feat(metax): add graph capture and replay support

CMakeLists.txt

@@ -34,6 +34,7 @@ option(WITH_NVIDIA "Enable CUDA backend" OFF)
 option(WITH_ILUVATAR "Enable Iluvatar GPU backend" OFF)
 option(WITH_HYGON "Enable Hygon GPU backend" OFF)
 option(WITH_METAX "Enable MetaX backend" OFF)
+option(WITH_MARS "Enable Mars backend" OFF)
 option(WITH_CAMBRICON "Enable Cambricon backend" OFF)
 option(WITH_MOORE "Enable Moore backend" OFF)
 option(WITH_ASCEND "Enable Ascend backend" OFF)
@@ -76,9 +77,48 @@ if(AUTO_DETECT_DEVICES)
         endif()
     endif()
 
-    if(DEFINED ENV{MACA_PATH})
+    set(_infinirt_hpcc_env_root "")
+    if(DEFINED ENV{HPCC_PATH} AND
+       EXISTS "$ENV{HPCC_PATH}/include/hcr/hc_runtime.h")
+        set(_infinirt_hpcc_env_root "$ENV{HPCC_PATH}")
+    endif()
+
+    set(_infinirt_maca_env_root "")
+    if(DEFINED ENV{MACA_PATH} AND
+       EXISTS "$ENV{MACA_PATH}/include/mcr/mc_runtime.h")
+        set(_infinirt_maca_env_root "$ENV{MACA_PATH}")
+    endif()
+
+    set(_infinirt_default_hpcc_available FALSE)
+    if(EXISTS "/opt/hpcc/include/hcr/hc_runtime.h")
+        set(_infinirt_default_hpcc_available TRUE)
+    endif()
+
+    set(_infinirt_default_maca_available FALSE)
+    if(EXISTS "/opt/maca/include/mcr/mc_runtime.h")
+        set(_infinirt_default_maca_available TRUE)
+    endif()
+
+    if(WITH_MARS OR WITH_METAX)
+        message(STATUS "Skipping MetaX-family auto-detection because a backend was explicitly enabled")
+    elseif(_infinirt_hpcc_env_root AND _infinirt_maca_env_root)
+        message(FATAL_ERROR
+                "Both HPCC_PATH and MACA_PATH identify valid SDKs. Select Mars with -DWITH_MARS=ON or MetaX with -DWITH_METAX=ON.")
+    elseif(_infinirt_hpcc_env_root)
+        set(WITH_MARS ON)
+        message(STATUS "Auto-detected Mars HPCC environment from HPCC_PATH")
+    elseif(_infinirt_maca_env_root)
         set(WITH_METAX ON)
-        message(STATUS "Auto-detected MetaX environment from MACA_PATH")
+        message(STATUS "Auto-detected MetaX MACA environment from MACA_PATH")
+    elseif(_infinirt_default_hpcc_available AND _infinirt_default_maca_available)
+        message(FATAL_ERROR
+                "Both /opt/hpcc and /opt/maca contain valid SDKs. Select Mars with -DWITH_MARS=ON or MetaX with -DWITH_METAX=ON.")
+    elseif(_infinirt_default_hpcc_available)
+        set(WITH_MARS ON)
+        message(STATUS "Auto-detected Mars HPCC environment from /opt/hpcc")
+    elseif(_infinirt_default_maca_available)
+        set(WITH_METAX ON)
+        message(STATUS "Auto-detected MetaX MACA environment from /opt/maca")
     else()
         execute_process(
             COMMAND sh -c "grep -h 9999 /sys/bus/pci/devices/*/vendor 2>/dev/null"
@@ -89,8 +129,7 @@ if(AUTO_DETECT_DEVICES)
         string(FIND "${_pci_vendor_output}" "9999" _found_pos)
 
         if(_found_pos GREATER -1)
-            set(WITH_METAX ON)
-            message(STATUS "Detected MetaX GPU from PCI vendor ID 0x9999")
+            message(STATUS "Detected a MetaX-family GPU from PCI vendor ID 0x9999, but no MACA or HPCC SDK was found")
         else()
             set(WITH_METAX OFF)
             message(STATUS "No MetaX GPU detected")
@@ -118,7 +157,7 @@ if(AUTO_DETECT_DEVICES)
 
     if(WITH_NVIDIA)
         set(_non_nvidia_gpu_detected FALSE)
-        foreach(_gpu_backend WITH_ILUVATAR WITH_HYGON WITH_METAX WITH_CAMBRICON WITH_MOORE WITH_ASCEND)
+        foreach(_gpu_backend WITH_ILUVATAR WITH_HYGON WITH_METAX WITH_MARS WITH_CAMBRICON WITH_MOORE WITH_ASCEND)
             if(${_gpu_backend})
                 set(_non_nvidia_gpu_detected TRUE)
             endif()
@@ -135,14 +174,14 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/src)
 
 # Only one CUDA-like GPU backend can be enabled at a time.
 set(_gpu_backend_count 0)
-foreach(_gpu_backend WITH_NVIDIA WITH_ILUVATAR WITH_HYGON WITH_METAX WITH_MOORE WITH_ASCEND)
+foreach(_gpu_backend WITH_NVIDIA WITH_ILUVATAR WITH_HYGON WITH_METAX WITH_MARS WITH_MOORE WITH_ASCEND)
     if(${_gpu_backend})
         math(EXPR _gpu_backend_count "${_gpu_backend_count} + 1")
     endif()
 endforeach()
 
 if(_gpu_backend_count GREATER 1)
-    message(FATAL_ERROR "`WITH_NVIDIA`, `WITH_ILUVATAR`, `WITH_HYGON`, `WITH_METAX`, `WITH_MOORE`, and `WITH_ASCEND` are mutually exclusive. Build one GPU backend at a time.")
+    message(FATAL_ERROR "`WITH_NVIDIA`, `WITH_ILUVATAR`, `WITH_HYGON`, `WITH_METAX`, `WITH_MARS`, `WITH_MOORE`, and `WITH_ASCEND` are mutually exclusive. Build one GPU backend at a time.")
 endif()
 
 if(WITH_NVIDIA)
@@ -220,18 +259,38 @@ endif()
 if(WITH_METAX)
     add_compile_definitions(WITH_METAX=1)
 
-    # Normally can be found at: `/opt/maca/`.
-    set(MACA_PATH $ENV{MACA_PATH})
-    set(CMAKE_C_COMPILER   ${CMAKE_CURRENT_SOURCE_DIR}/scripts/mxcc_wrapper.sh)
+    set(MACA_PATH "$ENV{MACA_PATH}" CACHE PATH "Path to the MetaX MACA SDK")
+    if(NOT MACA_PATH)
+        set(MACA_PATH "/opt/maca")
+    endif()
+    if(NOT EXISTS "${MACA_PATH}/include/mcr/mc_runtime.h")
+        message(FATAL_ERROR "`WITH_METAX` requires a MACA SDK. Set MACA_PATH to a root containing include/mcr/mc_runtime.h.")
+    endif()
     set(CMAKE_CXX_COMPILER ${CMAKE_CURRENT_SOURCE_DIR}/scripts/mxcc_wrapper.sh)
+    message(STATUS "MetaX: using MACA SDK at ${MACA_PATH}")
 
     include_directories("${MACA_PATH}/include")
     link_directories("${MACA_PATH}/lib")
-
-    # Libraries: mcruntime / mcdnn / mcblas.
     find_library(MACA_RUNTIME_LIB NAMES mcruntime HINTS "${MACA_PATH}/lib" REQUIRED)
 endif()
 
+if(WITH_MARS)
+    add_compile_definitions(WITH_MARS=1)
+
+    set(HPCC_PATH "$ENV{HPCC_PATH}" CACHE PATH "Path to the Mars HPCC SDK")
+    if(NOT HPCC_PATH)
+        set(HPCC_PATH "/opt/hpcc")
+    endif()
+    if(NOT EXISTS "${HPCC_PATH}/include/hcr/hc_runtime.h")
+        message(FATAL_ERROR "`WITH_MARS` requires an HPCC SDK. Set HPCC_PATH to a root containing include/hcr/hc_runtime.h.")
+    endif()
+    message(STATUS "Mars: using HPCC SDK at ${HPCC_PATH}")
+
+    include_directories("${HPCC_PATH}/include")
+    link_directories("${HPCC_PATH}/lib")
+    find_library(HPCC_RUNTIME_LIB NAMES hcruntime HINTS "${HPCC_PATH}/lib" REQUIRED)
+endif()
+
 if(WITH_MOORE)
     add_compile_definitions(WITH_MOORE=1)
 
@@ -294,7 +353,7 @@ if(WITH_ASCEND)
 endif()
 
