Dynamic weight handling

src/include/migraphx/op/convolution.hpp

@@ -62,23 +62,21 @@ struct convolution
             MIGRAPHX_THROW("CONVOLUTION: input and attribute size mismatch!");
         }
 
-        const shape& input   = inputs.at(0);
-        const shape& weights = inputs.at(1);
-        if(weights.dynamic())
-        {
-            MIGRAPHX_THROW("CONVOLUTION: dynamic weights not supported");
-        }
+        const shape& input            = inputs.at(0);
+        const shape& weights          = inputs.at(1);
         const size_t num_spatial_dims = input_size - 2;
         if(num_spatial_dims != this->kdims())
         {
             MIGRAPHX_THROW("CONVOLUTION: input k-dims does not match attribute size");
         }
 
-        if(!input.dynamic() and input.lens().at(1) != (weights.lens().at(1) * group))
+        if(not input.dynamic() and not weights.dynamic() and
+           input.lens().at(1) != (weights.lens().at(1) * group))
             MIGRAPHX_THROW("CONVOLUTION: mismatched channel numbers");
 
         auto calc_output_lens =
-            [this, &weights, &num_spatial_dims, &padding_size](std::vector<std::size_t> lens) {
+            [this, &num_spatial_dims, &padding_size](std::vector<std::size_t> i_lens,
+                                                     std::vector<std::size_t> w_lens) {
                 std::vector<size_t> ret = {};
                 // calculate the output shape of the convolution: ((W - K + 2P) / S) + 1
                 for(size_t i = 0; i < num_spatial_dims; i++)
@@ -91,8 +89,7 @@ struct convolution
                     }
                     ret.push_back(std::size_t(std::max<std::ptrdiff_t>(
                         1,
-                        (lens[i + 2] - (1 + dilation[i] * (weights.lens()[i + 2] - 1)) +
-                         padding_factor) /
+                        (i_lens[i + 2] - (1 + dilation[i] * (w_lens[i + 2] - 1)) + padding_factor) /
                                 stride[i] +
                             1)));
                 }
@@ -103,9 +100,9 @@ struct convolution
         {
             std::vector<shape::dynamic_dimension> output_dyn_dims = {input.dyn_dims().at(0),
                                                                      input.dyn_dims().at(1)};
-            auto min_spatial_dims = calc_output_lens(input.min_lens());
-            auto max_spatial_dims = calc_output_lens(input.max_lens());
-            auto opt_spatial_dims = calc_output_lens(input.opt_lens());
+            auto min_spatial_dims = calc_output_lens(input.min_lens(), weights.min_lens());
+            auto max_spatial_dims = calc_output_lens(input.max_lens(), weights.max_lens());
+            auto opt_spatial_dims = calc_output_lens(input.opt_lens(), weights.opt_lens());
             for(size_t i = 0; i < num_spatial_dims; ++i)
             {
                 output_dyn_dims.push_back(shape::dynamic_dimension{
@@ -116,7 +113,7 @@ struct convolution
         else
         {
             std::vector<size_t> output_lens{input.lens()[0], weights.lens()[0]};
-            auto spatial_lens = calc_output_lens(input.lens());
+            auto spatial_lens = calc_output_lens(input.lens(), weights.lens());
             std::for_each(spatial_lens.begin(), spatial_lens.end(), [&output_lens](auto x) {
                 output_lens.push_back(x);
             });

src/program.cpp

@@ -312,7 +312,9 @@ std::vector<argument> generic_eval(const module* mod,
                     return shapes;
                 };
                 // TODO: Consider how this will be handled when memoized.
-                // Could memoize these output shapes now so not recalculating
+                // Could memoize these output shapes into a map so not recalculating
+                // TODO: Issue with incompatible input tensor to kernel and needing to set
+                // padding/strides
                 output_shape = ins->get_operator().compute_shape(to_shapes(values));
             }
             else
@@ -333,7 +335,6 @@ std::vector<argument> generic_eval(const module* mod,
                             }));
         }
         assert(results.find(ins) != results.end());
-        // TODO: update this assert for dynamic shapes
         if(not ins->get_shape().dynamic())
         {
             assert(results.at(ins).get_shape() == ins->get_shape());