Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into mlir-attention

CMakeLists.txt

@@ -27,20 +27,18 @@ if("${CMAKE_SOURCE_DIR}" STREQUAL "${CMAKE_BINARY_DIR}")
     message(FATAL_ERROR "The binary and source directroy cannot be the same")
 endif()
 
-get_property(_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
+# Setup valid strings for build type
+if (NOT CMAKE_CONFIGURATION_TYPES)
+    set(CMAKE_CONFIGURATION_TYPES "Debug;Release;RelWithDebInfo;MinSizeRel" CACHE STRING "Configs")
+endif()
 
+get_property(MIGRAPHX_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
 # This has to be initialized before the project() command appears
 # Set the default of CMAKE_BUILD_TYPE to be release, unless user specifies with -D.  MSVC_IDE does not use CMAKE_BUILD_TYPE
-if(_GENERATOR_IS_MULTI_CONFIG)
-    if (NOT CMAKE_CONFIGURATION_TYPES)
-        set(CMAKE_CONFIGURATION_TYPES "Debug;Release;RelWithDebInfo;MinSizeRel" CACHE STRING
-            "Available build types (configurations) on multi-config generators")
-    endif()
-else()
-    if(NOT CMAKE_BUILD_TYPE)
-        set(CMAKE_BUILD_TYPE Release CACHE STRING
-            "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel.")
-    endif()
+if(NOT MIGRAPHX_GENERATOR_IS_MULTI_CONFIG)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING
+        "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel.")
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS ${CMAKE_CONFIGURATION_TYPES})
 endif()
 
 set(CMAKE_INSTALL_PREFIX "/opt/rocm" CACHE PATH "")
@@ -78,7 +76,7 @@ include(ROCMSetupVersion)
 option(BUILD_DEV "Build for development purpose only" OFF)
 
 rocm_setup_version(VERSION 2.8.0)
-set(MIGRAPHX_SO_VERSION ${PROJECT_VERSION_MAJOR}.${PROJECT_VERSION_MINOR}.${PROJECT_VERSION_PATCH})
+set(MIGRAPHX_SO_VERSION ${PROJECT_VERSION_MAJOR}.${PROJECT_VERSION_MINOR})
 
 option( BUILD_SHARED_LIBS "Build as a shared library" ON )
 

docs/.sphinx/requirements.txt

@@ -35,7 +35,7 @@ fastjsonschema==2.16.3
     # via rocm-docs-core
 gitdb==4.0.10
     # via gitpython
-gitpython==3.1.32
+gitpython==3.1.37
     # via rocm-docs-core
 idna==3.4
     # via requests
@@ -87,7 +87,7 @@ requests==2.28.2
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==0.24.2
+rocm-docs-core==0.26.0
     # via -r requirements.in
 smmap==5.0.0
     # via gitdb
@@ -130,7 +130,7 @@ sphinxcontrib-serializinghtml==1.1.5
     # via sphinx
 typing-extensions==4.5.0
     # via pydata-sphinx-theme
-urllib3==1.26.15
+urllib3==1.26.18
     # via requests
 wrapt==1.15.0
     # via deprecated

src/driver/argument_parser.hpp

@@ -187,6 +187,13 @@ struct value_parser
     }
 };
 
+// version for std::optional object
+template <class T>
+struct value_parser<std::optional<T>>
+{
+    static T apply(const std::string& x) { return value_parser<T>::apply(x); }
+};
+
 struct argument_parser
 {
     struct argument

src/driver/main.cpp

@@ -540,22 +540,17 @@ struct params : command<params>
 struct verify : command<verify>
 {
     compiler c;
-    // Set to -1. as nonsense initial value
-    double rms_tol       = -1.0;
-    double atol          = -1.0;
-    double rtol          = -1.0;
+    std::optional<double> rms_tol;
+    std::optional<double> atol;
+    std::optional<double> rtol;
     bool per_instruction = false;
     bool reduce          = false;
     void parse(argument_parser& ap)
     {
         c.parse(ap);
-        ap(rms_tol, {"--rms-tol"}, ap.help("Tolerance for the RMS error (Default: 0.001)"));
-        ap(atol,
-           {"--atol"},
-           ap.help("Tolerance for the elementwise absolute difference (Default: 0.001)"));
-        ap(rtol,
-           {"--rtol"},
-           ap.help("Tolerance for the elementwise relative difference (Default: 0.001)"));
+        ap(rms_tol, {"--rms-tol"}, ap.help("Tolerance for the RMS error"));
+        ap(atol, {"--atol"}, ap.help("Tolerance for the elementwise absolute difference"));
+        ap(rtol, {"--rtol"}, ap.help("Tolerance for the elementwise relative difference"));
         ap(per_instruction,
            {"-i", "--per-instruction"},
            ap.help("Verify each instruction"),
@@ -572,33 +567,6 @@ struct verify : command<verify>
         auto t = c.ct.get_target();
         auto m = c.parameters.generate(p, t, true, c.l.batch);
 
-        // TODO remove this and make the driver able to figure out datatype most used in the model
-        //  then set the tolerances appropriately. Need to check here because c.to_fp16 only set
-        //  after argument_parser.parse() is run. This code is complicated because there's not a
-        //  good way to change the default tolerances after reading `--fp16` but before reading
-        //  `--rms-tol`, `--atol`, and `--rtol`.
-        migraphx::verify::tolerance tols{};
-        if(c.to_fp16)
-        {
-            tols = migraphx::verify::tolerance{8e-2, 4e-2, 4e-2};
-        }
-        if(not float_equal(this->rms_tol, -1.0))
-        {
-            tols.rms_tol = this->rms_tol;
-        }
-        if(not float_equal(this->atol, -1.0))
-        {
-            tols.atol = this->atol;
-        }
-        if(not float_equal(this->rtol, -1.0))
-        {
-            tols.rtol = this->rtol;
-        }
-
-        std::cout << "rms_tol: " << tols.rms_tol << std::endl;
-        std::cout << "atol: " << tols.atol << std::endl;
-        std::cout << "rtol: " << tols.rtol << std::endl;
-
         auto quantize = precision::fp32;
         if(c.to_fp16)
         {
@@ -609,6 +577,11 @@ struct verify : command<verify>
             quantize = precision::int8;
         }
 
+        auto tols = get_tolerances(p, quantize, rms_tol, atol, rtol);
+        std::cout << "rms_tol: " << tols.rms_tol << std::endl;
+        std::cout << "atol: " << tols.atol << std::endl;
+        std::cout << "rtol: " << tols.rtol << std::endl;
+
         if(per_instruction)
         {
             verify_instructions(p, t, c.co, quantize, tols);

src/driver/verify.cpp

@@ -36,6 +36,42 @@ namespace migraphx {
 namespace driver {
 inline namespace MIGRAPHX_INLINE_NS {
 
+/**
+ * Gives tolerances based on user input (`rms_tol`, `atol`, `rtol` parameters) and defaults.
+ * Sets to fp16 tolerances if `quantize` input is fp16 or any fp16 instruction in found in the
+ * model.
+ */
+verify::tolerance get_tolerances(const program& p,
+                                 precision quantize,
+                                 std::optional<double> rms_tol,
+                                 std::optional<double> atol,
+                                 std::optional<double> rtol)
+{
+    bool has_fp16 = any_of(p.get_modules(), [](auto&& m) {
+        return any_of(*m, [](auto&& ins) { return (ins.get_shape().type() == shape::half_type); });
+    });
+    migraphx::verify::tolerance result{};
+    if(has_fp16 or quantize == precision::fp16)
+    {
+        result.rms_tol = 8e-2;
+        result.atol    = 4e-2;
+        result.rtol    = 4e-2;
+    }
+    if(rms_tol)
+    {
+        result.rms_tol = *rms_tol;
+    }
+    if(atol)
+    {
+        result.atol = *atol;
+    }
+    if(rtol)
+    {
+        result.rtol = *rtol;
+    }
+    return result;
+}
+
 std::vector<argument> run_ref(program p, const parameter_map& inputs)
 {
     p.compile(migraphx::make_target("ref"));

src/driver/verify.hpp

@@ -32,6 +32,12 @@ namespace migraphx {
 namespace driver {
 inline namespace MIGRAPHX_INLINE_NS {
 
+verify::tolerance get_tolerances(const program& p,
+                                 precision quantize,
+                                 std::optional<double> rms_tol,
+                                 std::optional<double> atol,
+                                 std::optional<double> rtol);
+
 void verify_program(const std::string& name,
                     const program& p,
                     const target& t,

src/include/migraphx/argument.hpp

@@ -46,7 +46,7 @@ struct MIGRAPHX_EXPORT argument : raw_data<argument>
 {
     argument() = default;
 
