Merge branch 'master' into migraphx

src/include/migraphx/operators.hpp

@@ -285,6 +285,12 @@ struct transpose
     int output_alias(const std::vector<shape>&) const { return 0; }
 };
 
+/// The contiguous operator takes a non-standard input tensor and returns
+/// the same tensor but in standard form. For example, if input tensor A which has lens = (4,5)
+/// is first transposed, i.e. lens = (5,4), this tensor's data layout remained the same
+/// during the transpose operation; only it's shape lengths and strides were changed.
+/// This leaves the tensor in a non-standard form. The contiguous operator copies the
+/// underlying data such that resulting tensor is returned to a standard form.
 struct contiguous
 {
     std::string name() const { return "contiguous"; }
@@ -400,15 +406,6 @@ struct slice
         auto t                  = input_shape.type();
         const auto& old_lens    = input_shape.lens();
         const auto& old_strides = input_shape.strides();
-        // std::vector<int64_t> t_axes(old_lens.size());
-        // if(axes.size() == 0)
-        // {
-        //     std::iota(t_axes.begin(), t_axes.end(), 0);
-        // }
-        // else
-        // {
-        //     std::copy(axes.begin(), axes.end(), t_axes.begin());
-        // }
         if(starts.size() != axes.size() || axes.size() != ends.size())
         {
             MIGRAPHX_THROW("inconsistent sizes");
@@ -718,6 +715,15 @@ struct flatten
     }
     int output_alias(const std::vector<shape>&) const { return 0; }
 };
+
+/// The broadcast operator performs the numpy-style broadcasting of an axis of a given tensor. This
+/// is achieved primarily by setting the stride of the broadcasted axis to zero. Linear indicies are
+/// computed from multi-indicies by computing the inner product on the multi-index with the strides.
+/// For example, if we have a tensor A(2,3) it has lengths of (2,3) and strides of (3,1). If we want
+/// to compute the linear offset that corresponds to the element on the 2nd row (i = 1) and 3rd
+/// column (j = 2), we compute the following inner product (1,2) dot (3, 1) = 1*3 + 2*1 = 5. It is
+/// obvious from there that we can negate the effects of a given axis by setting the stride of that
+/// axis to zero.
 struct broadcast
 {
     uint64_t axis = 0;

src/onnx/onnx.cpp

@@ -44,6 +44,7 @@ struct onnx_parser
     node_map nodes;
     std::unordered_map<std::string, instruction_ref> instructions;
     program prog    = program();
+    bool is_pytorch = false;
 
     std::unordered_map<std::string, op_func> ops;
 
@@ -138,7 +139,7 @@ struct onnx_parser
                     std::swap(s0, s1);
 
                 // Copy the larger vector to output_lens
-                std::vector<std::size_t> output_lens(s1->size());
+                std::vector<std::size_t> output_lens = *s1;
                 auto offset                          = s1->size() - s0->size();
                 std::transform(s0->begin(),
                                s0->end(),
@@ -181,7 +182,22 @@ struct onnx_parser
         op::convolution op;
         if(contains(attributes, "pads"))
         {
-            copy(attributes["pads"].ints(), op.padding.begin());
+            if(contains(attributes, "auto_pad"))
+            {
+                MIGRAPH_THROW("auto_pad and padding cannot be specified simultaneously");
+            }
+            std::vector<std::size_t> padding(4);
+            copy(attributes["pads"].ints(), padding.begin());
+            if(padding.size() != 4)
+            {
+                MIGRAPH_THROW("padding should have 4 values");
+            }
+            if(padding[0] != padding[2] || padding[1] != padding[3])
+            {
+                MIGRAPH_THROW("migraphx does not support asymetric padding");
+            }
+            op.padding[0] = padding[0];
+            op.padding[1] = padding[1];
         }
         if(contains(attributes, "strides"))
         {
@@ -191,6 +207,19 @@ struct onnx_parser
         {
             copy(attributes["dilations"].ints(), op.dilation.begin());
         }
+        if(contains(attributes, "auto_pad"))
+        {
+            auto s = attributes["auto_pad"].s();
+            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
+            {
+                MIGRAPH_THROW("auto_pad and padding cannot be specified simultaneously");
+            }
+
+            if(s.find("SAME") != std::string::npos)
+            {
+                op.padding_mode = op::convolution::same;
+            }
+        }
         if(args.size() == 3)
         {
             uint64_t axis = 1;
@@ -213,7 +242,18 @@ struct onnx_parser
         }
         if(contains(attributes, "pads"))
         {
-            copy(attributes["pads"].ints(), op.padding.begin());
+            std::vector<std::size_t> padding(4);
+            copy(attributes["pads"].ints(), padding.begin());
+            if(padding.size() != 4)
+            {
+                MIGRAPH_THROW("padding should have 4 values");
+            }
+            if(padding[0] != padding[2] || padding[1] != padding[3])
+            {
+                MIGRAPH_THROW("migraphx does not support asymetric padding");
+            }
+            op.padding[0] = padding[0];
+            op.padding[1] = padding[1];
         }
         if(contains(attributes, "strides"))
         {
@@ -223,6 +263,15 @@ struct onnx_parser
         {
             copy(attributes["kernel_shape"].ints(), op.lengths.begin());
         }
+        if(contains(attributes, "auto_pad"))
+        {
+            auto s = attributes["auto_pad"].s();
+            if(to_upper(s) != "NOTSET")
+            {
+                MIGRAPH_THROW("auto_pad is not supported for pooling");
+            }
+        }
+
         return prog.add_instruction(op, std::move(args));
     }