Merge branch 'develop' into add_parity_check_ci

.githooks/pre-commit

@@ -4,62 +4,7 @@
 # are installed, and if so, uses the installed version to format
 # the staged changes.
 
-base=clang-format-10
-format=""
-yapf_base=yapf
-yapf_format=""
-
-use_yapf=true
-
-# Redirect output to stderr.
-exec 1>&2
-
-# check if clang-format is installed
-type "$base" >/dev/null 2>&1 && format="$base"
-
-# no versions of clang-format are installed
-if [ -z "$format" ]
-then
-    echo "$base is not installed. Pre-commit hook will not be executed."
-    exit 0
-fi
-
-# check if yapf is installed
-type "$yapf_base" >/dev/null 2>&1 && yapf_format="$yapf_base"
-
-# no versions of yapf are installed
-if [ -z "$yapf_format" ]
-then
-    echo "$yapf_base is not installed. Pre-commit hook for python files will not be executed"
-    use_yapf=false
-fi
-
 # Do everything from top - level
 cd $(git rev-parse --show-toplevel)
 
-if git rev-parse --verify HEAD >/dev/null 2>&1
-then
-    against=HEAD
-else
-    # Initial commit: diff against an empty tree object
-    against=16bbb57
-fi
-
-# do the formatting
-for file in $(git diff-index --cached --name-only $against | grep -E '\.h$|\.hpp$|\.cpp$|\.cl$|\.c$|\.h\.in$|\.hpp\.in$|\.cpp\.in$|\.py$')
-do
-    if [ -e "$file" ]
-    then
-        if [ $(echo $file | grep -E '\.py$') ]
-        then
-            if $use_yapf
-            then
-                echo "$yapf_format $file"
-                "$yapf_format" -i "$file"
-            fi
-        else
-            echo "$format $file"
-            "$format" -i -style=file "$file"
-        fi
-    fi
-done
+python3 tools/format.py -q -i HEAD

.github/workflows/ci.yaml

@@ -144,6 +144,8 @@ jobs:
     runs-on: ROCM-Ubuntu
     steps:
     - uses: actions/checkout@v3
+      with:
+        fetch-depth: 0
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -164,19 +166,8 @@ jobs:
       shell: bash -c "docker run -i -v=$GITHUB_WORKSPACE:/data -w /data migraphx bash < {0}"
       run: |
         set -e
-        find . -iname '*.h' \
-              -o -iname '*.hpp' \
-              -o -iname '*.cpp' \
-              -o -iname '*.h.in' \
-              -o -iname '*.hpp.in' \
-              -o -iname '*.cpp.in' \
-              -o -iname '*.cl' \
-              -o -iname '*.c' \
-          | grep -v 'build/' \
-          | xargs -n 1 -P 1 -I{} -t sh -c 'clang-format-10 -style=file {} | diff - {}'
-          find . -iname '*.py' \
-          | grep -v 'build/'  \
-          | xargs -n 1 -P 1 -I{} -t sh -c 'yapf {} | diff - {}'
+        git config --global --add safe.directory /data
+        python3 tools/format.py origin/${{ github.event_name == 'pull_request' && github.base_ref || 'develop' }}
 
   pyflakes:
     runs-on: ubuntu-20.04
@@ -323,6 +314,7 @@ jobs:
       env:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       run: |
+        set +x
         gh extension install actions/gh-actions-cache --pin v1.0.1
         gh actions-cache delete ${{ steps.ccache_restore.outputs.cache-matched-key }} --confirm
 
@@ -439,6 +431,7 @@ jobs:
       env:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       run: |
+        set +x
         gh extension install actions/gh-actions-cache
         gh actions-cache delete ${{ steps.ccache_restore_fpga.outputs.cache-matched-key }} --confirm
       continue-on-error: true

.github/workflows/performance.yaml

@@ -12,7 +12,7 @@ on:
       rocm_release:
         description: ROCm Version
         required: true
-        default: '5.5'
+        default: '5.6'
       performance_reports_repo:
         description: Repository where performance reports are stored
         required: true
@@ -42,13 +42,15 @@ on:
         required: true
         default: '-r'
 