-    argument(const shape& s);
+    explicit argument(const shape& s);
 
     template <class F, MIGRAPHX_REQUIRES(std::is_pointer<decltype(std::declval<F>()())>{})>
     argument(shape s, F d)

src/include/migraphx/op/allocate.hpp

@@ -88,13 +88,13 @@ struct allocate
     {
         if(args.empty())
         {
-            return {output_shape};
+            return argument{output_shape};
         }
         else
         {
             std::vector<std::size_t> output_dims(output_shape.ndim());
             args.at(0).visit([&](auto a) { output_dims.assign(a.begin(), a.end()); });
-            return {shape{buf_type, output_dims}};
+            return argument{shape{buf_type, output_dims}};
         }
     }
 };

src/include/migraphx/op/argmax.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -31,6 +31,7 @@
 #include <migraphx/value.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
 #include <migraphx/dyn_output.hpp>
+#include <migraphx/float_equal.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -38,12 +39,13 @@ namespace op {
 
 struct argmax
 {
-    int64_t axis = 0;
+    int64_t axis           = 0;
+    bool select_last_index = false;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.axis, "axis"));
+        return pack(f(self.axis, "axis"), f(self.select_last_index, "select_last_index"));
     }
 
     value attributes() const
@@ -87,6 +89,10 @@ struct argmax
                 max_val   = cur_val;
                 max_index = i;
             }
+            else if(select_last_index and float_equal(max_val, cur_val))
+            {
+                max_index = i;
+            }
         }
         return max_index;
     }

src/include/migraphx/op/argmin.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -30,6 +30,7 @@
 #include <migraphx/config.hpp>
 #include <migraphx/value.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
+#include <migraphx/float_equal.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -38,11 +39,12 @@ namespace op {
 struct argmin
 {
     int64_t axis = 0;
+    bool select_last_index = false;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.axis, "axis"));
+        return pack(f(self.axis, "axis"), f(self.select_last_index, "select_last_index"));
     }
 
     value attributes() const
@@ -78,6 +80,10 @@ struct argmin
                 min_val   = cur_val;
                 min_index = i;
             }
+            else if(select_last_index and float_equal(min_val, cur_val))
+            {
+                min_index = i;
+            }
         }
 
         return min_index;

src/include/migraphx/op/pooling.hpp

@@ -411,7 +411,7 @@ struct pooling
             // for dynamic GlobalPooling, there's no padding
             kernel_dims.insert(kernel_dims.end(), input_lens.begin() + 2, input_lens.end());
             output_shape = dyn_out.computed_shape;
-            result       = dyn_out.computed_shape;
+            result       = argument{dyn_out.computed_shape};
         }
         else if((padding_mode != op::padding_mode_t::default_))
         {
@@ -439,7 +439,7 @@ struct pooling
         {
             kernel_dims  = this->lengths;
             output_shape = dyn_out.computed_shape;
-            result       = dyn_out.computed_shape;
+            result       = argument{dyn_out.computed_shape};
         }
 
         // Perform the computation and populate result

src/include/migraphx/op/reshape.hpp

@@ -36,6 +36,22 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * 1 input version:
+ * reshape(input_data)
+ * this.dims = output_dims
+ * Makes a copy of input_data to the output shape.
+ *
+ * 2 input version:
+ * reshape(input_data, output_buffer)
+ * this.dims = unset
+ * Copies input_data to output_buffer; output_buffer already has the output shape.
+ * This version will not fail gracefully if the input shape and output_buffer shape are
+ * incompatible. There's a throw that will catch when the number of elements do not match at
+ * runtime. This version should only be used for dynamic reshapes (output dimensions only known at
+ * runtime). If output_buffer has a static shape during compile/parse, you can use the 1 input
+ * version.
+ */
 struct reshape
 {
     std::vector<int64_t> dims;
@@ -215,32 +231,56 @@ struct reshape
 
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this, true}.has(1);
+        check_shapes{inputs, *this, true}.has(1, 2);
 
         auto n_neg_dims = std::count(dims.begin(), dims.end(), -1);
         if(n_neg_dims > 1)
             MIGRAPHX_THROW("reshape: Dimensions for reshape can only have one -1 dim");
 
         auto s0 = inputs.front();
-        if(s0.dynamic())
+        if(inputs.size() == 1)
         {
-            return dyn_compute_shape(s0);
+            if(s0.dynamic())
+            {
+                return dyn_compute_shape(s0);
+            }
+            else
+            {
+                return static_compute_shape(inputs, n_neg_dims);
+            }
         }
         else
         {
-            return static_compute_shape(inputs, n_neg_dims);
+            return inputs.back();
         }
     }
 
     argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
         assert(dyn_out.computed_shape.standard());
-        argument result{dyn_out.computed_shape};
+        if(args.size() == 1)
+        {
+            argument result{dyn_out.computed_shape};
 
