feat: add Moore Threads platform support across runtime, device, build, and launcher stack

.github/pull_request_template.md

@@ -56,6 +56,7 @@ Please check all the platforms and/or backends this PR affects (i.e., code is to
 - [ ] NVIDIA GPU
 - [ ] Iluvatar GPU
 - [ ] MetaX GPU
+- [ ] Moore Threads GPU
 - [ ] Cambricon MLU
 
 ### Backend
@@ -102,6 +103,7 @@ See `CONTRIBUTING.md` § Pull Requests for the official testing requirements and
 - [ ] NVIDIA GPU
 - [ ] Iluvatar GPU
 - [ ] MetaX GPU
+- [ ] Moore Threads GPU
 - [ ] Cambricon MLU
 
 ### Test Involved Backend

CMakeLists.txt

@@ -10,6 +10,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 option(WITH_NVIDIA      "Enable NVIDIA GPU support"     OFF)
 option(WITH_ILUVATAR    "Enable ILUVATAR GPU support"   OFF)
 option(WITH_METAX       "Enable MetaX GPU support"      OFF)
+option(WITH_MOORE       "Enable Moore GPU support"      OFF)
 option(WITH_CAMBRICON   "Enable Cambricon MLU support"  OFF)
 
 set(WITH_CPU ON CACHE INTERNAL "CPU backend is always enabled")
@@ -114,6 +115,38 @@ if(AUTO_DETECT_DEVICES)
         endif()
     endif()
 
+    # MThreads (Moore)
+    set(MOORE_FOUND FALSE)
+
+    file(GLOB MOORE_DEV_FILES "/dev/mtgpu*")
+
+    if(DEFINED ENV{MUSA_ROOT} OR DEFINED ENV{MUSA_HOME} OR DEFINED ENV{MUSA_PATH})
+        set(MOORE_FOUND TRUE)
+    elseif(MOORE_DEV_FILES)
+        set(MOORE_FOUND TRUE)
+    elseif(NOT MOORE_FOUND)
+        find_program(MOORE_SMI_PATH mthreads-gmi)
+
+        if(MOORE_SMI_PATH)
+            execute_process(
+                COMMAND ${MOORE_SMI_PATH} -L
+                RESULT_VARIABLE SMI_RESULT
+                OUTPUT_QUIET
+                ERROR_QUIET
+            )
+            if(SMI_RESULT EQUAL 0)
+                set(MOORE_FOUND TRUE)
+            endif()
+        endif()
+    endif()
+
+    if(MOORE_FOUND)
+        set(WITH_MOORE ON)
+        message(STATUS "Auto-detected Moore environment.")
+    else()
+        message(STATUS "Moore environment not detected.")
+    endif()
+
     # Cambricon
     set(CAMBRICON_FOUND FALSE)
 
@@ -248,6 +281,40 @@ if(WITH_METAX)
     find_library(MACA_RUNTIME_LIB NAMES mcruntime HINTS "${MACA_PATH}/lib" REQUIRED)
 endif()
 