-concurrency: "perftest-${{ github.head_ref ||  github.base_ref || 'schedule' }}"
+concurrency: 
+  group: "perftest-${{ github.head_ref ||  github.base_ref || 'schedule' }}"
+  cancel-in-progress: true
 
 jobs:
   release:
     uses: ROCmSoftwarePlatform/migraphx-benchmark/.github/workflows/perf-test.yml@main
     with:
-      rocm_release: ${{ github.event.inputs.rocm_release || '5.5' }}
+      rocm_release: ${{ github.event.inputs.rocm_release || '5.6' }}
       result_number: ${{ github.event.inputs.result_number || '10' }}
       flags: ${{ github.event.inputs.flags || '-r' }}
       performance_reports_repo: ${{ github.event.inputs.performance_reports_repo || 'ROCmSoftwarePlatform/migraphx-reports' }}

.gitignore

@@ -70,5 +70,13 @@ docs/_images
 docs/_static
 docs/_templates
 docs/.doxygen/docBin
+docs/.doxygen/DoxygenWarningLog.txt
 docs/_doxygen
+docs/html
 /_readthedocs
+
+# JetBrains config directories (ignoring symlinks)
+.idea/
+cmake-build*/
+build*/
+

CMakeLists.txt

@@ -43,6 +43,8 @@ else()
     endif()
 endif()
 
+set(CMAKE_INSTALL_PREFIX "/opt/rocm" CACHE PATH "")
+
 set(CMAKE_BUILD_RPATH "${CMAKE_BINARY_DIR}/lib")
 
 project(migraphx LANGUAGES C CXX)
@@ -114,6 +116,7 @@ rocm_enable_clang_tidy(
         llvm-namespace-comment
         misc-*
 	-misc-confusable-identifiers
+        -misc-use-anonymous-namespace
         modernize-*
         performance-*
         readability-*

Dockerfile

@@ -7,10 +7,10 @@ RUN dpkg --add-architecture i386
 
 # Install rocm key
 RUN apt-get update && apt-get install -y gnupg2 --no-install-recommends curl && \
-    curl -sL http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - 
+    curl -sL http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
 
 # Add rocm repository
-RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.5/ focal main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.6/ focal main > /etc/apt/sources.list.d/rocm.list'
 
 # From docs.amd.com for installing rocm. Needed to install properly
 RUN sh -c "echo 'Package: *\nPin: release o=repo.radeon.com\nPin-priority: 600' > /etc/apt/preferences.d/rocm-pin-600"
@@ -19,7 +19,6 @@ RUN sh -c "echo 'Package: *\nPin: release o=repo.radeon.com\nPin-priority: 600'
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
     apt-utils \
     build-essential \
-    clang-format-10 \
     cmake \
     curl \
     doxygen \
@@ -97,7 +96,8 @@ 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 -DENABLE_TESTING=OFF
-RUN cget -p /opt/cmake install kitware/cmake@v3.24.3
+RUN cget -p /opt/cmake install kitware/cmake@v3.26.4
+
 
 COPY ./test/onnx/.onnxrt-commit /
 
@@ -115,7 +115,8 @@ ADD tools/build_and_install_onnxrt.sh /onnxruntime/build_and_install_onnxrt.sh
 ADD tools/test_onnxrt_unit_tests.sh /onnxruntime/test_onnxrt_unit_tests.sh
 ADD tools/test_onnxrt_parity_tests.sh /onnxruntime/test_onnxrt_parity_tests.sh
 
-RUN cget -p /usr/local install ROCmSoftwarePlatform/rocMLIR@8d25af3b3721c159bb41cc6388e9453b1018c126 -DBUILD_MIXR_TARGET=On -DLLVM_ENABLE_ZSTD=Off -DLLVM_ENABLE_THREADS=Off
+# Use the /opt/cmake install because LLVM/MLIR need cmake >= 3.20
+RUN env PATH=/opt/cmake/bin:$PATH cget -p /usr/local install ROCmSoftwarePlatform/rocMLIR@1ad9d6df32acc6d29d58e8ed6710e36746d0a4d6 -DBUILD_FAT_LIBROCKCOMPILER=On
 
 ENV MIOPEN_FIND_DB_PATH=/tmp/miopen/find-db
 ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db