-        visit_all(result, args[0])([&](auto output, auto input) {
-            std::copy(input.begin(), input.end(), output.begin());
-        });
-        return result;
+            visit_all(result, args[0])([&](auto output, auto input) {
+                std::copy(input.begin(), input.end(), output.begin());
+            });
+            return result;
+        }
+        else
+        {
+            // 2 arg
+            if(args[0].get_shape().elements() != args[1].get_shape().elements())
+            {
+                MIGRAPHX_THROW("Reshape: Number of elements must match at runtime. Input: " +
+                               std::to_string(args[0].get_shape().elements()) +
+                               " Output buffer: " + std::to_string(args[1].get_shape().elements()));
+            }
+            visit_all(args[1], args[0])([&](auto output, auto input) {
+                std::copy(input.begin(), input.end(), output.begin());
+            });
+            return args[1];
+        }
     }
 };
 

src/onnx/parse_arg_op.cpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -50,14 +50,25 @@ struct parse_arg_op : op_parser<parse_arg_op>
             keep_dims = parser.parse_value(info.attributes.at("keepdims")).at<int>();
         }
 
+        bool select_last_index = false;
+        if(contains(info.attributes, "select_last_index"))
+        {
+            select_last_index = static_cast<bool>(
+                parser.parse_value(info.attributes.at("select_last_index")).at<int>());
+        }
+
         if(keep_dims == 0)
         {
-            auto ins = info.add_instruction(make_op(opd.op_name, {{"axis", axis}}), args);
+            auto ins = info.add_instruction(
+                make_op(opd.op_name, {{"axis", axis}, {"select_last_index", select_last_index}}),
+                args);
             return info.add_instruction(make_op("squeeze", {{"axes", {axis}}}), ins);
         }
         else
         {
-            return info.add_instruction(make_op(opd.op_name, {{"axis", axis}}), args);
+            return info.add_instruction(
+                make_op(opd.op_name, {{"axis", axis}, {"select_last_index", select_last_index}}),
+                args);
         }
     }
 };

