Merge branch 'jit-layernorm-merge' of...

Merge branch 'jit-layernorm-merge' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into layernorm_eps

CMakeLists.txt

@@ -63,7 +63,7 @@ set(CMAKE_EXTRA_INCLUDE_FILES)
 
 include(ROCMSetupVersion)
 
-rocm_setup_version(VERSION 2.3)
+rocm_setup_version(VERSION 2.4)
 set(MIGRAPHX_SO_VERSION ${PROJECT_VERSION_MAJOR}.${PROJECT_VERSION_MINOR})
 
 option( BUILD_SHARED_LIBS "Build as a shared library" ON )

src/include/migraphx/op/fmod.hpp

@@ -24,17 +24,8 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_FMOD_HPP
 #define MIGRAPHX_GUARD_OPERATORS_FMOD_HPP
 
-#include <array>
 #include <migraphx/op/binary.hpp>
-#include <migraphx/check_shapes.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/streamutils.hpp>
-#include <migraphx/literal.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
 #include <cmath>
-#include <utility>
-#include <type_traits>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/include/migraphx/op/mod.hpp

@@ -24,17 +24,8 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_MOD_HPP
 #define MIGRAPHX_GUARD_OPERATORS_MOD_HPP
 
-#include <array>
 #include <migraphx/op/binary.hpp>
-#include <migraphx/check_shapes.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/streamutils.hpp>
-#include <migraphx/literal.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
 #include <cmath>
-#include <utility>
-#include <type_traits>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/fuse_mlir.cpp

@@ -61,13 +61,25 @@ struct mlir_conv
 MIGRAPHX_REGISTER_OP(mlir_conv);
 
 namespace {
+
+MIGRAPHX_PRED_MATCHER(is_mlir_conv, instruction_ref ins)
+{
+    if(ins->name() != "convolution")
+        return false;
+    value v    = ins->get_operator().to_value();
+    auto group = v.at("group").to<int>();
+    if(group != 1)
+        return false;
+    return true;
+}
+
 struct find_conv_pointwise
 {
     // Find a convolution followed by a pointwise operation.
     auto matcher() const
     {
         auto convolution =
-            match::skip(match::name("contiguous"))(match::name("convolution").bind("convolution"));
+            match::skip(match::name("contiguous"))(is_mlir_conv().bind("convolution"));
         return match::name("pointwise")(match::any_of[match::inputs()](convolution.bind("x")));
     }
 

src/targets/gpu/fuse_ops.cpp

@@ -26,7 +26,6 @@
 #include <migraphx/gpu/fuse_ops.hpp>
 #include <migraphx/matcher.hpp>
 #include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/convolution.hpp>
 #include <migraphx/gpu/device_name.hpp>
 #include <migraphx/gpu/oper.hpp>
@@ -50,8 +49,6 @@
 #include <migraphx/array.hpp>
 #include <migraphx/permutation.hpp>
 #include <migraphx/make_op.hpp>
-#include <migraphx/op/clip.hpp>
-#include <migraphx/op/contiguous.hpp>
 #include <cmath>
 #include <set>
 
@@ -1036,7 +1033,7 @@ struct find_gemm_pointwise
         // const-fold input if not standard shape since rocblas can't handle it
         if(not c_ins->get_shape().standard())
         {
-            auto c = op::contiguous{};
+            auto c = make_op("contiguous");
             auto l = c.compute(c.compute_shape({c_ins->get_shape()}), {c_ins->eval()});
             c_ins  = m.add_literal(l.get_shape(), l.data());
         }

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

@@ -27,7 +27,7 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/reflect.hpp>
 #include <migraphx/op/gather.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/include/migraphx/gpu/int8_conv_pack.hpp