Jenkinsfile

@@ -15,13 +15,11 @@ def rocmtestnode(Map conf) {
         def compiler = bconf.get("compiler", "/opt/rocm/llvm/bin/clang++")
         def flags = bconf.get("flags", "")
         def gpu_debug = bconf.get("gpu_debug", "0")
-        def hiprtc_workarounds = bconf.get("hiprtc_workarounds", "0")
         def cmd = """
             ulimit -c unlimited
             echo "leak:dnnl::impl::malloc" > suppressions.txt
             export LSAN_OPTIONS="suppressions=\$(pwd)/suppressions.txt"
             export MIGRAPHX_GPU_DEBUG=${gpu_debug}
-            export MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS=${hiprtc_workarounds}
             export CXX=${compiler}
             export CXXFLAGS='-Werror'
             env
@@ -90,9 +88,9 @@ def rocmnodename(name) {
     } else if(name == "navi21") {
         node_name = "${rocmtest_name} && navi21";
     } else if(name == "mi100+") {
-        node_name = "${rocmtest_name} && (gfx908 || gfx90a)";
-    } else if(name == "anygpu") {
-        node_name = "${rocmtest_name} && (gfx908 || gfx90a || vega)";
+        node_name = "${rocmtest_name} && (gfx908 || gfx90a) && !vm";
+    } else if(name == "cdna") {
+        node_name = "${rocmtest_name} && (gfx908 || gfx90a || vega) && !vm";
     } else if(name == "nogpu") {
         node_name = "${rocmtest_name} && nogpu";
     }
@@ -105,35 +103,29 @@ def rocmnode(name, body) {
     }
 }
 
-rocmtest clang_debug: rocmnode('vega') { cmake_build ->
-    stage('Hip Clang Debug') {
+rocmtest clang_debug: rocmnode('cdna') { cmake_build ->
+    stage('hipRTC Debug') {
         def sanitizers = "undefined"
         def debug_flags = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
-        cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}'")
+        cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}' -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true)
     }
-}, clang_gpu_debug: rocmnode('vega') { cmake_build ->
-    stage('Hip Clang GPU Debug') {
-        cmake_build(flags: "-DCMAKE_BUILD_TYPE=release", gpu_debug: true)
-    }
-}, clang_release: rocmnode('vega') { cmake_build ->
+}, clang_release: rocmnode('cdna') { cmake_build ->
     stage('Hip Clang Release') {
         cmake_build(flags: "-DCMAKE_BUILD_TYPE=release")
         stash includes: 'build/*.deb', name: 'migraphx-package'
     }
-}, hiprtc_gpu_debug: rocmnode('vega') { cmake_build ->
-    stage('HipRTC GPU Debug') {
-        cmake_build(flags: "-DCMAKE_BUILD_TYPE=release -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true, hiprtc_workarounds: true)
-    }
-}, all_targets_debug : rocmnode('vega') { cmake_build ->
+}, all_targets_debug : rocmnode('cdna') { cmake_build ->
     stage('All targets Release') {
         cmake_build(flags: "-DCMAKE_BUILD_TYPE=release -DMIGRAPHX_ENABLE_GPU=On -DMIGRAPHX_ENABLE_CPU=On -DMIGRAPHX_ENABLE_FPGA=On") 
     }
-}, mlir_debug: rocmnode('vega') { cmake_build ->
+}, mlir_debug: rocmnode('cdna') { cmake_build ->
     stage('MLIR Debug') {
         withEnv(['MIGRAPHX_ENABLE_MLIR=1']) {
             def sanitizers = "undefined"
-            def debug_flags = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
-            cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DMIGRAPHX_ENABLE_MLIR=On -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}'")
+            // Note: the -fno-sanitize= is copied from upstream LLVM_UBSAN_FLAGS.
+            def debug_flags_cxx = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize=vptr,function -fno-sanitize-recover=${sanitizers}"
+            def debug_flags = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize=vptr -fno-sanitize-recover=${sanitizers}"
+            cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DMIGRAPHX_ENABLE_MLIR=On -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags_cxx}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}'")
         }
     }
 }, ck_release: rocmnode('mi100+') { cmake_build ->
@@ -163,6 +155,7 @@ def onnxnode(name, body) {
     }
 }
 
+
 rocmtest onnxbuild onnxnode('anygpu') { cmake_build ->
     stage("Onnxruntime Build and Install") {
         sh '''

docs/dev_intro.rst

@@ -131,7 +131,7 @@ In this case, we can create `argument <migraphx::argument>` objects directly fro
     std::vector<float> results_vector(64);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
 
-    EXPECT(migraphx::verify_range(results_vector, sol));
+    EXPECT(migraphx::verify::verify_range(results_vector, sol));
 
 An `argument <migraphx::argument>` can handle memory buffers from either the GPU or the CPU.
 By default when running the `program <migraphx::program>`, buffers are allocated on the corresponding target.

hip-clang.docker

@@ -6,13 +6,12 @@ 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/5.5/ focal main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/5.6/ focal main > /etc/apt/sources.list.d/rocm.list'
 
 # Install dependencies
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
     apt-utils \
     build-essential \
-    clang-format-10 \
     cmake \
     curl \
     doxygen \

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
-ROCmSoftwarePlatform/half@rocm-5.4.2
+ROCmSoftwarePlatform/half@rocm-5.6.0
 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

@@ -129,10 +129,11 @@ register_migraphx_ops(
     contiguous
     convert
     convolution
+    convolution_backwards
     cosh
     cos
-    deconvolution
     dequantizelinear
+    dimensions_of
     div
     dot
     elu

src/api/include/migraphx/migraphx.hpp

@@ -1487,13 +1487,17 @@ quantize_int8(const program& prog, const target& ptarget, const quantize_int8_op
 
 struct experimental_custom_op_base
 {
+    experimental_custom_op_base()                                   = default;
+    experimental_custom_op_base(const experimental_custom_op_base&) = default;
+    experimental_custom_op_base& operator=(const experimental_custom_op_base&) = default;
+    virtual ~experimental_custom_op_base()                                     = default;
+
     virtual std::string name() const                                            = 0;
     virtual argument compute(context ctx, shape output, arguments inputs) const = 0;
     virtual shape compute_shape(shapes inputs) const                            = 0;
     virtual std::vector<size_t> output_alias(shapes) const { return {}; }
     // TODO: Return target string instead of bool
     virtual bool runs_on_offload_target() const = 0;
-    virtual ~experimental_custom_op_base()      = default;
 };
 
 struct experimental_custom_op : interface_base<MIGRAPHX_HANDLE_BASE(experimental_custom_op)>

src/common.cpp

@@ -150,8 +150,8 @@ insert_common_args(module& m, instruction_ref ins, std::vector<instruction_ref>
         auto c_type       = compute_common_types(input_shapes);
         auto c_dyn_dims   = compute_common_dyn_dims(input_shapes);
 
-        // following should work for a static or dynamic shape
-        if(inputs[0]->get_shape().dyn_dims() != c_dyn_dims)
+        auto s0 = inputs[0]->get_shape();
+        if(not s0.dynamic() or s0.dyn_dims() != c_dyn_dims)
         {
             inputs[0] = m.insert_instruction(
                 ins, make_op("multibroadcast", {{"out_dyn_dims", to_value(c_dyn_dims)}}), inputs);
@@ -159,7 +159,8 @@ insert_common_args(module& m, instruction_ref ins, std::vector<instruction_ref>
         std::transform(inputs.begin() + 1, inputs.end(), inputs.begin() + 1, [&](auto input) {
             // uses previous input to avoid recalculating the common shape from the
             // full set of input shapes at runtime
-            if(input->get_shape().dyn_dims() != c_dyn_dims)
+            auto s = input->get_shape();
+            if(not s.dynamic() or s.dyn_dims() != c_dyn_dims)
             {
                 return m.insert_instruction(
                     ins,

src/generate.cpp

@@ -26,7 +26,7 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-argument fill_argument(shape s, unsigned long value)
+argument fill_argument(shape s, double value)
 {
     argument result;
     if(s.type() == shape::tuple_type)

src/include/migraphx/generate.hpp

@@ -117,14 +117,14 @@ auto generate_tensor_data(const migraphx::shape& s, unsigned long seed = 0)
 }
 
 template <class T>
-auto fill_tensor_data(const migraphx::shape& s, unsigned long value = 0)
+auto fill_tensor_data(const migraphx::shape& s, double value = 0)
 {
     auto result = make_shared_array<T>(s.element_space());
     std::generate(result.get(), result.get() + s.element_space(), [=] { return value; });
     return result;
 }
 
-MIGRAPHX_EXPORT argument fill_argument(shape s, unsigned long value = 0);
+MIGRAPHX_EXPORT argument fill_argument(shape s, double value = 0);
 
 MIGRAPHX_EXPORT argument generate_argument(shape s, unsigned long seed = 0);
 

src/include/migraphx/normalize_attributes.hpp

@@ -43,7 +43,7 @@ template <class T, class... Ts>
 using dependent_type = typename select_dependent_type<T, Ts...>::type;
 