+if(WITH_MOORE)  
+    set(MUSA_ROOT "")
+    foreach(_musa_env MUSA_ROOT MUSA_HOME MUSA_PATH)
+        if(NOT MUSA_ROOT AND DEFINED ENV{${_musa_env}} AND NOT "$ENV{${_musa_env}}" STREQUAL "")
+            set(MUSA_ROOT "$ENV{${_musa_env}}")
+        endif()
+    endforeach()
+
+    if(NOT MUSA_ROOT AND EXISTS "/usr/local/musa")
+        set(MUSA_ROOT "/usr/local/musa")
+    endif()
+
+    if(NOT MUSA_ROOT)
+        message(FATAL_ERROR "`WITH_MOORE` is `ON` but `MUSA_ROOT`/`MUSA_HOME`/`MUSA_PATH` is not set and `/usr/local/musa` was not found.")
+    endif()
+
+    if(NOT EXISTS "${MUSA_ROOT}/bin/mcc")
+        message(FATAL_ERROR "Could not find `mcc` under `${MUSA_ROOT}/bin`.")
+    endif()
+
+    get_filename_component(MCC_WRAPPER_ABS "${CMAKE_CURRENT_SOURCE_DIR}/scripts/devices/mcc_wrapper.sh" ABSOLUTE)
+    file(CHMOD "${MCC_WRAPPER_ABS}"
+     PERMISSIONS OWNER_READ OWNER_WRITE OWNER_EXECUTE GROUP_READ GROUP_EXECUTE WORLD_READ WORLD_EXECUTE)
+    set(CMAKE_C_COMPILER   "${MCC_WRAPPER_ABS}")
+    set(CMAKE_CXX_COMPILER "${MCC_WRAPPER_ABS}")
+
+    include_directories("${MUSA_ROOT}/include")
+    link_directories("${MUSA_ROOT}/lib")
+
+    find_library(MUSA_LIB   NAMES musa   HINTS "${MUSA_ROOT}/lib" REQUIRED)
+    find_library(MUSART_LIB NAMES musart HINTS "${MUSA_ROOT}/lib" REQUIRED)
+    find_library(MUBLAS_LIB NAMES mublas HINTS "${MUSA_ROOT}/lib" REQUIRED)
+endif()
+
 if(WITH_CAMBRICON)
     set(NEUWARE_HOME $ENV{NEUWARE_HOME})
 

README.md

@@ -135,6 +135,7 @@ cmake .. -DWITH_NVIDIA=ON -DWITH_OMPI=ON
 | `WITH_NVIDIA`     | Enable NVIDIA GPU support | `OFF` |
 | `WITH_ILUVATAR`   | Enable Iluvatar GPU support | `OFF` |
 | `WITH_METAX`      | Enable MetaX GPU support  | `OFF` |
+| `WITH_MOORE`      | Enable Moore Threads GPU support  | `OFF` |
 | `WITH_CAMBRICON`  | Enable Cambricon MLU support  | `OFF` |
 | `WITH_CPU`        | CPU support (always enabled) | `ON` (internal, not user‑settable) |
 | **Backend (Communication) Options** |||
@@ -282,6 +283,7 @@ export LD_LIBRARY_PATH=${INFINI_INSTALL}/lib:$LD_LIBRARY_PATH
 | **NVIDIA**     | Full | Requires CUDA Toolkit. |
 | **Iluvatar**   | Full | Requires Iluvatar CoreX SDK. |
 | **MetaX**      | Full | Requires MACA SDK and `MACA_PATH` (default `/opt/maca`) to be set. |
+| **Moore Threads** | Full | Requires MUSA SDK and at least one of `MACA_ROOT` (default `/usr/local/musa`), `MACA_PATH`, and `MUSA_HOME` to be set. |
 | **Cambricon**  | Full | Requires CNToolKit and `NEUWARE_HOME` to be set. |
 
 </details>

examples/CMakeLists.txt

@@ -16,6 +16,11 @@ foreach(source_file ${EXAMPLE_SOURCES})
         target_link_libraries(${example_name} PRIVATE ${MACA_RUNTIME_LIB})
     endif()
 
+    if(WITH_MOORE)
+        target_link_libraries(${example_name} PRIVATE ${MUSART_LIB})
+        target_compile_options(${example_name} PRIVATE "-x" "musa")
+    endif()
+
     if(WITH_CAMBRICON)
         target_link_libraries(${example_name} PRIVATE ${CAMBRICON_RUNTIME_LIB})
     endif()

