Check for flag

src/targets/gpu/include/migraphx/gpu/quant_convolution.hpp

@@ -24,8 +24,7 @@ struct miopen_quant_convolution
     {
         // TODO: Add algo
         // return op::quant_convolution::reflect(self.op, f);
-        return pack(f(self.op, "op"),
-                    f(self.int8_x4_format, "int8_x4_format"));
+        return pack(f(self.op, "op"), f(self.int8_x4_format, "int8_x4_format"));
     }
 
     std::string name() const { return "gpu::quant_convolution"; }

src/targets/gpu/pack_int8_args.cpp

@@ -156,6 +156,7 @@ void pack_int8_args::apply(module& m) const
         }
         else if(ins->name() == "gpu::quant_convolution")
         {
+            // TODO: Check for int8_x4_format
             auto inputs   = ins->inputs();
             auto packed_x = m.insert_instruction(
                 ins,

src/targets/gpu/quant_convolution.cpp

@@ -16,8 +16,8 @@ argument miopen_quant_convolution::compute(context& ctx,
                                            const shape& output_shape,
                                            const std::vector<argument>& args) const
 {
-    auto x_desc = make_tensor(args[0].get_shape(), true);
-    auto w_desc = make_tensor(args[1].get_shape(), true);
+    auto x_desc = make_tensor(args[0].get_shape(), int8_x4_format);
+    auto w_desc = make_tensor(args[1].get_shape(), int8_x4_format);
     auto y_desc = make_tensor(output_shape);
 
     float alpha = 1;
@@ -49,8 +49,8 @@ shape miopen_quant_convolution::compile(context& ctx,
                                         std::vector<shape> inputs)
 {
     shape workspace_shape{};
-    auto x_desc = make_tensor(inputs[0], true);
-    auto w_desc = make_tensor(inputs[1], true);
+    auto x_desc = make_tensor(inputs[0], int8_x4_format);
+    auto w_desc = make_tensor(inputs[1], int8_x4_format);
     auto y_desc = make_tensor(output_shape);
 
     std::size_t workspace_size = 0;
@@ -62,6 +62,7 @@ shape miopen_quant_convolution::compile(context& ctx,
                                              &workspace_size);
     workspace_shape = shape{shape::int8_type, {workspace_size}};
 
+    // TODO: use pack_int8_shape only when pack_int8_shape is true
     auto arg_vec4_x = to_gpu(generate_argument(pack_int8_shape(inputs[0])));
     auto arg_vec4_w = to_gpu(generate_argument(pack_int8_shape(inputs[1])));
     auto y          = allocate_gpu(output_shape);