@@ -25,7 +25,6 @@
 #define MIGRAPHX_GUARD_RTGLIB_INT8_CONV_PACK_HPP
 
 #include <migraphx/argument.hpp>
-#include <migraphx/op/quant_dot.hpp>
 #include <migraphx/config.hpp>
 #include <utility>
 

src/targets/gpu/include/migraphx/gpu/int8_gemm_pack.hpp

@@ -25,7 +25,6 @@
 #define MIGRAPHX_GUARD_RTGLIB_INT8_GEMM_PACK_HPP
 
 #include <migraphx/argument.hpp>
-#include <migraphx/op/quant_dot.hpp>
 #include <migraphx/config.hpp>
 #include <utility>
 

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

@@ -24,22 +24,10 @@
 #ifndef MIGRAPHX_GUARD_RTGLIB_LOGSOFTMAX_HPP
 #define MIGRAPHX_GUARD_RTGLIB_LOGSOFTMAX_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
 #include <migraphx/op/logsoftmax.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
+#include <migraphx/shape.hpp>
+#include <migraphx/reflect.hpp>
 #include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

@@ -26,7 +26,7 @@
 
 #include <migraphx/shape.hpp>
 #include <migraphx/reflect.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

@@ -25,7 +25,7 @@
 #define MIGRAPHX_GUARD_GPU_PREFUSE_OPS_HPP
 
 #include <migraphx/config.hpp>
-#include <migraphx/gpu/context.hpp>
+#include <string>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

@@ -27,7 +27,7 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/reflect.hpp>
 #include <migraphx/op/reverse.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

@@ -24,22 +24,10 @@
 #ifndef MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
 #define MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
 #include <migraphx/op/softmax.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
+#include <migraphx/shape.hpp>
+#include <migraphx/reflect.hpp>
 #include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

@@ -25,8 +25,6 @@
 #define MIGRAPHX_GUARD_RTGLIB_GPU_SYNC_DEVICE_HPP
 
 #include <string>
-#include <migraphx/instruction_ref.hpp>
-#include <migraphx/gpu/context.hpp>
 #include <migraphx/config.hpp>
 
 namespace migraphx {

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

@@ -27,7 +27,7 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/reflect.hpp>
 #include <migraphx/op/topk.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/jit/gathernd.cpp

@@ -27,13 +27,6 @@
 
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compile_hip.hpp>
-#include <migraphx/ranges.hpp>
-#include <migraphx/reduce_dims.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/eliminate_common_subexpression.hpp>
-#include <migraphx/module.hpp>
-#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/jit/layernorm.cpp

@@ -27,8 +27,6 @@
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/compile_gen.hpp>
 
-#include <migraphx/cpp_generator.hpp>
-#include <migraphx/ranges.hpp>
 #include <migraphx/reduce_dims.hpp>
 #include <migraphx/stringutils.hpp>
 

src/targets/gpu/jit/pointwise.cpp

@@ -26,16 +26,7 @@
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/compile_gen.hpp>
-
-#include <migraphx/cpp_generator.hpp>
-#include <migraphx/ranges.hpp>
 #include <migraphx/reduce_dims.hpp>
-#include <migraphx/permutation.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/eliminate_common_subexpression.hpp>
-#include <migraphx/module.hpp>
-#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/jit/reduce.cpp

@@ -26,15 +26,7 @@
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/compile_gen.hpp>
-
-#include <migraphx/cpp_generator.hpp>
-#include <migraphx/ranges.hpp>
 #include <migraphx/reduce_dims.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/eliminate_common_subexpression.hpp>
-#include <migraphx/module.hpp>
-#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/jit/roialign.cpp

@@ -24,16 +24,7 @@
 #include <migraphx/gpu/compiler.hpp>
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/context.hpp>
-
 #include <migraphx/gpu/compile_hip.hpp>
-#include <migraphx/cpp_generator.hpp>
-#include <migraphx/ranges.hpp>
-#include <migraphx/reduce_dims.hpp>
-#include <migraphx/stringutils.hpp>
-#include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/eliminate_common_subexpression.hpp>
-#include <migraphx/module.hpp>
-#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {