clean up the quantize_ins implementation

src/include/migraphx/op/fp_conversion.hpp

@@ -31,7 +31,7 @@ struct fp_conversion
     shape compute_shape(std::vector<shape> inputs) const
     {
         check_shapes{inputs, *this}.has(1);
-        return {targe_type, inputs.front().lens()};
+        return {targe_type, inputs.front().lens(), inputs.front().strides()};
     }
 
     argument compute(const shape& output_shape, std::vector<argument> args) const

src/quantize_ins.cpp

@@ -90,10 +90,8 @@ void quantize_ins(program& prog, const std::vector<std::string>& ins_names)
         if(inputs != converted_inputs)
         {
             auto op = ins->get_operator();
-            instruction::replace(ins, op, compute_shape(op, converted_inputs), converted_inputs);
-        }
-
-        if(ins->get_shape().type() != orig_type)
+            auto ins_shape = compute_shape(op, converted_inputs);
+            if (ins_shape.type() != orig_type)
             {
                 // insert another fp_conversion instruction to convert it back
                 if(ins == std::prev(prog.end()))
@@ -107,6 +105,9 @@ void quantize_ins(program& prog, const std::vector<std::string>& ins_names)
                     prog.replace_instruction(ins, ins_orig_type);
                 }
             }
+
+            instruction::replace(ins, op, compute_shape(op, converted_inputs), converted_inputs);
+        }
     }
 }
 

src/targets/gpu/fp_conversion.cpp

@@ -9,7 +9,7 @@ namespace gpu {
 shape hip_fp_conversion::compute_shape(std::vector<shape> inputs) const
 {
     inputs.pop_back();
-    check_shapes{inputs}.not_broadcasted().not_transposed();
+    check_shapes{inputs}.packed();
     return op.compute_shape(inputs);
 }