Fix TRACE_EVAL > 1 (#1835)

* add fix for the trace_eval

* Add throw for the debug builds

* Formatting

---------
Co-authored-by: Chris Austen <causten@users.noreply.github.com>

src/program.cpp

@@ -539,7 +539,8 @@ std::vector<argument> program::eval(parameter_map params, execution_environment
             ins_out[x] = ss.str();
         });
 
-        ret = generic_eval(*this,
+        ret = generic_eval(
+            *this,
             ctx,
             std::move(params),
             with_check_context([&](auto& ins, auto f, auto&& check_context) {
@@ -551,15 +552,28 @@ std::vector<argument> program::eval(parameter_map params, execution_environment
                 ctx.finish();
                 double t2 = t.record<milliseconds>();
                 std::cout << "Time: " << t1 << "ms, " << t2 << "ms" << std::endl;
-                               if(trace_level > 1 and ins->name().front() != '@' and
-                                  ins->name() != "load" and not result.empty())
+                if(trace_level > 1 and ins->name().front() != '@' and ins->name() != "load" and
+                   not result.empty())
+                {
+                    migraphx::argument buffer;
+                    try
                     {
                         target tgt = make_target(this->impl->target_name);
-                                   auto buffer = tgt.copy_from(result);
+                        buffer     = tgt.copy_from(result);
+                    }
+                    catch(const migraphx::exception&)
+                    {
+                        // instruction was run on host then no need to copy buffer from target
+                        buffer = result;
+                    }
+                    catch(...)
+                    {
+                        MIGRAPHX_THROW(
+                            "MIGraphX program execution with MIGRAPHX_TRACE_EVAL failed.\n");
+                    }
                     if(trace_level == 2)
                     {
-                                       std::cout << "Output has "
-                                                 << to_string_range(classify_argument(buffer))
+                        std::cout << "Output has " << to_string_range(classify_argument(buffer))
                                   << std::endl;
                         std::cout << "Output: ";
                         preview_argument(std::cout, buffer);

src/targets/gpu/hip.cpp

@@ -146,7 +146,11 @@ std::vector<T> read_from_gpu(const void* x, std::size_t sz)
     gpu_sync();
     std::vector<T> result(sz);
     assert(not is_device_ptr(result.data()));
-    assert(is_device_ptr(x));
+    if(not is_device_ptr(x))
+    {
+        MIGRAPHX_THROW(
+            "read_from_gpu() requires Src buffer to be on the GPU, Copy from gpu failed\n");
+    }
     auto status = hipMemcpy(result.data(), x, sz * sizeof(T), hipMemcpyDeviceToHost);
     if(status != hipSuccess)
         MIGRAPHX_THROW("Copy from gpu failed: " + hip_error(status)); // NOLINT