src/onnx/parse_groupnorm.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_groupnorm : op_parser<parse_groupnorm>
+{
+    std::vector<op_desc> operators() const { return {{"GroupNormalization"}}; }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          std::vector<instruction_ref> args) const
+    {
+        float epsilon = 1e-5f;
+        if(contains(info.attributes, "epsilon"))
+        {
+            epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
+        }
+        size_t num_groups;
+        if(contains(info.attributes, "num_groups"))
+        {
+            num_groups = parser.parse_value(info.attributes.at("num_groups")).at<size_t>();
+        }
+        else
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: num_groups must be available");
+        }
+
+        if(args.size() != 3)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: invalid input count");
+        }
+
+        auto x     = args.at(0);
+        auto scale = args.at(1);
+        auto bias  = args.at(2);
+
+        auto x_shape = x->get_shape();
+        auto x_dtype = x_shape.type();
+        auto x_dims  = x_shape.lens();
+
+        if(x_shape.ndim() <= 2)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: invalid input shape");
+        }
+
+        auto c = x_shape.lens().at(1);
+        if(c % num_groups != 0)
+        {
+            MIGRAPHX_THROW(
+                "PARSE_GROUPNORM: num_groups should be a divisor of the number of channels");
+        }
+        auto group_size = c / num_groups;
+        if(scale->get_shape().ndim() != 1 or scale->get_shape().lens().at(0) != num_groups)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: scale tensor shape should be num_groups");
+        }
+        if(bias->get_shape().ndim() != 1 or bias->get_shape().lens().at(0) != num_groups)
+        {
+            MIGRAPHX_THROW("PARSE_GROUPNORM: bias tensor shape should be num_groups");
+        }
+
+        // Original shape: N x C x D1 x ... x Dn
+        // New shape: N x num_groups x C // num_groups x D1 x ... x Dn
+
+        std::vector<size_t> dims = {x_dims.at(0), num_groups, group_size};
+        std::copy(x_dims.begin() + 2, x_dims.end(), std::back_inserter(dims));
+        auto x_reshaped = info.add_instruction(make_op("reshape", {{"dims", dims}}), x);
+
+        // Axes for D1 x ... x Dn
+        std::vector<size_t> axes(dims.size() - 2);
+        std::iota(axes.begin(), axes.end(), 2);
+
+        // y = (x - mean) * rsqrt(variance + epsilon) * scale + bias
+        // mean = reduce_mean({D1, D2, ... Dk}, x)
+        // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
+
+        auto mean = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_reshaped);
+        auto x_sub_mean    = info.add_common_op("sub", x_reshaped, mean);
+        auto x_sqdiff_mean = info.add_common_op("sqdiff", x_reshaped, mean);
+        auto variance =
+            info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_sqdiff_mean);
+        epsilon =
+            (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon;
+        auto eps     = info.add_literal(migraphx::literal{migraphx::shape{x_dtype}, {epsilon}});
+        auto var_eps = info.add_common_op("add", variance, eps);
+        auto rsqrt   = info.add_instruction(make_op("rsqrt"), var_eps);
+        auto result  = info.add_common_op("mul", x_sub_mean, rsqrt);
+        auto scale_bcast =
+            info.add_instruction(make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), scale);
+        auto bias_bcast =
+            info.add_instruction(make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), bias);
+        auto scaled     = info.add_instruction(make_op("mul"), result, scale_bcast);
+        auto y          = info.add_instruction(make_op("add"), scaled, bias_bcast);
+        auto y_reshaped = info.add_instruction(make_op("reshape", {{"dims", x_dims}}), y);
+        return y_reshaped;
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_layernorm.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_layernorm : op_parser<parse_layernorm>
+{
+    std::vector<op_desc> operators() const { return {{"LayerNormalization"}}; }
+
+    std::vector<instruction_ref> parse(const op_desc& /*opd*/,
+                                       const onnx_parser& parser,
+                                       const onnx_parser::node_info& info,
+                                       std::vector<instruction_ref> args) const
+    {
+        int64_t axis = -1;
+        if(contains(info.attributes, "axis"))
+        {
+            axis = parser.parse_value(info.attributes.at("axis")).at<int64_t>();
+        }
+        float epsilon = 1e-5f;
+        if(contains(info.attributes, "epsilon"))
+        {
+            epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
+        }
+        if(contains(info.attributes, "stash_type"))
+        {
+            std::cerr << "WARNING: LAYERNORM does not support stash_type, it will be ignored.\n";
+        }
+
+        if(args.size() < 2 or args.size() > 3)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid input count");
+        }
+
+        auto x         = args.at(0);
+        auto scale     = args.at(1);
+        bool skip_bias = args.size() == 2;
+        instruction_ref bias;
+        if(not skip_bias)
+        {
+            bias = args.at(2);
+        }
+
+        auto x_shape   = x->get_shape();
+        auto x_dtype   = x_shape.type();
+        int64_t x_rank = x_shape.ndim();
+
+        if(x_rank < 2)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid input shape");
+        }
+
+        // If rank(X) is r, axis' allowed range is [-r, r)
+        if(axis < -x_rank or axis >= x_rank)
+        {
+            MIGRAPHX_THROW("PARSE_LAYERNORM: invalid axis");
+        }
+
+        // y = (x - mean) * rsqrt(variance + epsilon) * scale + bias
+        // mean = reduce_mean({D1, D2, ... Dk}, x)
+        // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
+
+        // axis can be negative
+        axis = axis < 0 ? axis + x_rank : axis;
+
+        auto kdims = x_rank - axis;
+        std::vector<int64_t> axes(kdims);
+        std::iota(axes.begin(), axes.end(), axis);
+        auto skipped_axes = x_rank - kdims;
+
+        auto mean          = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x);
+        auto x_sub_mean    = info.add_common_op("sub", x, mean);
+        auto x_sqdiff_mean = info.add_common_op("sqdiff", x, mean);
+        auto variance =
+            info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), x_sqdiff_mean);
+        epsilon =
+            (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon;
+        auto eps     = info.add_literal(migraphx::literal{migraphx::shape{x_dtype}, {epsilon}});
+        auto var_eps = info.add_common_op("add", variance, eps);
+        auto rsqrt   = info.add_instruction(make_op("rsqrt"), var_eps);
+        auto result  = info.add_common_op("mul", x_sub_mean, rsqrt);
+
+        instruction_ref scale_bcast = scale;
+        instruction_ref bias_bcast  = bias;
+        if(skipped_axes > 0)
+        {
+            auto x_dims = x_shape.lens();
+            scale_bcast = info.add_instruction(
+                make_op("broadcast", {{"axis", skipped_axes}, {"out_lens", x_dims}}), scale);
+            if(not skip_bias)
+            {
+                bias_bcast = info.add_instruction(
+                    make_op("broadcast", {{"axis", skipped_axes}, {"out_lens", x_dims}}), bias);
+            }
+        }
+        auto scaled = info.add_instruction(make_op("mul"), result, scale_bcast);
+        auto y      = skip_bias ? scaled : info.add_instruction(make_op("add"), scaled, bias_bcast);
+        return {y, mean, rsqrt};
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_mean_variance_normalization.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/onnx/checks.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_mean_variance_normalization : op_parser<parse_mean_variance_normalization>
+{
+    std::vector<op_desc> operators() const { return {{"MeanVarianceNormalization"}}; }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& /*parser*/,
+                          onnx_parser::node_info info,
+                          std::vector<instruction_ref> args) const
+    {
+        auto&& data    = args.front();
+        auto data_rank = data->get_shape().ndim();
+        std::vector<int64_t> axes{0, 2, 3};
+
+        if(contains(info.attributes, "axes"))
+        {
+            const auto& axes_attr = info.attributes["axes"].ints();
+            axes.assign(axes_attr.begin(), axes_attr.end());
+        }
+        else if(data_rank != 4)
+        {
+            MIGRAPHX_THROW(
+                "Input tensor needs to be rank 4 when axes is not specified. Instead it is rank " +
+                std::to_string(data_rank));
+        }
+
+        if(axes.size() != data_rank - 1)
+        {
+            MIGRAPHX_THROW("Length of axes array needs to be equal to input tensor rank - 1");
+        }
+
+        auto data_mean = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), data);
+        auto data_mean_squared = info.add_common_op("mul", data_mean, data_mean);
+
+        auto data_squared = info.add_common_op("mul", data, data);
+        auto data_squared_mean =
+            info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), data_squared);
+
+        auto mean_sub = info.add_common_op("sub", data_squared_mean, data_mean_squared);
+        auto std      = info.add_common_op("sqrt", mean_sub);
+
+        auto dividend = info.add_common_op("sub", data, data_mean);
+        auto epsilon =
+            info.add_literal({data->get_shape().type(),
+                              {data->get_shape().type() == shape::half_type ? 1e-7 : 1e-9}});
+        auto divisor = info.add_common_op("add", std, epsilon);
+
+        return info.add_common_op("div", dividend, divisor);
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_pad.cpp