scripts/devices/mcc_wrapper.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Filter out flags unsupported by `mcc`.
+ARGS=()
+skip_next=0
+linking=1
+for arg in "$@"; do
+    if [ $skip_next -eq 1 ]; then
+        skip_next=0
+        continue
+    fi
+    case "$arg" in
+        -c|-E|-S)
+            linking=0
+            ARGS+=("$arg")
+            ;;
+        -pthread)
+            ;;
+        -B)
+            skip_next=1
+            ;;
+        -B*)
+            ;;
+        *)
+            ARGS+=("$arg")
+            ;;
+    esac
+done
+
+MUSA_ROOT_DIR="${MUSA_ROOT:-${MUSA_HOME:-${MUSA_PATH:-/usr/local/musa}}}"
+
+if command -v g++ >/dev/null 2>&1; then
+    GXX_MAJOR="$(g++ -dumpversion | cut -d. -f1)"
+    if [ -d "/usr/include/c++/${GXX_MAJOR}" ]; then
+        ARGS=(
+            "-isystem" "/usr/include/c++/${GXX_MAJOR}"
+            "-isystem" "/usr/include/x86_64-linux-gnu/c++/${GXX_MAJOR}"
+            "-isystem" "/usr/include/c++/${GXX_MAJOR}/backward"
+            "${ARGS[@]}"
+        )
+    fi
+
+    STDCPP_LIB="$(g++ -print-file-name=libstdc++.so)"
+    if [ $linking -eq 1 ] && [ -f "${STDCPP_LIB}" ]; then
+        ARGS=("-L$(dirname "${STDCPP_LIB}")" "${ARGS[@]}")
+    fi
+fi
+
+exec "${MUSA_ROOT_DIR}/bin/mcc" "${ARGS[@]}"

scripts/icclrun_logic.py

@@ -124,6 +124,11 @@ def ensure_launcher_exists(self):
                     'grep -l "9999" /sys/bus/pci/devices/*/vendor >/dev/null 2>&1'
                 )
 
+            elif n_type == "moore":
+                condition = (
+                    '[ -c "/dev/mtgpu.0" ] || [ -x "$(command -v mthreads-gmi)" ]'
+                )
+
             elif n_type == "cambricon":
                 condition = (
                     '[ -n "${NEUWARE_HOME}" ] || command -v cnmon >/dev/null 2>&1'

scripts/run_wrapper.sh

@@ -13,6 +13,8 @@ if [ -c "/dev/nvidia0" ] || [ -x "$(command -v nvidia-smi)" ]; then
     ARCH="nvidia"
 elif grep -l "9999" /sys/bus/pci/devices/*/vendor >/dev/null 2>&1 || [ -d "/opt/maca" ]; then
     ARCH="metax"
+elif [ -c "/dev/mtgpu.0" ] || [ -x "$(command -v mthreads-gmi)" ]; then
+    ARCH="moore"
 elif [ -n "${NEUWARE_HOME}" ] || [ -x "$(command -v cnmon)" ]; then
     ARCH="cambricon"
 else

src/CMakeLists.txt

@@ -98,6 +98,29 @@ if(WITH_METAX)
     target_compile_options(infiniccl PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-x maca>)
 endif()
 
+# Mthreads (Moore)
+if(WITH_MOORE)
+    list(APPEND DEVICE_LIST "moore")
+
+    set(MOORE_PATTERNS
+        "cuda/*.cc"
+        "cuda/*.cpp"
+        "moore/*.cc"
+        "moore/*.cpp"
+        "moore/*.mu"
+    )
+    file(GLOB MOORE_SOURCES ${MOORE_PATTERNS})
+
+    set_source_files_properties(${MOORE_SOURCES} PROPERTIES LANGUAGE CXX)
+
+    # target_compile_options(infiniccl PRIVATE "-x" "musa")
+    target_sources(infiniccl PRIVATE ${MOORE_SOURCES})
+
+    target_include_directories(infiniccl PRIVATE "${MUSA_ROOT}/include")
+    target_link_libraries(infiniccl PRIVATE ${MUSA_LIB} ${MUSART_LIB} ${MUBLAS_LIB})
+    target_compile_options(infiniccl PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-x musa>)
+endif()
+
 # Cambricon
 if(WITH_CAMBRICON)
     list(APPEND DEVICE_LIST "cambricon")

src/device.h

@@ -145,6 +145,11 @@ struct DevicePriority<Device::Type::kMetax> {
   static constexpr int value = 5;
 };
 
