Merge branch 'perk-kernel' into bert-opt2

src/targets/gpu/code_object_op.cpp

@@ -51,7 +51,8 @@ code_object_op::compute(context& ctx, const shape&, const std::vector<argument>&
     std::vector<void*> kargs(args.size());
     std::transform(
         args.begin(), args.end(), kargs.begin(), [](const argument& a) { return a.data(); });
-    k.launch(ctx.get_stream().get(), global, local, std::move(kargs));
+    auto [start, stop] = ctx.get_perf_events();
+    k.launch(ctx.get_stream().get(), global, local, std::move(kargs), start, stop);
     return args[get_output_arg(args.size())];
 }
 void code_object_op::finalize(context&, const shape&, const std::vector<shape>&)

src/targets/gpu/driver/compile_op.cpp

@@ -38,8 +38,11 @@ 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);
-        double t    = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << t << "ms" << std::endl;
+        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";
+        std::cout << std::endl;
     }
 };
 

src/targets/gpu/driver/include/migraphx/gpu/driver/perf.hpp

@@ -33,7 +33,8 @@ inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 namespace driver {
 
-double time_op(context& ctx, operation op, const std::vector<shape>& inputs, int n = 100);
+std::pair<double, double>
+time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n = 100);
 
 } // namespace driver
 } // namespace gpu

src/targets/gpu/driver/perf.cpp

@@ -42,22 +42,31 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
 }
 
 using milliseconds = std::chrono::duration<double, std::milli>;
-double time_op(context& ctx, operation op, const std::vector<shape>& inputs, int n)
+std::pair<double, double>
+time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
 {
+
     // TODO: Use std::ref
-    migraphx::context gctx = ctx;
-    auto output            = op.compute_shape(inputs);
-    op.finalize(gctx, output, inputs);
+    migraphx::context ctx = ictx;
+    auto& gctx            = any_cast<migraphx::gpu::context>(ctx);
+    auto output           = op.compute_shape(inputs);
+    op.finalize(ctx, output, inputs);
     auto args = generate_arguments(inputs);
     auto run  = [&] {
-        op.compute(gctx, output, args);
-        gctx.finish();
+        op.compute(ctx, output, args);
+        ctx.finish();
     };
+    gctx.enable_perf_measurement();
     run();
-    auto r   = range(n);
-    double t = std::accumulate(
-        r.begin(), r.end(), double{0.0}, [&](auto x, auto) { return x + time<milliseconds>(run); });
-    return t / n;
+    double host_time   = 0.0;
+    double device_time = 0.0;
+    for(auto i : range(n))
+    {
+        (void)i;
+        host_time += time<milliseconds>(run);
+        device_time += gctx.get_elapsed_ms();
+    }
+    return std::make_pair(host_time / n, device_time / n);
 }
 
 } // namespace driver

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"));
-        double t = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
-        std::cout << op << ": " << t << "ms" << std::endl;
+        auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
+        std::cout << op << ": " << host_time << "ms" << std::endl;
     }
 };
 

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

@@ -244,6 +244,15 @@ struct context
         return hip_event_ptr{event};
     }
 
+    static hip_event_ptr create_event_for_timing()
+    {
+        hipEvent_t event;
+        auto status = hipEventCreate(&event);
+        if(status != hipSuccess)
+            MIGRAPHX_THROW("Failed to create event");
+        return hip_event_ptr{event};
+    }
+
     value to_value() const
     {
         value result;
@@ -267,10 +276,49 @@ 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)
+            return std::make_pair(start_event.get(), stop_event.get());
+        return std::make_pair(nullptr, nullptr);
+    }
+
+    float get_elapsed_ms() const
+    {
+        float result = 0;
+        if(start_event != nullptr and stop_event != nullptr)
+        {
+            auto status = hipEventElapsedTime(&result, start_event.get(), stop_event.get());
+            if(status != hipSuccess)
+                MIGRAPHX_THROW("Failed hipEventElapsedTime: " + hip_error(status));
+        }
+        return result;
+    }
+
     private:
     // TODO: Make this a vector to support multiple devices
     std::shared_ptr<hip_device> current_device;
     std::vector<shared<hip_event_ptr>> events;
+    bool measure_perf                 = false;
+    shared<hip_event_ptr> start_event = nullptr;
+    shared<hip_event_ptr> stop_event  = nullptr;
 };
 
 inline void migraphx_to_value(value& v, const context& ctx) { v = ctx.to_value(); }