@@ -115,34 +115,9 @@ struct parse_pad : op_parser<parse_pad>
 {
     std::vector<op_desc> operators() const { return {{"Pad"}}; }
 
-    instruction_ref parse(const op_desc& /*opd*/,
-                          const onnx_parser& parser,
-                          onnx_parser::node_info info,
-                          std::vector<instruction_ref> args) const
+    std::string parse_mode(const onnx_parser::node_info& info,
+                           const std::vector<instruction_ref>& args) const
     {
-        std::vector<int64_t> pads{};
-        if(args.size() >= 2)
-        {
-            auto pad_arg = args.at(1)->eval();
-            check_arg_empty(pad_arg, "PARSE_PAD: pad input must be constant");
-            pad_arg.visit([&](auto v) { pads.assign(v.begin(), v.end()); });
-        }
-        else if(contains(info.attributes, "pads"))
-        {
-            auto&& pad_vals = info.attributes["pads"].ints();
-            pads            = std::vector<int64_t>(pad_vals.begin(), pad_vals.end());
-        }
-        else
-        {
-            MIGRAPHX_THROW("PARSE_PAD: pad must be available");
-        }
-
-        // check if padding is actually being done (at least one value is nonzero)
-        if(std::all_of(pads.begin(), pads.end(), [](const int& i) { return i == 0; }))
-        {
-            return info.add_instruction(make_op("identity"), args.front());
-        }
-
         if(contains(info.attributes, "mode"))
         {
             auto mode = info.attributes.at("mode").s();
@@ -152,28 +127,59 @@ struct parse_pad : op_parser<parse_pad>
                 {
                     MIGRAPHX_THROW("PARSE_PAD: reflect padding with dynamic shape not supported");
                 }
-                return reflect_pad(info, pads, args.front());
             }
-            if(mode != "constant")
+            else if(mode != "constant")
             {
                 MIGRAPHX_THROW(
                     "PARSE_PAD: migraphx currently only supports constant and reflect padding");
             }
+            return mode;
+        }
+        else
+        {
+            // default mode
+            return "constant";
         }
+    }
 