 MIGRAPHX_EXPORT
-bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens);
+bool normalize_attributes(operation& op, const shape& input_shape);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/op/convert.hpp

@@ -66,7 +66,19 @@ struct convert : unary<convert>
         auto type = target_type;
         return [type](auto x) {
             auto y = x;
-            shape::visit(type, [&](auto as) { y = as(x); });
+            shape::visit(type, [&](auto as) {
+                // clamping value between target_type's max and min doesn't work for NaNs,
+                if(std::isnan(x))
+                {
+                    y = as.nan();
+                }
+                else
+                {
+                    // clamp overflowing/underflowing values to min()/max() instead of +/-infinity
+                    // during downcasting
+                    y = std::min(std::max(as(x), as.min()), as.max());
+                }
+            });
             return y;
         };
     }

src/include/migraphx/op/convolution.hpp

@@ -79,17 +79,17 @@ struct convolution
         check_shapes{inputs, *this, true}.has(2).same_type().same_ndims().min_ndims(3);
         check_attribute_size();
         // num of dims of input and attribute should match
-        const auto input_size   = inputs[0].max_lens().size();
+        const auto input_ndim   = inputs[0].ndim();
         const auto padding_size = padding.size();
 
-        if(input_size != padding_size / 2 + 2 && input_size != padding_size + 2)
+        if(input_ndim != padding_size / 2 + 2 && input_ndim != padding_size + 2)
         {
             MIGRAPHX_THROW("CONVOLUTION: input and attribute size mismatch!");
         }
 
         const shape& x_shape          = inputs.at(0);
         const shape& w_shape          = inputs.at(1);
-        const size_t num_spatial_dims = input_size - 2;
+        const size_t num_spatial_dims = input_ndim - 2;
         if(num_spatial_dims != this->kdims())
         {
             MIGRAPHX_THROW("CONVOLUTION: input k-dims does not match attribute size");
@@ -105,7 +105,7 @@ struct convolution
         }
         else
         {
-            return fixed_compute_shape(x_shape, w_shape);
+            return static_compute_shape(x_shape, w_shape);
         }
     }
 
@@ -143,23 +143,10 @@ struct convolution
     shape dynamic_compute_shape(shape x_shape, shape w_shape) const
     {
         std::vector<shape::dynamic_dimension> output_dyn_dims = {};
+        output_dyn_dims.push_back(x_shape.to_dynamic().dyn_dims().at(0));
+        output_dyn_dims.push_back(w_shape.to_dynamic().dyn_dims().at(0));
 
-        auto dynamic_shape_push_back = [&](const shape& input_shape) {
-            if(input_shape.dynamic())
-            {
-                output_dyn_dims.push_back(input_shape.dyn_dims().at(0));
-            }
-            else
-            {
-                auto l = input_shape.lens().at(0);
-                output_dyn_dims.push_back({l, l});
-            }
-        };
-
-        dynamic_shape_push_back(x_shape);
-        dynamic_shape_push_back(w_shape);
-
-        const size_t num_spatial_dims = x_shape.max_lens().size() - 2;
+        const size_t num_spatial_dims = x_shape.ndim() - 2;
         if(padding_mode != default_)
         {
             for(std::size_t i = 0; i < num_spatial_dims; ++i)
@@ -198,7 +185,7 @@ struct convolution
         return shape{x_shape.type(), output_dyn_dims};
     }
 
-    shape fixed_compute_shape(shape x_shape, shape w_shape) const
+    shape static_compute_shape(shape x_shape, shape w_shape) const
     {
         std::vector<size_t> output_lens{x_shape.lens()[0], w_shape.lens()[0]};
         auto spatial_lens = calc_conv_lens(x_shape.lens(), w_shape.lens());

src/include/migraphx/op/deconvolution.hpp → src/include/migraphx/op/convolution_backwards.hpp

@@ -21,9 +21,11 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#ifndef MIGRAPHX_GUARD_OPERATORS_DECONVOLUTION_HPP
-#define MIGRAPHX_GUARD_OPERATORS_DECONVOLUTION_HPP
+#ifndef MIGRAPHX_GUARD_OPERATORS_CONVOLUTION_BACKWARDS_HPP
+#define MIGRAPHX_GUARD_OPERATORS_CONVOLUTION_BACKWARDS_HPP
 
