masahi commented on a change in pull request #4418: [RUNTIME] Add cudnn conv3d
URL: https://github.com/apache/incubator-tvm/pull/4418#discussion_r352275702
##########
File path: src/runtime/contrib/cudnn/conv_forward.cc
##########
@@ -287,6 +355,91 @@ TVM_REGISTER_GLOBAL("tvm.contrib.cudnn.conv2d.find_algo")
}
ret[0] = best_algo;
+}
+
+
+TVM_REGISTER_GLOBAL("tvm.contrib.cudnn.conv.forward")
+.set_body([](TVMArgs args, TVMRetValue *ret) {
+ int mode = args[0];
+ int format = args[1];
+ int algo = args[2];
+ int dims = args[3];
+
+ Array<Expr> pads = args[4];
+ Array<Expr> strides = args[5];
+ Array<Expr> dilations = args[6];
+
+ std::vector<int> pad_v(dims), stride_v(dims), dilation_v(dims);
+
+ for (auto &pad : pads) {
+ int i = 0;
+ pad_v[i++] = pad.as<IntImm>()->value;
+ }
+ for (auto &stride : strides) {
+ int i = 0;
+ pad_v[i++] = strides.as<IntImm>()->value;
+ }
+ for (auto &dilation : dilations) {
+ int i = 0;
+ pad_v[i++] = dilations.as<IntImm>()->value;
+ }
+
+ DLTensor* x = args[7];
+ DLTensor* w = args[8];
+ DLTensor* y = args[9];
+ std::string conv_dtype = args[10];
+
+ ConvolutionForward(mode, format, algo, pad_v, stride_v, dilation_v,
+ x, w, y, conv_dtype);
+});
+
+
+TVM_REGISTER_GLOBAL("tvm.contrib.cudnn.conv.output_shape")
+.set_body([](TVMArgs args, TVMRetValue *ret) {
+ int format = args[0];
+ int dims = args[1];
+ int* pad = static_cast<int*>(static_cast<void*>(args[2]));
+ int* stride = static_cast<int*>(static_cast<void*>(args[3]));
+ int* dilation = static_cast<int*>(static_cast<void*>(args[4]));
+ int* x_dim = static_cast<int*>(static_cast<void*>(args[5]));
+ int* w_dim = static_cast<int*>(static_cast<void*>(args[6]));
+ std::vector<int> pad_v(pad, pad + dims);
+ std::vector<int> stride_v(stride, stride + dims);
+ std::vector<int> dilation_v(dilation, dilation + dims);
+ std::vector<int> x_dim_v(x_dim, x_dim + dims + 2);
+ std::vector<int> w_dim_v(w_dim, w_dim + dims + 2);
+
Review comment:
All you do with these vectors in OutputShape and FindAlgo is just .data()
right? So why not pass pad, stride (of int*) etc directly rather than pad_v,
stride_v (you need to update OutputShape and FindAlgo, of course). Is there
something you can only do with vectors?
----------------------------------------------------------------
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
