You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/11/13 05:50:35 UTC
[GitHub] [incubator-tvm] anijain2305 opened a new pull request #6912: [ShapeFunc] Handle weights in shape func
anijain2305 opened a new pull request #6912:
URL: https://github.com/apache/incubator-tvm/pull/6912
When compiling a dynamic model with TRT, the primitive function has constants already bound in the graph. TVM codegen typically lifts those into variables in ANF pass, so we never see this problem in TVM codegen.
@icemelon9 @zhiics @codeislife99 @rohanmukh @trevor-m
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [incubator-tvm] anijain2305 merged pull request #6912: [ShapeFunc] Handle weights in shape func
Posted by GitBox <gi...@apache.org>.
anijain2305 merged pull request #6912:
URL: https://github.com/apache/incubator-tvm/pull/6912
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [incubator-tvm] icemelon9 commented on a change in pull request #6912: [ShapeFunc] Handle weights in shape func
Posted by GitBox <gi...@apache.org>.
icemelon9 commented on a change in pull request #6912:
URL: https://github.com/apache/incubator-tvm/pull/6912#discussion_r523310191
##########
File path: src/relay/backend/compile_engine.cc
##########
@@ -420,38 +420,59 @@ class MakeShapeFunc : public backend::MemoizedExprTranslator<Array<te::Tensor>>
Array<te::Tensor> VisitExpr_(const ConstantNode* op) final {
using tir::make_const;
ICHECK(data_dependants_.size());
- ICHECK(op->is_scalar());
bool data_dependant = data_dependants_.back();
- if (data_dependant) {
- void* data = op->data->data;
- DataType dtype = DataType(op->data->dtype);
- auto value = tvm::te::compute(
- {},
- [&](const Array<tvm::tir::Var>&) {
- if (dtype == DataType::Int(32)) {
- return make_const(dtype, static_cast<const int32_t*>(data)[0]);
- } else if (dtype == DataType::Int(64)) {
- return make_const(dtype, static_cast<const int64_t*>(data)[0]);
- } else if (dtype == DataType::Float(32)) {
- return make_const(dtype, static_cast<const float*>(data)[0]);
- } else if (dtype == DataType::Float(64)) {
- return make_const(dtype, static_cast<const double*>(data)[0]);
- } else if (dtype == DataType::Bool()) {
- return make_const(dtype, static_cast<const uint8_t*>(data)[0]);
- } else {
- LOG(FATAL) << "not handled";
- return tvm::PrimExpr();
+ if (!op->is_scalar()) {
+ // This is a constant weight, extract the shape of the weight tensor.
+ // This can not be data dependent.
+ CHECK(!data_dependant);
+ auto ttype = op->checked_type().as<TensorTypeNode>();
+ int ndim = static_cast<int>(ttype->shape.size());
+ Array<PrimExpr> out_shape{ndim};
+ te::Tensor value = tvm::te::compute(
+ out_shape,
+ [&](const Array<tvm::tir::Var>& indices) {
+ auto idx = indices[0];
+ PrimExpr ret = make_const(DataType::Int(64), 0);
+ for (int i = 0; i < ndim; i++) {
+ ret = tvm::if_then_else(idx == i, ttype->shape[i], ret);
}
+ return ret;
},
- "data_const", topi::kBroadcast);
- scalars_.push_back(value);
- return {value};
- } else {
- auto value = tvm::te::compute(
- {}, [&](const Array<tvm::tir::Var>&) { return tir::make_const(DataType::Int(64), 0); },
"shape_const", topi::kBroadcast);
scalars_.push_back(value);
return {value};
+ } else {
Review comment:
No need for else since you already return.
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [incubator-tvm] anijain2305 commented on pull request #6912: [ShapeFunc] Handle weights in shape func
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on pull request #6912:
URL: https://github.com/apache/incubator-tvm/pull/6912#issuecomment-727246034
Thanks @zhiics @icemelon9 This is merged.
----------------------------------------------------------------
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:
users@infra.apache.org