+    std::vector<int64_t> parse_pads(const onnx_parser::node_info& info,
+                                    const std::vector<instruction_ref>& args) const
+    {
+        std::vector<int64_t> pads{};
+        if(args.size() >= 2)
+        {
+            auto pad_arg = args.at(1)->eval();
+            check_arg_empty(pad_arg, "PARSE_PAD: `pads` input must be constant");
+            pad_arg.visit([&](auto v) { pads.assign(v.begin(), v.end()); });
+        }
+        else if(contains(info.attributes, "pads"))
+        {
+            auto&& pad_vals = info.attributes.at("pads").ints();
+            pads            = std::vector<int64_t>(pad_vals.begin(), pad_vals.end());
+        }
+        else
+        {
+            MIGRAPHX_THROW("PARSE_PAD: `pads` must be available");
+        }
+        return pads;
+    }
+
+    float parse_constant_value(const onnx_parser& parser,
+                               const onnx_parser::node_info& info,
+                               const std::vector<instruction_ref>& args) const
+    {
         float value = 0.0f;
-        // third input is the value
-        if(args.size() == 3)
+        if(args.size() >= 3 and args.at(2)->get_shape().scalar())
         {
             auto val_ins = args.at(2);
             if(not val_ins->can_eval())
             {
-                MIGRAPHX_THROW("PARSE_PAD: input value must be constant");
+                MIGRAPHX_THROW("PARSE_PAD: input `value` must be constant");
             }
             auto val_arg = val_ins->eval();
             if(val_arg.get_shape().elements() != 1)
             {
-                MIGRAPHX_THROW("PARSE_PAD: value should contain only one element");
+                MIGRAPHX_THROW("PARSE_PAD: `value` should contain only one element");
             }
             value = val_arg.at<float>();
         }
@@ -181,6 +187,81 @@ struct parse_pad : op_parser<parse_pad>
         {
             value = parser.parse_value(info.attributes.at("value")).at<float>();
         }
+        return value;
+    }
+
+    std::vector<int64_t> parse_axes(const std::vector<instruction_ref>& args,
+                                    bool is_constant_mode) const
+    {
+        std::vector<int64_t> axes{};
+        // axes is 3rd or 4th, depending on constant mode
+        auto pos = is_constant_mode ? 4 : 3;
+        if(args.size() >= pos)
+        {
+            auto axes_arg = args.at(pos - 1)->eval();
+            check_arg_empty(axes_arg, "PARSE_PAD: variable `axes` input not supported");
+            axes_arg.visit([&](auto v) { axes.assign(v.begin(), v.end()); });
+        }
+        return axes;
+    }
+
+    std::vector<int64_t> calculate_pads_with_axes(const std::vector<int64_t>& pads,
+                                                  const std::vector<int64_t>& axes,
+                                                  size_t input_rank) const
+    {
+        size_t num_axes = axes.size();
+        if(num_axes * 2 != pads.size())
+        {
+            MIGRAPHX_THROW("PARSE_PAD: number of elements of pads should be equal to 2 * "
+                           "number of elements of axes");
+        }
+
+        std::vector<int64_t> new_pads(input_rank * 2);
+        for(size_t idx{0}; idx < num_axes; ++idx)
+        {
+            // axis can be negative
+            int64_t axis = axes[idx] < 0 ? input_rank + axes[idx] : axes[idx];
+            // pad format is x1_begin, x2_begin, ... , x3_end, x4_end
+            new_pads[axis]              = pads[idx];
+            new_pads[axis + input_rank] = pads[idx + num_axes];
+        }
+        return new_pads;
+    }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        std::vector<int64_t> pads = parse_pads(info, args);
+
+        // check if padding is actually being done (at least one value is nonzero)
+        if(std::all_of(pads.begin(), pads.end(), [](const int& i) { return i == 0; }))
+        {
+            return info.add_instruction(make_op("identity"), args.front());
+        }
+
+        std::string mode      = parse_mode(info, args);
+        bool is_constant_mode = mode == "constant";
+        float value           = is_constant_mode ? parse_constant_value(parser, info, args) : 0.0f;
+        std::vector<int64_t> axes = parse_axes(args, is_constant_mode);
+        size_t input_rank         = args.front()->get_shape().ndim();
+
+        if(not axes.empty())
+        {
+            pads = calculate_pads_with_axes(pads, axes, input_rank);
+        }
+
+        if(pads.size() != input_rank * 2)
+        {
+            MIGRAPHX_THROW("PARSE_PAD: number of elements of pads should be equal to 2 * "
+                           "input rank");
+        }
+
+        if(mode == "reflect")
+        {
+            return reflect_pad(info, pads, args.front());
+        }
 
         return info.add_instruction(migraphx::make_op("pad", {{"pads", pads}, {"value", value}}),
                                     args.front());

src/onnx/parse_reshape.cpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -45,15 +45,25 @@ struct parse_reshape : op_parser<parse_reshape>
         {
             literal s = parser.parse_value(info.attributes.at("shape"));
             s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
+            return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
         }
-        if(args.size() == 2)
+        else
         {
+            // 2 inputs
             auto s = args[1]->eval();
-            check_arg_empty(s, "Reshape: non-constant shape input is not supported");
-            s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
+            if(s.empty())
+            {
+                // arg[1] not eval-able
+                auto alloc_ins = info.add_instruction(
+                    make_op("allocate", {{"buf_type", args[0]->get_shape().type()}}), args[1]);
+                return info.add_instruction(make_op("reshape"), args[0], alloc_ins);
+            }
+            else
+            {
+                s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
+                return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
+            }
         }
-
-        return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
     }
 };
 

src/onnx/parse_shrink.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/onnx/checks.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_shrink : op_parser<parse_shrink>
+{
+    std::vector<op_desc> operators() const { return {{"Shrink"}}; }
+
+    instruction_ref parse(const op_desc&,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          std::vector<instruction_ref> args) const
+    {
+        float bias = 0.0;
+        if(contains(info.attributes, "bias"))
+        {
+            bias = parser.parse_value(info.attributes.at("bias")).at<float>();
+        }
+        float lambd = 0.5;
+        if(contains(info.attributes, "lambd"))
+        {
+            lambd = parser.parse_value(info.attributes.at("lambd")).at<float>();
+        }
+
+        auto x             = args[0];
+        auto x_shape       = x->get_shape();
+        auto x_type        = x_shape.type();
+        auto lit_bias      = info.add_literal(bias);
+        auto lit_neg_lambd = info.add_literal(-lambd);
+        auto lit_lambd     = info.add_literal(lambd);
+
+        auto x_plus_bias = info.add_common_op("add", x, lit_bias);
+        auto x_min_bias  = info.add_common_op("sub", x, lit_bias);
+
+        auto cond1   = info.add_common_op("less", x, lit_neg_lambd);
+        auto cond2_a = info.add_common_op("not", cond1);
+        auto cond2_b = info.add_common_op("greater", x, lit_lambd);
+        auto cond2   = info.add_common_op("logical_and", cond2_a, cond2_b);
+
+        auto mul1 = info.add_instruction(make_op("convert", {{"target_type", x_type}}), cond1);
+        auto mul2 = info.add_instruction(make_op("convert", {{"target_type", x_type}}), cond2);
+
+        auto first  = info.add_common_op("mul", mul1, x_plus_bias);
+        auto second = info.add_common_op("mul", mul2, x_min_bias);
+        auto ret    = info.add_common_op("add", first, second);
+        if(ret->get_shape().type() != x_type)
+        {
+            ret = info.add_instruction(make_op("convert", {{"target_type", x_type}}), ret);
+        }
+        return ret;
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_trilu.cpp

@@ -56,9 +56,6 @@ struct parse_trilu : op_parser<parse_trilu>
             k = arg_k.at<int>();
         }
 
-        if(k < 0)
-            MIGRAPHX_THROW("PARSE_TRILU: negative k values not supported");
-
         if(contains(info.attributes, "upper"))
         {
             upper = static_cast<bool>(info.attributes.at("upper").i());
@@ -69,9 +66,12 @@ struct parse_trilu : op_parser<parse_trilu>
         // when creating the mask, if upper == 1,
         // the inner triangle will have values set to 0
         std::vector<bool> mask_mat(num_rows * num_cols, upper);
+        // if upper == 0, kth diagonal must also be masked
+        if(not upper)
+            k++;
         for(size_t i = 0; i < num_rows; i++)
         {
-            for(size_t j = 0; j < std::min(k, static_cast<int>(num_cols)); j++)
+            for(int j = 0; j < std::min(k, static_cast<int>(num_cols)); j++)
             {
                 mask_mat[i * num_cols + j] = not upper;
             }