Merge from develop

.github/workflows/ci.yaml

@@ -7,17 +7,27 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Cancel Previous Runs
-        uses: styfle/cancel-workflow-action@0.6.0
+        uses: styfle/cancel-workflow-action@0.11.0
         with:
           access_token: ${{ github.token }}
   tidy:
-    runs-on: ubuntu-18.04
+    runs-on: ubuntu-20.04
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-
-    - uses: actions/checkout@v2
+      run: |
+        sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku 
+        du . --max-depth=1 -h
+        ls -la
+        cd /usr/local
+        du . --max-depth=1 -h
+        ls -la
+        cd /usr/local/lib
+        echo $(pwd)
+        du . --max-depth=1 -h
+        ls -la
+
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -34,7 +44,7 @@ jobs:
         message("::set-output name=timestamp::${current_date}")
 
     - name: Cache files for tidy
-      uses: pat-s/always-upload-cache@v2.1.3
+      uses: pat-s/always-upload-cache@v3.0.11
       with:
         path: tidy-cache
         key: tidy-cache-${{ steps.cache_timestamp.outputs.timestamp }}
@@ -61,12 +71,12 @@ jobs:
         make -j2 -k onnx-proto tf-proto tidy
 
   cppcheck:
-    runs-on: ubuntu-18.04
+    runs-on: ubuntu-20.04
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -106,12 +116,12 @@ jobs:
         make -j2 cppcheck
 
   format:
-    runs-on: ubuntu-18.04
+    runs-on: ubuntu-20.04
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku 
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -142,14 +152,14 @@ jobs:
           | xargs -n 1 -P 1 -I{} -t sh -c 'yapf {} | diff - {}'
 
   pyflakes:
-    runs-on: ubuntu-18.04
+    runs-on: ubuntu-20.04
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.8
     - name: Install pyflakes
@@ -163,14 +173,14 @@ jobs:
         mypy tools/api.py
 
   licensing:
-    runs-on: ubuntu-18.04
+    runs-on: ubuntu-20.04
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.8
     - name: run License Check
@@ -190,7 +200,6 @@ jobs:
     strategy:
       matrix:
         os:
-          - ubuntu-18.04
           - ubuntu-20.04
         configuration:
           - debug
@@ -199,16 +208,16 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws*  /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
-        python-version: 3.6
+        python-version: 3.7
     - name: Cache dependencies
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
-      uses: actions/cache@v2
+      uses: actions/cache@v3
       with:
         # This path is specific to Ubuntu
         path: ${{ github.workspace }}/cget
@@ -287,7 +296,6 @@ jobs:
     strategy:
       matrix:
         os:
-          - ubuntu-18.04
           - ubuntu-20.04
         configuration:
           - debug
@@ -296,16 +304,16 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
-        python-version: 3.6
+        python-version: 3.7
     - name: Cache dependencies
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
-      uses: actions/cache@v2
+      uses: actions/cache@v3
       with:
         # This path is specific to Ubuntu
         path: ${{ github.workspace }}/cget

.github/workflows/performance.yaml

@@ -5,14 +5,14 @@ on:
     branches: [develop]
     types: [opened, synchronize, closed]
   schedule:
-    - cron: "0 5 * * 1-6"
+    - cron: "0 6 * * 1-6"
 
   workflow_dispatch:
     inputs:
       rocm_release:
         description: ROCm Version
         required: true
-        default: '5.2'
+        default: '5.3'
       performance_reports_repo:
         description: Result repository
         required: true
@@ -30,9 +30,9 @@ concurrency: "perftest-${{ github.head_ref ||  github.base_ref || 'schedule' }}"
 
 jobs:
   release:
-    uses: rocmsoftwareplatform/migraphx-benchmark/.github/workflows/perf-test.yml@main
+    uses: ROCmSoftwarePlatform/migraphx-benchmark/.github/workflows/perf-test.yml@main
     with:
-      rocm_release: ${{ github.event.inputs.rocm_release || '5.2' }}
+      rocm_release: ${{ github.event.inputs.rocm_release || '5.3' }}
       result_number: ${{ github.event.inputs.result_number || '10' }}
       flags: ${{ github.event.inputs.flags || '-s' }} 
       performance_reports_repo: ${{ github.event.inputs.performance_reports_repo || 'ROCmSoftwarePlatform/migraphx-reports' }} 