+#include <cmath>
+#include <utility>
 #include <migraphx/op/common.hpp>
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/config.hpp>
@@ -31,14 +33,13 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/par_dfor.hpp>
 #include <migraphx/shape_for_each.hpp>
-#include <cmath>
-#include <utility>
+#include <migraphx/dyn_output.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-struct deconvolution
+struct convolution_backwards
 {
     std::vector<std::size_t> padding  = {0, 0};
     std::vector<std::size_t> stride   = {1, 1};
@@ -57,45 +58,91 @@ struct deconvolution
                     f(self.group, "group"));
     }
 
-    std::string name() const { return "deconvolution"; }
+    std::string name() const { return "convolution_backwards"; }
 
     void check_attribute_size() const
     {
-        if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
-           stride.size() != dilation.size())
+        if(padding.size() != stride.size() or stride.size() != dilation.size())
         {
-            MIGRAPHX_THROW("deconvolution: inconsistent attribute sizes");
+            MIGRAPHX_THROW("CONVOLUTION_BACKWARDS: inconsistent attribute sizes");
         }
     }
 
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(2).same_type().same_ndims().min_ndims(3);
+        check_shapes{inputs, *this, true}.has(2).same_type().same_ndims().min_ndims(3);
 
-        const shape& input   = inputs.at(0);
-        const shape& weights = inputs.at(1);
-        size_t kdims         = input.lens().size() - 2;
-        if(kdims != this->kdims())
+        const shape& x_shape = inputs.at(0);
+        const shape& w_shape = inputs.at(1);
+        if(x_shape.ndim() - 2 != this->kdims())
         {
-            MIGRAPHX_THROW("deconvolution: input k-dims does not match attribute size");
+            MIGRAPHX_THROW("CONVOLUTION_BACKWARDS: input k-dims does not match attribute size");
         }
 
-        std::vector<size_t> output_lens{input.lens()[0], weights.lens()[1]};
+        if(not x_shape.dynamic() and not w_shape.dynamic() and
+           x_shape.lens().at(1) != (w_shape.lens().at(0) * group))
+        {
+            MIGRAPHX_THROW("CONVOLUTION_BACKWARDS: mismatched channel numbers");
+        }
 
-        for(size_t i = 0; i < kdims; i++)
+        if(x_shape.dynamic() or w_shape.dynamic())
         {
-            output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
+            return dynamic_compute_shape(x_shape, w_shape);
+        }
+        else
+        {
+            return static_compute_shape(x_shape, w_shape);
+        }
+    }
+
+    std::vector<std::size_t> calc_spatial_lens(std::vector<std::size_t> x_lens,
+                                               std::vector<std::size_t> w_lens) const
+    {
+        std::vector<size_t> spatial_lens(x_lens.size() - 2);
+
+        // stride * (input - 1) + output_padding + ((kernel - 1) * dilation + 1) - padding_L -
+        // padding_R. This assumes padding_L = padding_R and output_padding handled in parser.
+        for(size_t i = 0; i < spatial_lens.size(); i++)
+        {
+            spatial_lens.at(i) = (std::size_t(std::max<std::ptrdiff_t>(
                 1,
-                stride[i] * (input.lens()[i + 2] - 1) +
-                    ((weights.lens()[i + 2] - 1) * dilation[i] + 1) - 2 * padding[i])));
+                stride[i] * (x_lens[i + 2] - 1) + ((w_lens[i + 2] - 1) * dilation[i] + 1) -
+                    2 * padding[i])));
         }
-        return inputs[0].with_lens(output_lens);
+        return spatial_lens;
+    }
+
+    shape dynamic_compute_shape(shape x_shape, shape w_shape) const
+    {
+        std::vector<shape::dynamic_dimension> output_dyn_dims = {};
+        output_dyn_dims.push_back(x_shape.to_dynamic().dyn_dims().at(0));
+        output_dyn_dims.push_back(w_shape.to_dynamic().dyn_dims().at(1));
+        const std::size_t num_spatial_dims = x_shape.ndim() - 2;
+        // Does not compute for optimals
+        auto min_spatial_dims = calc_spatial_lens(x_shape.min_lens(), w_shape.min_lens());
+        auto max_spatial_dims = calc_spatial_lens(x_shape.max_lens(), w_shape.max_lens());
+        for(size_t i = 0; i < num_spatial_dims; ++i)
+        {
+            output_dyn_dims.push_back(
+                shape::dynamic_dimension{min_spatial_dims[i], max_spatial_dims[i], {}});
+        }
+        return shape{x_shape.type(), output_dyn_dims};
+    }
+
+    shape static_compute_shape(shape x_shape, shape w_shape) const
+    {
+        std::vector<size_t> output_lens{x_shape.lens()[0], w_shape.lens()[1]};
+        auto spatial_lens = calc_spatial_lens(x_shape.lens(), w_shape.lens());
+        std::for_each(spatial_lens.begin(), spatial_lens.end(), [&output_lens](auto x) {
+            output_lens.push_back(x);
+        });
+        return x_shape.with_lens(output_lens);
     }
 