 # If all other platforms are not enabled, CPU is enabled by default.
-if(NOT WITH_NVIDIA AND NOT WITH_ILUVATAR AND NOT WITH_HYGON AND NOT WITH_METAX AND NOT WITH_MOORE AND NOT WITH_CAMBRICON AND NOT WITH_ASCEND)
+if(NOT WITH_NVIDIA AND NOT WITH_ILUVATAR AND NOT WITH_HYGON AND NOT WITH_METAX AND NOT WITH_MARS AND NOT WITH_MOORE AND NOT WITH_CAMBRICON AND NOT WITH_ASCEND)
     set(WITH_CPU ON)
     add_compile_definitions(WITH_CPU=1)
 endif()
@@ -315,6 +374,9 @@ endif()
 if(WITH_METAX)
     list(APPEND INFINI_RT_PUBLIC_HEADER_DEVICES metax)
 endif()
+if(WITH_MARS)
+    list(APPEND INFINI_RT_PUBLIC_HEADER_DEVICES mars)
+endif()
 if(WITH_MOORE)
     list(APPEND INFINI_RT_PUBLIC_HEADER_DEVICES moore)
 endif()
@@ -347,7 +409,7 @@ if(WITH_NVIDIA OR WITH_ILUVATAR OR WITH_HYGON)
 endif()
 
 set(INFINI_RT_PACKAGE_VENDOR_BACKEND "")
-foreach(_backend METAX MOORE CAMBRICON ASCEND)
+foreach(_backend METAX MARS MOORE CAMBRICON ASCEND)
     if(WITH_${_backend})
         set(INFINI_RT_PACKAGE_VENDOR_BACKEND "${_backend}")
     endif()

CONTRIBUTING.md

@@ -117,6 +117,7 @@ Common backend options are:
 -DWITH_ILUVATAR=ON
 -DWITH_HYGON=ON
 -DWITH_METAX=ON
+-DWITH_MARS=ON
 -DWITH_MOORE=ON
 -DWITH_CAMBRICON=ON
 -DWITH_ASCEND=ON
@@ -200,7 +201,8 @@ ruff check .
 
 1. **CMake cannot find a backend SDK**: Check the backend-specific environment
    variables documented in `docs/backends.md`, such as `DTK_ROOT`,
-   `MACA_PATH`, `MUSA_ROOT`, `NEUWARE_HOME`, or `ASCEND_HOME_PATH`.
+   `MACA_PATH`, `HPCC_PATH`, `MUSA_ROOT`, `NEUWARE_HOME`, or
+   `ASCEND_HOME_PATH`.
 2. **Switching between backends gives stale build errors**: Use a separate
    build directory per platform or delete the stale build directory.
 3. **Generated headers are stale**: Re-run CMake configure. Public headers are

cmake/InfiniRTConfig.cmake.in

@@ -10,6 +10,7 @@ set(InfiniRT_WITH_NVIDIA @WITH_NVIDIA@)
 set(InfiniRT_WITH_ILUVATAR @WITH_ILUVATAR@)
 set(InfiniRT_WITH_HYGON @WITH_HYGON@)
 set(InfiniRT_WITH_METAX @WITH_METAX@)
+set(InfiniRT_WITH_MARS @WITH_MARS@)
 set(InfiniRT_WITH_MOORE @WITH_MOORE@)
 set(InfiniRT_WITH_CAMBRICON @WITH_CAMBRICON@)
 set(InfiniRT_WITH_ASCEND @WITH_ASCEND@)
@@ -34,6 +35,7 @@ if(_InfiniRT_VENDOR_BACKEND STREQUAL "METAX")
     if(DEFINED ENV{MACA_PATH})
         list(APPEND _InfiniRT_ROOT_HINTS "$ENV{MACA_PATH}")
     endif()
+    list(APPEND _InfiniRT_ROOT_HINTS /opt/maca)
     find_path(_InfiniRT_BACKEND_INCLUDE_DIR
         NAMES mcr/mc_runtime.h
         HINTS ${_InfiniRT_ROOT_HINTS}
@@ -52,6 +54,30 @@ if(_InfiniRT_VENDOR_BACKEND STREQUAL "METAX")
     _infinirt_define_backend_runtime(
         "${_InfiniRT_BACKEND_INCLUDE_DIR}"
         "${_InfiniRT_BACKEND_RUNTIME_LIBRARY}")
+elseif(_InfiniRT_VENDOR_BACKEND STREQUAL "MARS")
+    set(_InfiniRT_ROOT_HINTS "${InfiniRT_MARS_ROOT}")
+    if(DEFINED ENV{HPCC_PATH})
+        list(APPEND _InfiniRT_ROOT_HINTS "$ENV{HPCC_PATH}")
+    endif()
+    list(APPEND _InfiniRT_ROOT_HINTS /opt/hpcc)
+    find_path(_InfiniRT_BACKEND_INCLUDE_DIR
+        NAMES hcr/hc_runtime.h
+        HINTS ${_InfiniRT_ROOT_HINTS}
+        PATH_SUFFIXES include)
+    find_library(_InfiniRT_BACKEND_RUNTIME_LIBRARY
+        NAMES hcruntime
+        HINTS ${_InfiniRT_ROOT_HINTS}
+        PATH_SUFFIXES lib lib64)
+    if(NOT _InfiniRT_BACKEND_INCLUDE_DIR OR
+       NOT _InfiniRT_BACKEND_RUNTIME_LIBRARY)
+        set(InfiniRT_FOUND FALSE)
+        set(InfiniRT_NOT_FOUND_MESSAGE
+            "Mars HPCC SDK not found. Set InfiniRT_MARS_ROOT or HPCC_PATH.")
+        return()
+    endif()
+    _infinirt_define_backend_runtime(
+        "${_InfiniRT_BACKEND_INCLUDE_DIR}"
+        "${_InfiniRT_BACKEND_RUNTIME_LIBRARY}")
 elseif(_InfiniRT_VENDOR_BACKEND STREQUAL "MOORE")
     set(_InfiniRT_ROOT_HINTS "${InfiniRT_MOORE_ROOT}")
     foreach(_InfiniRT_ENV MUSA_ROOT MUSA_HOME MUSA_PATH)