+template <>
+struct DevicePriority<Device::Type::kMoore> {
+  static constexpr int value = 5;
+};
+
 template <>
 struct DevicePriority<Device::Type::kCambricon> {
   static constexpr int value = 5;

src/moore/data_type_.h

@@ -0,0 +1,26 @@
+#ifndef INFINI_CCL_MOORE_DATA_TYPE__H_
+#define INFINI_CCL_MOORE_DATA_TYPE__H_
+
+// clang-format off
+#include <musa_bf16.h>
+#include <musa_fp16.h>
+// clang-format on
+
+#include "data_type_impl.h"
+#include "moore/device_.h"
+
+namespace infini::ccl {
+
+template <>
+struct TypeMap<Device::Type::kMoore, DataType::kFloat16> {
+  using type = half;
+};
+
+template <>
+struct TypeMap<Device::Type::kMoore, DataType::kBFloat16> {
+  using type = __mt_bfloat16;
+};
+
+}  // namespace infini::ccl
+
+#endif  // INFINI_CCL_MOORE_DATA_TYPE__H_

src/moore/device_.h

@@ -0,0 +1,13 @@
+#ifndef INFINI_CCL_MOORE_DEVICE__H_
+#define INFINI_CCL_MOORE_DEVICE__H_
+
+#include "device.h"
+
+namespace infini::ccl {
+
+template <>
+struct DeviceEnabled<Device::Type::kMoore> : std::true_type {};
+
+}  // namespace infini::ccl
+
+#endif  // INFINI_CCL_MOORE_DEVICE__H_

src/moore/runtime_.h

@@ -0,0 +1,62 @@
+#ifndef INFINI_CCL_MOORE_RUNTIME_H_
+#define INFINI_CCL_MOORE_RUNTIME_H_
+
+#include <utility>
+
+// clang-format off
+#include <musa_runtime.h>
+// clang-format on
+
+#include "cuda/runtime_.h"
+#include "logging.h"
+#include "moore/device_.h"
+#include "return_status_impl.h"
+
+namespace infini::ccl {
+
+template <>
+struct Runtime<Device::Type::kMoore>
+    : CudaRuntime<Runtime<Device::Type::kMoore>> {
+  using Stream = musaStream_t;
+
+  static constexpr Device::Type kDeviceType = Device::Type::kMoore;
+
+  static constexpr auto Check =
+      [](auto status, ReturnStatus err_code = ReturnStatus::kSystemError) {
+        if (status != musaSuccess) {
+          LOG(musaGetErrorString(static_cast<musaError_t>(status)));
+          return err_code;
+        }
+        return ReturnStatus::kSuccess;
+      };
+
+  static constexpr auto Malloc = [](auto &&...args) {
+    return musaMalloc(std::forward<decltype(args)>(args)...);
+  };
+
+  static constexpr auto Memcpy = musaMemcpy;
+
+  static constexpr auto Free = [](auto &&...args) {
+    return musaFree(std::forward<decltype(args)>(args)...);
+  };
+
+  static constexpr auto MemcpyHostToDevice = musaMemcpyHostToDevice;
+
+  static constexpr auto MemcpyDeviceToHost = musaMemcpyDeviceToHost;
+
+  static constexpr auto Memset = musaMemset;
+
+  static constexpr auto SetDevice = musaSetDevice;
+
+  static constexpr auto DeviceSynchronize = [](auto &&...args) {
+    return musaDeviceSynchronize(std::forward<decltype(args)>(args)...);
+  };
+
+  static constexpr auto StreamSynchronize = musaStreamSynchronize;
+};
+
+static_assert(Runtime<Device::Type::kMoore>::Validate());
+
+}  // namespace infini::ccl
+
+#endif  // INFINI_CCL_MOORE_RUNTIME_H_