feat(mgb/core): add comp node for cambricon

add testcase for cambricon comp node

GitOrigin-RevId: 7794faa47f

CMakeLists.txt

@@ -261,6 +261,7 @@ if(MGE_WITH_CUDA)
     set(MGE_CUDA_LIBS "${MGE_CUDA_LIBS}")
 endif()
 
+
 find_program(CCACHE_BIN ccache)
 if(CCACHE_BIN)
     set(CMAKE_CXX_COMPILER_LAUNCHER ${CCACHE_BIN})

dnn/src/CMakeLists.txt

@@ -56,4 +56,3 @@ target_link_libraries(megdnn ${MGE_BLAS_LIBS})
 if(CMAKE_THREAD_LIBS_INIT)
     target_link_libraries(megdnn Threads::Threads)
 endif()
-

dnn/src/common/megcore/common/device_context.cpp

@@ -16,7 +16,6 @@
 #include "src/cuda/megcore/cuda_device_context.hpp"
 #endif
 
-
 using namespace megcore;
 using namespace megdnn;
 

dnn/test/CMakeLists.txt

@@ -26,6 +26,7 @@ if(MGE_WITH_CUDA)
 endif()
 
 
+
 add_executable(megdnn_test ${SOURCES})
 target_link_libraries(megdnn_test gtest)
 target_link_libraries(megdnn_test megdnn)

python_module/test/run.sh

@@ -9,5 +9,6 @@ pushd $(dirname "${BASH_SOURCE[0]}")/.. >/dev/null
         --ignore test/unit/data \
         --ignore test/integration/manual \
         --ignore megengine/module/pytorch \
+        --ignore test/unit/module/test_external.py \
         megengine test
 popd >/dev/null

python_module/test/unit/module/.gitattributes

@@ -0,0 +1 @@
+*.mlu binary

src/CMakeLists.txt

@@ -31,6 +31,7 @@ if(MGE_WITH_CUDA AND MGE_WITH_TRT)
     list(APPEND SOURCES ${SOURCES_})
 endif()
 
+
 set(MGB_DEF ${MGB_DEF} PARENT_SCOPE)
 add_library(megbrain STATIC EXCLUDE_FROM_ALL ${SOURCES})
 target_link_libraries(megbrain mgb_opr_param_defs)

src/core/impl/comp_node_env.cpp

@@ -22,7 +22,6 @@
 #endif
 #endif
 
-
 using namespace mgb;
 
 /* =================== MegDNNHandle =================== */

src/core/include/megbrain/comp_node/alloc.h

@@ -200,7 +200,6 @@ class DevMemAlloc: virtual public MemAllocBase {
 #endif
 
 
-
         virtual ~DevMemAlloc() = default;
 
         /*!

src/core/include/megbrain/comp_node_env.h

@@ -41,7 +41,7 @@
         }                                                               \
     } while (0)
 
-#endif //MGB_ENABLE_LOGGING
+#endif  // MGB_ENABLE_LOGGING
 
 #endif
 

src/core/include/megbrain_build_config.h.in

@@ -97,6 +97,10 @@
 #endif
 
 
+#ifndef MGB_CAMBRICON
+#define MGB_CAMBRICON 0
+#endif
+
 // whether to enable TensorRT support
 #ifndef MGB_ENABLE_TENSOR_RT
 #define MGB_ENABLE_TENSOR_RT    MGB_CUDA

src/core/test/comp_node.cpp

@@ -49,7 +49,8 @@ TEST(TestCompNode, Parse) {
     ASSERT_EQ(L::parse("cpu2:23"), make_lc(D::CPU, 2, 23));
     ASSERT_EQ(L::parse("cpu21:23"), make_lc(D::CPU, 21, 23));
 
-    ASSERT_EQ(L::parse("xpu"), make_lc(D::UNSPEC, -1, 0)); 
+    
+    ASSERT_EQ(L::parse("xpu"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpux"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpu23"), make_lc(D::UNSPEC, 23, 0));
     ASSERT_EQ(L::parse("xpu23:1"), make_lc(D::UNSPEC, 23, 1));
@@ -70,6 +71,7 @@ TEST(TestCompNode, Parse) {
     ASSERT_THROW(L::parse("cpu2:23x"), MegBrainError);
     ASSERT_THROW(L::parse("heaxgon0"), MegBrainError);
     ASSERT_THROW(L::parse("rcom0"), MegBrainError);
+    ASSERT_THROW(L::parse("cmabricon0"), MegBrainError);
 }
 
 TEST(TestCompNode, SetDefaultDev) {
@@ -546,6 +548,7 @@ TEST(TestCompNode, MultipleLoad) {
     }
 }
 
+
 namespace {
 class CompNodeDepedentObjectInst final : public CompNodeDepedentObject {
     int *m_dst, *m_timer;

src/core/test/mem_alloc.cpp

@@ -464,6 +464,7 @@ public:
     }
     void raw_dev_free(void* ptr) override { MGB_CUDA_CHECK(cudaFree(ptr)); }
 };
+#endif
 
 using Callback = std::function<void()>;
 void test_free_mem(CompNode cn0, CompNode cn1, DevicePolicy* policy,
@@ -529,7 +530,7 @@ void test_gather_other(CompNode cn0, CompNode cn1) {
     opr::Sleep::sleep(cn1, 0.7);
     func->execute();
 }
-#endif
+
 }  // namespace
 
 #if MGB_CUDA
@@ -562,4 +563,5 @@ TEST(TestCudaMemAlloc, FreeMem) {
 }
 #endif  // MGB_CUDA
 
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

test/src/helper.cpp

@@ -70,8 +70,8 @@ dtype, RandomDistribution::GAUSSIAN>::operator ()(
     auto ptr = ret->ptr<ctype>();
     auto mean = m_mean, std = m_std;
     for (size_t i = 0, it = shape.total_nr_elems(); i < it; i += 2) {
-        ctype u1 = (m_rng() + 1.0) / (m_rng.max() + 1.0),
-              u2 = (m_rng() + 1.0) / (m_rng.max() + 1.0),
+        ctype u1 = ctype((m_rng() + 1.0) / (m_rng.max() + 1.0)),
+              u2 = ctype((m_rng() + 1.0) / (m_rng.max() + 1.0)),
               r = ctype(std * std::sqrt(-2 * std::log(u1))),
               theta = ctype(2 * M_PI * u2),
               z0 = ctype(r * std::cos(theta) + mean),
@@ -105,6 +105,8 @@ namespace mgb {
     template class HostTensorGenerator<
         dtype::Float32, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<
+        dtype::Float16, RandomDistribution::GAUSSIAN>;
+    template class HostTensorGenerator<
         dtype::Int8, RandomDistribution::UNIFORM>;
     template class HostTensorGenerator<
         dtype::Uint8, RandomDistribution::UNIFORM>;

test/src/include/megbrain/test/helper.h

@@ -400,6 +400,9 @@ bool check_gpu_available(size_t num);
 //! check whether given number of AMD GPUs is available
 bool check_amd_gpu_available(size_t num);
 
+//! check whether given number of cambricon devices is available
+bool check_cambricon_device_available(size_t num);
+
 //! check current capability >= major.minor
 bool check_compute_capability(int major, int minor);
 
@@ -436,6 +439,7 @@ public:
         return; \
 } while(0)
 
+
 #if MGB_HAVE_THREAD
 #define REQUIRE_THREAD()
 #else