Update time op to more accurately get device time (#2104)

src/targets/gpu/compile_ops.cpp

@@ -185,8 +185,7 @@ struct compile_plan
             results.begin(), results.end(), std::back_inserter(times), [&](const auto& cr) {
                 if(not cr.has_value())
                     return std::numeric_limits<double>::max();
-                return time_op(*ctx, cr->replace.code_object, to_shapes(cr->ins->inputs()), 20)
-                    .first;
+                return time_op(*ctx, cr->replace.code_object, to_shapes(cr->ins->inputs()), 20);
             });
         auto i = std::distance(times.begin(), std::min_element(times.begin(), times.end()));
         std::cout << "Fastest solution: " << config->solutions.at(i) << std::endl;

src/targets/gpu/driver/compile_op.cpp

@@ -38,10 +38,8 @@ struct compile_op : action<compile_op>
         context ctx;
         auto inputs = p.parse_shapes(v.at("inputs"));
         auto op     = gpu::compile_op(v.at("name").to<std::string>(), ctx, inputs, v);
-        auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << host_time << "ms";
-        if(device_time > 0)
-            std::cout << ", " << device_time << "ms";
+        auto t      = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
+        std::cout << op << ": " << t << "ms";
         std::cout << std::endl;
     }
 };

src/targets/gpu/driver/run_op.cpp

@@ -43,8 +43,8 @@ struct run_op : action<run_op>
         auto op = make_op(name);
         if(v.contains("fields"))
             op.from_value(v.at("fields"));
-        auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << host_time << "ms" << std::endl;
+        auto t = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
+        std::cout << op << ": " << t << "ms" << std::endl;
     }
 };
 

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

@@ -299,23 +299,6 @@ struct context
 
     any_ptr get_queue() { return get_stream().get(); }
 
-    void enable_perf_measurement(bool b = true)
-    {
-        if(b)
-        {
-            start_event = create_event_for_timing();
-            stop_event  = create_event_for_timing();
-            get_stream().record(start_event.get());
-            get_stream().record(stop_event.get());
-        }
-        else
-        {
-            start_event = nullptr;
-            stop_event  = nullptr;
-        }
-        measure_perf = b;
-    }
-
     std::pair<hipEvent_t, hipEvent_t> get_perf_events() const
     {
         if(measure_perf)
@@ -323,12 +306,12 @@ struct context
         return std::make_pair(nullptr, nullptr);
     }
 
-    float get_elapsed_ms() const
+    static float get_elapsed_ms(hipEvent_t start, hipEvent_t stop)
     {
         float result = 0;
-        if(start_event != nullptr and stop_event != nullptr)
+        if(start != nullptr and stop != nullptr)
         {
-            auto status = hipEventElapsedTime(&result, start_event.get(), stop_event.get());
+            auto status = hipEventElapsedTime(&result, start, stop);
             if(status != hipSuccess)
                 MIGRAPHX_THROW("Failed hipEventElapsedTime: " + hip_error(status));
         }

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

@@ -32,7 +32,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPHX_GPU_EXPORT std::pair<double, double>
+MIGRAPHX_GPU_EXPORT double
 time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n = 100);
 
 } // namespace gpu

src/targets/gpu/time_op.cpp

@@ -41,8 +41,7 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
 }
 
 using milliseconds = std::chrono::duration<double, std::milli>;
-std::pair<double, double>
-time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
+double time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
 {
 
     // TODO: Use std::ref
@@ -51,21 +50,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
     auto output           = op.compute_shape(inputs);
     op.finalize(ctx, output, inputs);
     auto args = generate_arguments(inputs);
-    auto run  = [&] {
-        op.compute(ctx, output, args);
-        ctx.finish();
-    };
-    gctx.enable_perf_measurement();
+    auto start = context::create_event_for_timing();
+    auto stop  = context::create_event_for_timing();
+    auto run   = [&] { op.compute(ctx, output, args); };
     run();
-    double host_time   = 0.0;
-    double device_time = 0.0;
+    gctx.get_stream().record(start.get());
     for(auto i : range(n))
     {
         (void)i;
-        host_time += time<milliseconds>(run);
-        device_time += gctx.get_elapsed_ms();
+        run();
     }
-    return std::make_pair(host_time / n, device_time / n);
+    gctx.get_stream().record(stop.get());
+    gctx.finish();
+    return context::get_elapsed_ms(start.get(), stop.get()) / n;
 }
 
 } // namespace gpu