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

@@ -37,6 +37,8 @@ namespace gpu {
 
 struct context;
 
+std::string hip_error(int error);
+
 argument allocate_gpu(const shape& s, bool host = false);
 
 argument register_on_gpu(const argument& arg);

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

@@ -50,17 +50,22 @@ struct kernel
     void launch(hipStream_t stream,
                 std::size_t global,
                 std::size_t local,
-                const std::vector<kernel_argument>& args) const;
+                const std::vector<kernel_argument>& args,
+                hipEvent_t start = nullptr,
+                hipEvent_t stop  = nullptr) const;
 
     void launch(hipStream_t stream,
                 std::size_t global,
                 std::size_t local,
-                std::vector<void*> args) const;
+                std::vector<void*> args,
+                hipEvent_t start = nullptr,
+                hipEvent_t stop  = nullptr) const;
 
-    auto launch(hipStream_t stream, std::size_t global, std::size_t local) const
+    template <class... Ts>
+    auto launch(hipStream_t stream, std::size_t global, std::size_t local, Ts... zs) const
     {
         return [=](auto&&... xs) {
-            launch(stream, global, local, std::vector<kernel_argument>{xs...});
+            launch(stream, global, local, std::vector<kernel_argument>{xs...}, zs...);
         };
     }
 

src/targets/gpu/kernel.cpp

@@ -80,7 +80,9 @@ void launch_kernel(hipFunction_t fun,
                    std::size_t global,
                    std::size_t local,
                    void* kernargs,
-                   std::size_t size)
+                   std::size_t size,
+                   hipEvent_t start,
+                   hipEvent_t stop)
 {
     assert(global > 0);
     assert(local > 0);
@@ -97,34 +99,55 @@ void launch_kernel(hipFunction_t fun,
 #endif
     };
 
-    auto status = hipExtModuleLaunchKernel(
-        fun, global, 1, 1, local, 1, 1, 0, stream, nullptr, reinterpret_cast<void**>(&config));
+    auto status = hipExtModuleLaunchKernel(fun,
+                                           global,
+                                           1,
+                                           1,
+                                           local,
+                                           1,
+                                           1,
+                                           0,
+                                           stream,
+                                           nullptr,
+                                           reinterpret_cast<void**>(&config),
+                                           start,
+                                           stop);
     if(status != hipSuccess)
         MIGRAPHX_THROW("Failed to launch kernel: " + hip_error(status));
+    if(stop)
+    {
+        status = hipEventSynchronize(stop);
+        if(status != hipSuccess)
+            MIGRAPHX_THROW("Failed to sync event: " + hip_error(status));
+    }
 }
 
 void kernel::launch(hipStream_t stream,
                     std::size_t global,
                     std::size_t local,
-                    std::vector<void*> args) const
+                    std::vector<void*> args,
+                    hipEvent_t start,
+                    hipEvent_t stop) const
 {
     assert(impl != nullptr);
     void* kernargs   = args.data();
     std::size_t size = args.size() * sizeof(void*);
 
-    launch_kernel(impl->fun, stream, global, local, kernargs, size);
+    launch_kernel(impl->fun, stream, global, local, kernargs, size, start, stop);
 }
 
 void kernel::launch(hipStream_t stream,
                     std::size_t global,
                     std::size_t local,
-                    const std::vector<kernel_argument>& args) const
+                    const std::vector<kernel_argument>& args,
+                    hipEvent_t start,
+                    hipEvent_t stop) const
 {
     assert(impl != nullptr);
     std::vector<char> kernargs = pack_args(args);
     std::size_t size           = kernargs.size();
 
-    launch_kernel(impl->fun, stream, global, local, kernargs.data(), size);
+    launch_kernel(impl->fun, stream, global, local, kernargs.data(), size, start, stop);
 }
 
 } // namespace gpu