docs/api/core-types.md

@@ -22,6 +22,7 @@ Known device types include:
 - `Device::Type::kNvidia`
 - `Device::Type::kIluvatar`
 - `Device::Type::kMetax`
+- `Device::Type::kMars`
 - `Device::Type::kMoore`
 - `Device::Type::kHygon`
 - `Device::Type::kCambricon`

docs/backends.md

@@ -12,6 +12,7 @@ backend.
 | NVIDIA | `WITH_NVIDIA` | Requires CUDA toolkit. |
 | Iluvatar | `WITH_ILUVATAR` | CUDA-compatible backend using clang with ivcore flags. |
 | MetaX | `WITH_METAX` | Requires `MACA_PATH`. |
+| Mars | `WITH_MARS` | Requires the HPCC SDK from `HPCC_PATH` or `/opt/hpcc`. |
 | Moore | `WITH_MOORE` | Requires `MUSA_ROOT`, `MUSA_HOME`, or `MUSA_PATH`. |
 | Hygon | `WITH_HYGON` | Requires DTK and a DTK CUDA toolkit. |
 | Cambricon | `WITH_CAMBRICON` | Requires `NEUWARE_HOME`. |
@@ -20,6 +21,9 @@ backend.
 Only one accelerator backend can be enabled at a time. CPU can be enabled
 together with the selected accelerator backend.
 
+When both HPCC and MACA SDKs are installed, select `WITH_MARS` or `WITH_METAX`
+explicitly. Automatic detection rejects an ambiguous MetaX-family environment.
+
 ## Runtime API Support
 
 All enabled runtime backends provide the core device and memory entry points
@@ -33,7 +37,8 @@ Current test expectations are:
 | CPU | No | Yes | No | Yes | No | Yes | No |
 | NVIDIA | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
 | Iluvatar | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
-| MetaX | Yes | Yes | Yes | Yes | Yes | Yes | No |
+| MetaX | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
+| Mars | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
 | Moore | Yes | Yes | No | Yes | Yes | Yes | No |
 | Hygon | Yes | Yes | Yes | Yes | Yes | Yes | No |
 | Cambricon | Yes | No | No | No | No | No | No |

docs/build.md

@@ -9,6 +9,7 @@ InfiniRT uses CMake and C++17.
 -DWITH_NVIDIA=ON
 -DWITH_ILUVATAR=ON
 -DWITH_METAX=ON
+-DWITH_MARS=ON
 -DWITH_MOORE=ON
 -DWITH_HYGON=ON
 -DWITH_CAMBRICON=ON

