr<OType>(), outputs[0].dptr<DType>(),
+ shape.get<3>(), axis, static_cast<DType>(temperature));
+ }
}
} else {
- if (shape.ndim() == 2) {
- SoftmaxGrad<OP1, OP2, Req, negate, DType>(
+ MXNET_INT_TYPE_SWITCH(inputs[2].type_flag_, IType, {
Review comment:
It may not necessarily be int64_t, and what do you mean by "cast the type to
int64_t"? Allocating a new buffer on the fly within the operator to hold the
casted `length` input? I would consider that a big performance bottleneck.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
