Merge branch 'improve-gpu-sched' into jit-layernorm-merge

.github/workflows/performance.yaml

 name: MIGraphX Performance Tests
 
 on:
-  push:
-    branches: [develop]
-
   pull_request:
     branches: [develop]
+    types: [opened, synchronize, closed]
   schedule:
     - cron: "0 5 * * 1-6"
 

src/py/migraphx_py.cpp

@@ -40,6 +40,7 @@
 #include <migraphx/register_target.hpp>
 #include <migraphx/json.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/op/common.hpp>
 
 #ifdef HAVE_GPU
 #include <migraphx/gpu/hip.hpp>
@@ -82,7 +83,7 @@ void visit_py(T x, F f)
     {
         f(x.template cast<bool>());
     }
-    else if(py::isinstance<py::int_>(x))
+    else if(py::isinstance<py::int_>(x) || py::hasattr(x, "__index__"))
     {
         f(x.template cast<int>());
     }
@@ -324,6 +325,7 @@ MIGRAPHX_PYBIND11_MODULE(migraphx, m)
         .def("get_parameter_names", &migraphx::program::get_parameter_names)
         .def("get_parameter_shapes", &migraphx::program::get_parameter_shapes)
         .def("get_output_shapes", &migraphx::program::get_output_shapes)
+        .def("is_compiled", &migraphx::program::is_compiled)
         .def(
             "compile",
             [](migraphx::program& p, const migraphx::target& t, bool offload_copy, bool fast_math) {
@@ -358,8 +360,8 @@ MIGRAPHX_PYBIND11_MODULE(migraphx, m)
         .def("__ne__", std::not_equal_to<migraphx::program>{})
         .def("__repr__", [](const migraphx::program& p) { return migraphx::to_string(p); });
 
-    py::class_<migraphx::operation>(m, "op")
-        .def(py::init([](const std::string& name, py::kwargs kwargs) {
+    py::class_<migraphx::operation> op(m, "op");
+    op.def(py::init([](const std::string& name, py::kwargs kwargs) {
           migraphx::value v = migraphx::value::object{};
           if(kwargs)
           {
@@ -367,9 +369,26 @@ MIGRAPHX_PYBIND11_MODULE(migraphx, m)
           }
           return migraphx::make_op(name, v);
       }))
-
         .def("name", &migraphx::operation::name);
 
+    py::enum_<migraphx::op::pooling_mode>(op, "pooling_mode")
+        .value("average", migraphx::op::pooling_mode::average)
+        .value("max", migraphx::op::pooling_mode::max)
+        .value("lpnorm", migraphx::op::pooling_mode::lpnorm);
+
+    py::enum_<migraphx::op::rnn_direction>(op, "rnn_direction")
+        .value("forward", migraphx::op::rnn_direction::forward)
+        .value("reverse", migraphx::op::rnn_direction::reverse)
+        .value("bidirectional", migraphx::op::rnn_direction::bidirectional);
+
+    m.def(
+        "argument_from_pointer",
+        [](const migraphx::shape shape, const int64_t address) {
+            return migraphx::argument(shape, reinterpret_cast<void*>(address));
+        },
+        py::arg("shape"),
+        py::arg("address"));
+
     m.def(
         "parse_tf",
         [](const std::string& filename,

src/targets/gpu/compile_hip_code_object.cpp

@@ -138,16 +138,16 @@ compute_global_for(context& ctx, std::size_t n, std::size_t over)
         std::size_t groups     = (n + local - 1) / local;
         std::size_t max_blocks = max_global / local;
         std::size_t nglobal    = std::min(max_blocks * over, groups) * local;
-        return nglobal;
+        return std::min(nglobal, n);
     };
 }
 
 std::size_t compute_block_size(std::size_t n, std::size_t max_block_size)
 {
-    size_t block_size = 128;
-    while(block_size <= max_block_size and block_size <= n)
-        block_size *= 2;
-    return block_size / 2;
+    const std::size_t min_block_size  = 64;
+    const std::size_t base_block_size = 32;
+    auto block_size                   = (((n - 1) / base_block_size + 1)) * base_block_size;
+    return std::min(std::max(min_block_size, block_size), max_block_size);
 }
 
 operation compile_hip_code_object(const std::string& content, hip_compile_options options)