scripts/generate_public_headers.py

@@ -2,6 +2,7 @@
 import dataclasses
 import pathlib
 import re
+import shutil
 
 
 _DETAIL_PREFIX = "infini/rt/detail"
@@ -28,9 +29,14 @@
         ("hygon", "runtime_.h", "native/cuda/hygon/runtime_.h"),
     ),
     "metax": (
-        ("metax", "data_type_.h", "native/cuda/metax/data_type_.h"),
-        ("metax", "device_.h", "native/cuda/metax/device_.h"),
-        ("metax", "runtime_.h", "native/cuda/metax/runtime_.h"),
+        ("metax", "data_type_.h", "native/maca/metax/data_type_.h"),
+        ("metax", "device_.h", "native/maca/metax/device_.h"),
+        ("metax", "runtime_.h", "native/maca/metax/runtime_.h"),
+    ),
+    "mars": (
+        ("mars", "data_type_.h", "native/hpcc/mars/data_type_.h"),
+        ("mars", "device_.h", "native/hpcc/mars/device_.h"),
+        ("mars", "runtime_.h", "native/hpcc/mars/runtime_.h"),
     ),
     "moore": (
         ("moore", "data_type_.h", "native/cuda/moore/data_type_.h"),
@@ -55,6 +61,7 @@
     "iluvatar": "Device::Type::kIluvatar",
     "hygon": "Device::Type::kHygon",
     "metax": "Device::Type::kMetax",
+    "mars": "Device::Type::kMars",
     "moore": "Device::Type::kMoore",
     "cambricon": "Device::Type::kCambricon",
     "ascend": "Device::Type::kAscend",
@@ -65,6 +72,7 @@
     "iluvatar",
     "hygon",
     "metax",
+    "mars",
     "moore",
     "cambricon",
     "ascend",
@@ -663,6 +671,9 @@ def main():
     include_root = pathlib.Path(args.output_dir)
     source_root = pathlib.Path(args.runtime_header).parent
 
+    if include_root.exists():
+        shutil.rmtree(include_root)
+
     _write_detail_headers(include_root, source_root, devices)
     for device in devices:
         for wrapper_device, header_name, target in _DEVICE_HEADERS[device]:

scripts/run_performance_tests.py

@@ -12,6 +12,7 @@
     ("WITH_ILUVATAR", "iluvatar"),
     ("WITH_HYGON", "hygon"),
     ("WITH_METAX", "metax"),
+    ("WITH_MARS", "mars"),
     ("WITH_MOORE", "moore"),
     ("WITH_CAMBRICON", "cambricon"),
     ("WITH_ASCEND", "ascend"),

src/CMakeLists.txt

@@ -64,12 +64,19 @@ if(WITH_HYGON)
 endif()
 
 if(WITH_METAX)
-    target_compile_definitions(infinirt PRIVATE WITH_METAX=1)
+    target_compile_definitions(infinirt PUBLIC WITH_METAX=1)
     _infinirt_use_backend_runtime(
         "${MACA_PATH}/include"
         "${MACA_RUNTIME_LIB}")
 endif()
 
+if(WITH_MARS)
+    target_compile_definitions(infinirt PUBLIC WITH_MARS=1)
+    _infinirt_use_backend_runtime(
+        "${HPCC_PATH}/include"
+        "${HPCC_RUNTIME_LIB}")
+endif()
+
 if(WITH_MOORE)
     target_compile_definitions(infinirt PRIVATE WITH_MOORE=1)
     _infinirt_use_backend_runtime(

src/device.h

@@ -21,6 +21,7 @@ class Device {
     kMoore = 5,
     kIluvatar = 6,
     kHygon = 7,
+    kMars = 8,
     kCount
   };
 
@@ -64,6 +65,7 @@ class Device {
           {Type::kMoore, "moore"},
           {Type::kIluvatar, "iluvatar"},
           {Type::kHygon, "hygon"},
+          {Type::kMars, "mars"},
       }}};
 
   static constexpr ConstexprMap<std::string_view, Device::Type,
@@ -77,6 +79,7 @@ class Device {
           {"moore", Type::kMoore},
           {"iluvatar", Type::kIluvatar},
           {"hygon", Type::kHygon},
+          {"mars", Type::kMars},
       }}};
 
   int index_{0};
@@ -91,7 +94,7 @@ struct DeviceEnabled : std::false_type {};
 using AllDeviceTypes =
     List<Device::Type::kCpu, Device::Type::kNvidia, Device::Type::kCambricon,
          Device::Type::kAscend, Device::Type::kMetax, Device::Type::kMoore,
-         Device::Type::kIluvatar, Device::Type::kHygon>;
+         Device::Type::kIluvatar, Device::Type::kHygon, Device::Type::kMars>;
 
 // Deferred computation of active devices. The `Filter` and `FilterList`
 // evaluation are nested inside a class template so that `DeviceEnabled`