-    argument compute(shape output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        argument result{output_shape};
-        auto kdims = this->kdims();
+        argument result{dyn_out.computed_shape};
+        auto num_spatial_dims = this->kdims();
         visit_all(result, args[0], args[1])([&](auto output, auto input, auto weights) {
             using type = typename decltype(output)::value_type;
 
@@ -109,22 +156,22 @@ struct deconvolution
             auto wei_n = wei[0];
             auto wei_c = wei[1];
 
-            auto out_lens = output_shape.lens();
+            auto out_lens = dyn_out.computed_shape.lens();
 
             std::vector<std::size_t> win_size{in_c};
             std::copy(in_lens.begin() + 2, in_lens.end(), std::back_inserter(win_size));
             std::copy(wei.begin() + 2, wei.end(), std::back_inserter(win_size));
-            shape win_shape{output_shape.type(), win_size};
+            shape win_shape{dyn_out.computed_shape.type(), win_size};
 
             par_dfor(in_n, wei_c)([&](int o, int k) {
                 shape_for_each(win_shape, [&](auto idx_win) {
                     const int w = idx_win[0];
 
                     auto input_dims_start = idx_win.begin() + 1;
-                    auto wei_dims_start   = idx_win.begin() + kdims + 1;
+                    auto wei_dims_start   = idx_win.begin() + num_spatial_dims + 1;
 
                     std::vector<std::ptrdiff_t> win_start;
-                    for(std::size_t n = 0; n < kdims; ++n)
+                    for(std::size_t n = 0; n < num_spatial_dims; ++n)
                     {
                         win_start.push_back(std::ptrdiff_t(*(input_dims_start + n) * stride[n]) -
                                             std::ptrdiff_t(padding[n]));
@@ -135,7 +182,7 @@ struct deconvolution
 
                     std::vector<std::ptrdiff_t> idx_out{o, in_ch};
 
-                    for(size_t n = 0; n < kdims; n++)
+                    for(size_t n = 0; n < num_spatial_dims; n++)
                     {
                         idx_out.push_back(win_start[n] + *(wei_dims_start + n) * dilation[n]);
                     }

src/include/migraphx/op/dimensions_of.hpp

+/*
+ * 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.
+ */
+#ifndef MIGRAPHX_GUARD_OPERATORS_DIMENSIONS_OF_HPP
+#define MIGRAPHX_GUARD_OPERATORS_DIMENSIONS_OF_HPP
+
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/dyn_output.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+/**
+ * Returns the dimensions of the input argument from starting axis to ending axis.
+ * Atleast `end` must be set to use this operator (set `end` to ndim for default ONNX behavior of
+ * `Shape` operator) This should only be used for dynamic shapes as this can be simplified to a
+ * literal for static shapes.
+ */
+struct dimensions_of
+{
+    std::size_t start = 0;
+    std::size_t end   = 0;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.start, "start"), f(self.end, "end"));
+    }
+
+    std::string name() const { return "dimensions_of"; }
+
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        check_shapes{inputs, *this, true}.has(1);
+        if(start >= end)
+        {
+            MIGRAPHX_THROW("DIMENSIONS_OF: start >= end. start = " + std::to_string(start) +
+                           ", end = " + std::to_string(end));
+        }
+        return shape{shape::int64_type, {end - start}};
+    }
+
+    argument compute(const shape& output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        auto input_lens = args[0].get_shape().lens();
+        result.visit([&](auto output) {
+            std::copy(input_lens.cbegin() + start, input_lens.cbegin() + end, output.begin());
+        });
+        return result;
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif