add conv_transpose&&native maxpooling

CMakeLists.txt

@@ -20,6 +20,7 @@ endif()
 include(CMakeDependentOption)
 project(InfiniTensor C CXX)
 
+
 cmake_dependent_option(BUILD_TEST_CORE "Build tests for core components" ON BUILD_TEST OFF)
 cmake_dependent_option(BUILD_TEST_PET "Build tests for PET" OFF BUILD_TEST OFF)
 

Makefile

@@ -39,7 +39,7 @@ endif
 
 build:
 	mkdir -p build/$(TYPE)
-	cd build/$(TYPE) && cmake $(CMAKE_OPT) ../.. && make -j8
+	cd build/$(TYPE) && cmake $(CMAKE_OPT) ../.. && make -j
 
 clean:
 	rm -rf build

include/ascend/ascend_runtime.h

@@ -20,21 +20,16 @@ class ASCENDRuntimeObj : public RuntimeObj {
   private:
     aclrtContext context;
     aclrtStream stream;
+    std::unique_ptr<CommunicatorObj> comm;
     ASCENDPtr workspace = nullptr;
     size_t workspaceSize;
 
   public:
     ASCENDRuntimeObj(int deviceId = 0) : RuntimeObj(Device::ASCEND, deviceId) {
-        // #ifndef _ACL_INIT
+        auto ret = aclInit(nullptr);
-        // #define _ACL_INIT
+        CHECK_RET(ret == ACL_SUCCESS,
-        //         aclInit(nullptr);
+                  LOG_PRINT("aclInit failed. ERROR: %d\n", ret));
-        //         //  auto ret_init =
+        ret = aclrtSetDevice(deviceId);
-        //         //   CHECK_RET(ret == ACL_SUCCESS,
-        //         //             LOG_PRINT("aclInit failed. ERROR: %d\n",
-        //         ret));
-        // #endif
-        aclInit(nullptr);
-        auto ret = aclrtSetDevice(deviceId);
         CHECK_RET(ret == ACL_SUCCESS,
                   LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret));
         ret = aclrtCreateContext(&context, deviceId);
@@ -49,7 +44,7 @@ class ASCENDRuntimeObj : public RuntimeObj {
 
         // 10GB for Longformer
         // size_t longformerNum = 3lu * (1 << 30);
-        workspaceSize = 3ll << 30; // 3 GB
+        workspaceSize = 3ll << 33; // 3 GB
         // std::cout<<workspaceSize/1024/1024/1024<< std::endl;
         // std::cout<<std::bitset<64>(workspaceSize)<< std::endl;
         workspace = alloc(workspaceSize);
@@ -99,9 +94,9 @@ class ASCENDRuntimeObj : public RuntimeObj {
                     ACL_MEMCPY_DEVICE_TO_DEVICE);
     }
 
-    void initComm(const string &, int, int) override { IT_TODO_HALT(); }
+    void initComm(const string &name, int worldSize, int rank) final;
 
-    CommunicatorObj &getCommunicator() const override { IT_TODO_HALT(); }
+    CommunicatorObj &getCommunicator() const override { return *comm; }
 
   private:
     void runWithoutSync(const Graph &graph, bool tune, bool profiling) const;

pyinfinitensor/src/pyinfinitensor/onnx.py

@@ -184,7 +184,7 @@ class OnnxStub:
                     node,
                     {
                         "dilations": [1, 1],
-                        "pads": [0, 0],
+                        "pads": [0, 0, 0, 0],
                         "strides": [1, 1],
                         "output_padding": [0, 0],
                     },
@@ -193,8 +193,52 @@ class OnnxStub:
                     attributes[name]
                     for name in ["dilations", "pads", "strides", "output_padding"]
                 )
+                if p[0] != p[2] or p[1] != p[3]:
+                    adapt = "{}-adapt".format(node.output[0])
+                    tensors[adapt] = self.handler.pad(
+                        tensors[node.input[0]], None, p, [-2, -1]
+                    )
+                    p = [0, 0, 0, 0]
+                else:
+                    adapt = node.input[0]
+                
+                if len(node.input) > 2:
+                    bias = "{}-bias".format(node.output[0])
+                    reshape = "{}-reshape".format(node.output[0])
+                    tensors[bias] = self.handler.convTransposed2d(
+                        tensors[adapt],
+                        tensors[node.input[1]],
+                        None,
+                        p[0],
+                        p[1],
+                        s[0],
+                        s[1],
+                        d[0],
+                        d[1],
+                        op[0],
+                        op[1],
+                    )
+                    tensors[reshape] = self.handler.reshape(
+                        tensors[node.input[2]],
+                        None,
+                        [
+                            1,
+                            reduce(
+                                lambda acc, x: acc * x,
+                                tensors[node.input[2]].shape(),
+                            ),
+                            1,
+                            1,
+                        ],
+                    )
+                    tensors[node.output[0]] = self.handler.add(
+                        tensors[bias],
+                        tensors[reshape],
+                        tensors.get(node.output[0]),
+                    )
+                else:
                     tensors[node.output[0]] = self.handler.convTransposed2d(
-                    tensors[node.input[0]],
+                        tensors[adapt],
                         tensors[node.input[1]],
                         tensors.get(node.output[0]),
                         p[0],