CMakeLists.txt

@@ -63,7 +63,7 @@ set(CMAKE_EXTRA_INCLUDE_FILES)
 
 include(ROCMSetupVersion)
 
-rocm_setup_version(VERSION 2.4)
+rocm_setup_version(VERSION 2.5)
 set(MIGRAPHX_SO_VERSION ${PROJECT_VERSION_MAJOR}.${PROJECT_VERSION_MINOR})
 
 option( BUILD_SHARED_LIBS "Build as a shared library" ON )
@@ -114,6 +114,7 @@ rocm_enable_clang_tidy(
         hicpp-signed-bitwise
         llvm-namespace-comment
         misc-*
+	-misc-confusable-identifiers
         modernize-*
         performance-*
         readability-*

Dockerfile

@@ -6,7 +6,7 @@ ARG PREFIX=/usr/local
 RUN dpkg --add-architecture i386
 
 # Add rocm repository
-RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.0.2/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.3/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
@@ -71,10 +71,11 @@ RUN /download_models.sh && rm /download_models.sh
 
 # Install latest ccache version
 RUN cget -p $PREFIX install facebook/zstd@v1.4.5 -X subdir -DCMAKE_DIR=build/cmake
-RUN cget -p $PREFIX install ccache@v4.1
+RUN cget -p $PREFIX install ccache@v4.1 -DENABLE_TESTING=OFF
 
 # Install newer cmake for onnx runtime
-RUN cget -p /opt/cmake install kitware/cmake@v3.13.4
+ARG CMAKE_VERSION=3.24.2
+RUN cget -p /opt/cmake install -X binary https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-Linux-x86_64.tar.gz
 
 ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
 ARG ONNXRUNTIME_BRANCH=main
@@ -86,7 +87,7 @@ RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXR
 
 ADD tools/build_and_test_onnxrt.sh /onnxruntime/build_and_test_onnxrt.sh
 
-RUN cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@e8e77eb16be413d301ea8509726d47f265d9011f -DBUILD_MIXR_TARGET=On
+RUN cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@c0723a7e50043d973cb73ae51dc30d36679ee7e5 -DBUILD_MIXR_TARGET=On
 
 ENV MIOPEN_FIND_DB_PATH=/tmp/miopen/find-db
 ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db

dev-requirements.txt

@@ -21,9 +21,9 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 # THE SOFTWARE.
 #####################################################################################
-pfultz2/rocm-recipes
+ROCmSoftwarePlatform/rocm-recipes
 facebook/zstd@v1.4.5 -X subdir -DCMAKE_DIR=build/cmake
-ccache@v4.1
+ccache@v4.1 -DENABLE_TESTING=OFF
 pcre,pfultz2/pcre@8.45 -H sha256:d6f7182602a775a7d500a0cedca6449af0400c6493951513046d17615ed0bf11
 danmar/cppcheck@2.9 -DHAVE_RULES=1
 RadeonOpenCompute/rocm-cmake@1ebf7e7bc61bb5e949c171562b421264065230a7 --build

examples/migraphx/migraphx_driver/README.md

@@ -29,6 +29,7 @@ See below for a comprehensive list of commands and option arguments, as well as
 | --tf | Load file as a tensorflow graph |
 | --migraphx | Load file as a migraphx graph |
 | --migraphx-json | Load file as a migraphx JSON graph |
+| --batch | Set batch size for the model | 
 | --nhwc | Treat tensorflow format as nhwc | 
 | --nchw | Treat tensorflow format as nchw |
 | --skip-unknown-operators | Skip unknown operators when parsing and continue to parse |

hip-clang.docker

@@ -6,7 +6,7 @@ ARG PREFIX=/usr/local
 RUN dpkg --add-architecture i386
 
 # Add rocm repository
-RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.0.2/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.3/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \

rbuild.ini

@@ -2,7 +2,7 @@
 cxx = ${rocm_path}/llvm/bin/clang++
 cc = ${rocm_path}/llvm/bin/clang
 deps =
-    pfultz2/rocm-recipes
+    ROCmSoftwarePlatform/rocm-recipes
     -f requirements.txt
 
 [gh]
@@ -24,4 +24,4 @@ deps =
 define =
     CMAKE_C_COMPILER_LAUNCHER=${deps_dir}/bin/ccache
     CMAKE_CXX_COMPILER_LAUNCHER=${deps_dir}/bin/ccache
-    MIGRAPHX_ENABLE_CPU=On
+    MIGRAPHX_ENABLE_CPU=On
\ No newline at end of file

requirements.txt

@@ -24,7 +24,7 @@
 google/protobuf@v3.11.0 -DCMAKE_POSITION_INDEPENDENT_CODE=On -X subdir -Dprotobuf_BUILD_TESTS=Off
 nlohmann/json@v3.8.0
 live-clones/blaze@v3.8 -X header -DHEADER_DIR=blaze -H sha256:d0ff011f47538285178908ea5f2cab46bb6a8f55b1edb6e03224a82dbc1a3212
-half,https://github.com/pfultz2/half/archive/1.12.0.tar.gz -X header -H sha256:0a08660b68abb176ebc2a0cdf8de46e3182a7f46c66443bb80dbfaaec98cf969
+half,https://github.com/ROCmSoftwarePlatform/half/archive/1.12.0.tar.gz -X header -H sha256:0a08660b68abb176ebc2a0cdf8de46e3182a7f46c66443bb80dbfaaec98cf969
 pybind/pybind11@d159a563383d10c821ba7b2a71905d1207db6de4 --build
 msgpack/msgpack-c@cpp-3.3.0 -DMSGPACK_BUILD_TESTS=Off
 sqlite3@3.17 -DCMAKE_POSITION_INDEPENDENT_CODE=On

src/CMakeLists.txt

@@ -55,6 +55,7 @@ add_library(migraphx
     insert_pad.cpp
     instruction.cpp
     json.cpp
+    layout_nhwc.cpp
     load_save.cpp
     make_op.cpp
     module.cpp
@@ -81,7 +82,6 @@ add_library(migraphx
     replace_allocate.cpp
     simplify_qdq.cpp
     sqlite.cpp
-    rewrite_batchnorm.cpp
     rewrite_gelu.cpp
     rewrite_pooling.cpp
     rewrite_quantization.cpp
@@ -115,7 +115,6 @@ register_migraphx_ops(
     as_shape
     atanh
     atan
-    batch_norm_inference
     broadcast
     capture
     ceil
@@ -146,6 +145,7 @@ register_migraphx_ops(
     if_op
     im2col
     isnan
+    layout
     leaky_relu
     less
     load

src/api/include/migraphx/migraphx.hpp

@@ -32,6 +32,7 @@
 #include <memory>
 #include <numeric>
 #include <exception>
+#include <array>
 #include <vector>
 #include <cassert>
 #include <iostream>

src/auto_contiguous.cpp

@@ -59,6 +59,8 @@ void auto_contiguous::apply(module& m) const
     auto last = std::prev(m.end());
     for(auto ins : iterator_for(m))
     {
+        if(ins->name() == "layout")
+            continue;
         // for last instruction that is NOT a return
         if(ins->outputs().empty() and ins != last)
             continue;

src/common.cpp

@@ -27,6 +27,7 @@
 #include <migraphx/algorithm.hpp>
 #include <migraphx/stringutils.hpp>
 #include <migraphx/instruction.hpp>
+#include <migraphx/ranges.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -43,6 +44,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 // In this case we need to broadcast the (:,:,1:,:) axis
 // of s0 plus the 1st dimension of s1 giving
 // output_lens = (3,2,7,5)
+//
 std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
                                                   std::vector<std::size_t> s1)
 {
@@ -50,25 +52,63 @@ std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
         return s0;
     if(s0.size() > s1.size())
         s0.swap(s1);
-
     std::vector<std::size_t> out_lens(s1);
     auto offset = s1.size() - s0.size();
     std::transform(
         s0.begin(), s0.end(), s1.begin() + offset, out_lens.begin() + offset, [&](auto a, auto b) {
             if(a != b and a != 1 and b != 1)
             {
-                MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" + to_string_range(s0) + "} and {" +
-                               to_string_range(s1) + "} mismatch!");
+                MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" + migraphx::to_string_range(s0) +
+                               "} and {" + migraphx::to_string_range(s1) + "} mismatch!");
             }
             return std::max(a, b);
         });
-
     return out_lens;
 }
 
+std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, shape s1)
+{
+    // change both shapes to dynamic_dimension representation
+    s0 = s0.to_dynamic();
+    s1 = s1.to_dynamic();
+    if(s0.ndim() > s1.ndim())
+    {
+        std::swap(s0, s1);
+    }
+    auto offset = s1.ndim() - s0.ndim();
+    std::vector<shape::dynamic_dimension> out_dims(s1.dyn_dims());
+    shape::dynamic_dimension one_dyn_dim{1, 1, 0};
+    std::transform(
+        s0.dyn_dims().cbegin(),
+        s0.dyn_dims().cend(),
+        s1.dyn_dims().cbegin() + offset,
+        out_dims.begin() + offset,
+        [&](auto a, auto b) {
+            if(a == b)
+            {
+                return a;
+            }
+            else if(a == one_dyn_dim or b == one_dyn_dim)
+            {
+                // setting opt to 0, may need to be changed
+                return shape::dynamic_dimension{std::max(a.min, b.min), std::max(a.max, b.max), 0};
+            }
+            else
+            {
+                MIGRAPHX_THROW("COMPUTE_BROADCASTED_DYN_DIMS: dynamic shapes {" +
+                               migraphx::to_string_range(s0.dyn_dims()) + "} and {" +
+                               migraphx::to_string_range(s1.dyn_dims()) + "} mismatch!");
+            }
+        });
+    return out_dims;
+}
+
+// Compute the common (broadcasted) dimensions of a list of fixed shapes
 std::vector<std::size_t> compute_common_lens(const std::vector<shape>& shapes)
 {
     assert(not shapes.empty());
+    assert(
+        std::none_of(shapes.cbegin(), shapes.cend(), [](auto shape) { return shape.dynamic(); }));
     return transform_accumulate(shapes.begin() + 1,
                                 shapes.end(),
                                 shapes.front().lens(),
@@ -114,20 +154,63 @@ instruction_ref insert_common_op(module& m,
                                  const operation& op,
                                  std::vector<instruction_ref> inputs)
 {
-    auto common = common_shape(to_shapes(inputs));
-    std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto input) {
-        if(input->get_shape().lens() != common.lens())
+    if(std::any_of(
+           inputs.cbegin(), inputs.cend(), [](auto input) { return input->get_shape().dynamic(); }))
+    {
+        // currently only handles the binary case
+        if(inputs.size() != 2)
         {
-            input = m.insert_instruction(
-                ins, make_op("multibroadcast", {{"out_lens", common.lens()}}), input);
+            MIGRAPHX_THROW("INSERT_COMMON_OP: not handled; " + migraphx::to_string(inputs.size()) +
+                           "inputs, only handle two inputs if any are dynamic shape");
         }
-        if(input->get_shape().type() != common.type())
+
+        auto c_type = compute_common_types(to_shapes(inputs));
+        auto c_dyn_dims =
+            compute_broadcasted_dyn_dims(inputs[0]->get_shape(), inputs[1]->get_shape());
+
+        // following should work for a static or dynamic shape
+        if(inputs[0]->get_shape().dyn_dims() != c_dyn_dims)
         {
-            input = m.insert_instruction(
-                ins, make_op("convert", {{"target_type", common.type()}}), input);
+            inputs[0] = m.insert_instruction(
+                ins,
+                make_op("multibroadcast", {{"out_dyn_dims", to_value(c_dyn_dims)}}),
+                inputs[0],
+                inputs[1]);
         }
-        return input;
-    });
+        if(inputs[1]->get_shape().dyn_dims() != c_dyn_dims)
+        {
+            inputs[1] = m.insert_instruction(
+                ins,
+                make_op("multibroadcast", {{"out_dyn_dims", to_value(c_dyn_dims)}}),
+                inputs[1],
+                inputs[0]);
+        }
+        std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto input) {
+            if(input->get_shape().type() != c_type)
+            {
+                input =
+                    m.insert_instruction(ins, make_op("convert", {{"target_type", c_type}}), input);
+            }
+            return input;
+        });
+    }
+    else
+    {
+        auto common = common_shape(to_shapes(inputs));
+        std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto input) {
+            if(input->get_shape().lens() != common.lens())
+            {
+                input = m.insert_instruction(
+                    ins, make_op("multibroadcast", {{"out_lens", common.lens()}}), input);
+            }
+            if(input->get_shape().type() != common.type())
+            {
+                input = m.insert_instruction(
+                    ins, make_op("convert", {{"target_type", common.type()}}), input);
+            }
+            return input;
+        });
+    }
     return m.insert_instruction(ins, op, inputs);
 }
 

src/driver/alexnet.cpp

@@ -74,9 +74,9 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
     auto x_main_module_19 = mmain->add_literal(migraphx::generate_literal(
         migraphx::shape{migraphx::shape::float_type, {64, 3, 11, 11}}, 18));
     auto x_main_module_20 = mmain->add_instruction(
-        migraphx::make_json_op("convolution",
-                               "{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[4,"
-                               "4],use_dynamic_same_auto_pad:0}"),
+        migraphx::make_json_op(
+            "convolution",
+            "{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[4,4]}"),
         x_0,
         x_main_module_19);
     auto x_main_module_21 = mmain->add_instruction(
@@ -90,9 +90,9 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
             "{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,0,0],stride:[2,2]}"),
         x_main_module_23);
     auto x_main_module_25 = mmain->add_instruction(
-        migraphx::make_json_op("convolution",
-                               "{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[1,"
-                               "1],use_dynamic_same_auto_pad:0}"),
+        migraphx::make_json_op(
+            "convolution",
+            "{dilation:[1,1],group:1,padding:[2,2,2,2],padding_mode:0,stride:[1,1]}"),
         x_main_module_24,
         x_main_module_17);
     auto x_main_module_26 = mmain->add_instruction(
@@ -106,9 +106,9 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
             "{ceil_mode:0,lengths:[3,3],lp_order:2,mode:1,padding:[0,0,0,0],stride:[2,2]}"),
         x_main_module_28);
     auto x_main_module_30 = mmain->add_instruction(
-        migraphx::make_json_op("convolution",
-                               "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
-                               "1],use_dynamic_same_auto_pad:0}"),
+        migraphx::make_json_op(
+            "convolution",
+            "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,1]}"),
         x_main_module_29,
         x_main_module_15);
     auto x_main_module_31 = mmain->add_instruction(
@@ -117,9 +117,9 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
         mmain->add_instruction(migraphx::make_op("add"), x_main_module_30, x_main_module_31);
     auto x_main_module_33 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_32);
     auto x_main_module_34 = mmain->add_instruction(
-        migraphx::make_json_op("convolution",
-                               "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
-                               "1],use_dynamic_same_auto_pad:0}"),
+        migraphx::make_json_op(
+            "convolution",
+            "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,1]}"),
         x_main_module_33,
         x_main_module_13);
     auto x_main_module_35 = mmain->add_instruction(
@@ -128,9 +128,9 @@ migraphx::program alexnet(unsigned batch) // NOLINT(readability-function-size)
         mmain->add_instruction(migraphx::make_op("add"), x_main_module_34, x_main_module_35);
     auto x_main_module_37 = mmain->add_instruction(migraphx::make_op("relu"), x_main_module_36);
     auto x_main_module_38 = mmain->add_instruction(
-        migraphx::make_json_op("convolution",
-                               "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,"
-                               "1],use_dynamic_same_auto_pad:0}"),
+        migraphx::make_json_op(
+            "convolution",
+            "{dilation:[1,1],group:1,padding:[1,1,1,1],padding_mode:0,stride:[1,1]}"),
         x_main_module_37,
         x_main_module_11);
     auto x_main_module_39 = mmain->add_instruction(

src/driver/main.cpp

@@ -44,7 +44,6 @@
 #include <migraphx/propagate_constant.hpp>
 #include <migraphx/quantization.hpp>
 #include <migraphx/register_op.hpp>
-#include <migraphx/rewrite_batchnorm.hpp>
 #include <migraphx/simplify_algebra.hpp>
 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/register_target.hpp>
@@ -221,7 +220,6 @@ struct loader
         {
             migraphx::run_passes(*p.get_main_module(),
                                  {
-                                     migraphx::rewrite_batchnorm{},
                                      migraphx::eliminate_identity{},
                                      migraphx::dead_code_elimination{},
                                      migraphx::simplify_algebra{},

src/driver/verify.cpp

@@ -145,7 +145,7 @@ void verify_reduced(program p,
     auto* mm  = p.get_main_module();
     auto last = std::prev(mm->end(), n + 1);
     mm->remove_instructions(last, mm->end());
-    std::cout << "Verify: " << std::endl;
+    std::cout << "Verify: " << n << std::endl;
     std::cout << p << std::endl;
     verify_program(std::to_string(n), p, t, options, quantize, inputs, tolerance);
 }
@@ -159,6 +159,7 @@ void verify_reduced_program(const program& p,
 {
     const auto* mm = p.get_main_module();
     auto n         = std::distance(mm->begin(), mm->end());
+    std::cout << "Verify steps: " << n << std::endl;
     for(std::size_t i = 0; i < n; i++)
     {
         verify_reduced(p, i, t, options, quantize, inputs, tolerance);

src/eliminate_contiguous.cpp

@@ -42,6 +42,13 @@ static bool try_compute_shape(instruction_ref ins,
     try
     {
         shape new_shape = ins->get_operator().compute_shape(inputs, mods);
+
+        // Cannot tell if a dynamic shape will need to be made contiguous
+        if(new_shape.dynamic())
+        {
+            return false;
+        }
+
         // If the output shape is a standard shape, no need to try its output
         if(new_shape.standard())
         {
@@ -133,14 +140,20 @@ static void remove_contiguous(const std::string& op_name, module& m, F f)
         }
     }
 
-    // Perform evaluations in parallel
+    // Perform static contiguous evaluations in parallel
     std::vector<argument> literals(const_instructions.size());
     par_for(const_instructions.size(), 1, [&](const auto i) {
-        auto c      = op::contiguous{};
-        auto prev   = const_instructions[i]->inputs().front();
-        literals[i] = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()});
+        auto c    = op::contiguous{};
+        auto prev = const_instructions[i]->inputs().front();
+        // compute the output contiguous shape from the previous instruction shape
+        shape computed_shape                   = c.compute_shape({prev->get_shape()});
+        const std::vector<argument>& prev_eval = {prev->eval()};
+        // prev_eval should not be used in make_compute_output_shape() as computed_shape is static
+        auto co_shape = make_compute_output_shape(pack(c, computed_shape, prev_eval));
+        literals[i]   = c.compute(co_shape, prev_eval);
     });
 
+    // Replace static contiguous operations with a literal
     for(size_t i = 0; i < const_instructions.size(); i++)
     {
         auto l = m.add_literal(literals[i].get_shape(), literals[i].data());

src/fuse_pointwise.cpp

@@ -39,13 +39,22 @@ static literal get_scalar(instruction_ref ins)
     if(ins->name() == "contiguous")
         return get_scalar(ins->inputs().front());
     const auto& s = ins->get_shape();
-    if(not(s.elements() == 1 or s.scalar()))
+    if(s.elements() != 1 && not(s.scalar()))
         return {};
     if(not ins->can_eval())
         return {};
     auto e = ins->eval();
     literal r{};
-    e.visit_at([&](auto x) { r = literal{x}; });
+    // needed for bool as visit_at invokes as() which promotes bool to int8
+    // Without this we'll break type checks for logical ops that are fused.
+    if(e.get_shape().type() == shape::bool_type)
+    {
+        r = literal{e.at<bool>()};
+    }
+    else
+    {
+        e.visit_at([&](auto x) { r = literal{x}; });
+    }
     return r;
 }
 
@@ -56,6 +65,8 @@ static void create_pointwise_modules(module_pass_manager& mpm)
     {
         if(not ins->get_operator().attributes().get("pointwise", false))
             continue;
+        if(ins->get_operator().name() == "layout")
+            continue;
         assert(ins->get_operator().attributes().contains("point_op"));
         auto* pm = mpm.create_module(mpm.get_module().name() + ":pointwise" + std::to_string(n++));
         pm->set_bypass();