src/ascend/ascend_runtime.cc

@@ -56,4 +56,15 @@ void ASCENDRuntimeObj::sync() const { ; }
 
 string ASCENDRuntimeObj::toString() const { return "ASCEND Runtime"; }
 
+void ASCENDRuntimeObj::initComm(const string &name, int worldSize, int rank) {
+    IT_ASSERT(worldSize > 0);
+    IT_ASSERT(rank >= 0);
+    IT_ASSERT(rank < worldSize);
+    IT_ASSERT(!comm) << "communicator is already initialized.";
+#ifdef INFINI_USE_HCCL
+    comm = std::make_unique<HcclCommunicatorObj>(name, worldSize, rank);
+#else
+    IT_TODO_HALT_MSG("Not compiled with CNCL.");
+#endif
+}
 } // namespace infini

src/kernels/ascend/conv.cc

@@ -13,9 +13,9 @@ class ConvAclnn : public ASCENDKernelWithoutConfig {
         auto context = dynamic_cast<const ASCENDRuntimeObj *>(_context);
 
         const auto [ph, pw, sh, sw, dh, dw] = op->getPadStrideDilation();
-        // const auto [n, c, h, w, f, r, s] = op->getNCHWFRS();
+        const auto [n, c, h, w, f, r, s] = op->getNCHWFRS();
-        // const int cpg = op->getChannelPerGroup();
+        const int cpg = op->getChannelPerGroup();
-        // const int g = c / cpg;
+        const int g = c / cpg;
 
         std::vector<int64_t> pads = {ph, pw};
         // std::vector<int64_t> ksize = {r, s};
@@ -67,8 +67,8 @@ class ConvAclnn : public ASCENDKernelWithoutConfig {
 
         auto ret = aclnnConvolutionGetWorkspaceSize(
             inputTensor, weightTensor, nullptr, convstride, convpads,
-            convdilation, false, convOutputpadding, 1, outputTensor, 1,
+            convdilation, false, convOutputpadding, int64_t(g), outputTensor,
-            &workspaceSize, &executor);
+            int8_t(1), &workspaceSize, &executor);
         void *workspaceAddr = nullptr;
         if (workspaceSize > 0) {
             workspaceAddr = context->getWorkspace(workspaceSize);

src/kernels/ascend/pooling.cc

@@ -1,6 +1,6 @@
 #include "operators/pooling.h"
-#include "aclnnop/level2/aclnn_adaptive_max_pool2d.h"
 #include "aclnnop/level2/aclnn_avgpool2d.h"
+#include "aclnnop/level2/aclnn_max_pool.h"
 #include "ascend/ascend_kernel_without_config.h"
 #include "ascend/ascend_runtime.h"
 
@@ -75,40 +75,6 @@ class AvgPooling : public ASCENDKernelWithoutConfig {
 };
 
 class MaxPooling : public ASCENDKernelWithoutConfig {
-    // Only adaptiveMaxPool2d was found in the ACLNN doc.
-    int64_t GetShapeSize(const std::vector<int64_t> &shape) {
-        int64_t shapeSize = 1;
-        for (auto i : shape) {
-            shapeSize *= i;
-        }
-        return shapeSize;
-    }
-    template <typename T>
-    int CreateAclTensor(const std::vector<T> &hostData,
-                        const std::vector<int64_t> &shape, void **deviceAddr,
-                        aclDataType dataType, aclTensor **tensor) {
-        auto size = GetShapeSize(shape) * sizeof(T);
-        // 调用aclrtMalloc申请device侧内存
-        auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
-        assert(ret == ACL_SUCCESS);
-        // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
-        ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size,
-                          ACL_MEMCPY_HOST_TO_DEVICE);
-        assert(ret == ACL_SUCCESS);
-
-        // 计算连续tensor的strides
-        std::vector<int64_t> strides(shape.size(), 1);
-        for (int64_t i = shape.size() - 2; i >= 0; i--) {
-            strides[i] = shape[i + 1] * strides[i + 1];
-        }
-
-        // 调用aclCreateTensor接口创建aclTensor
-        *tensor = aclCreateTensor(shape.data(), shape.size(), dataType,
-                                  strides.data(), 0, aclFormat::ACL_FORMAT_NCHW,
-                                  shape.data(), shape.size(), *deviceAddr);
-        return 0;
-    }
-
     void compute(const Operator &_op,
                  const RuntimeObj *_context) const override {
         auto op = as<PoolingObj>(_op);
@@ -117,6 +83,15 @@ class MaxPooling : public ASCENDKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
 
+        auto [n, c, h, w, kh, kw] = op->getNCHWRS();
+        auto [ph, pw, sh, sw, dh, dw] = op->getPadStrideDilation();
+        int64_t ceilMode = int64_t(op->getCeilMode());
+
+        std::vector<int64_t> ksize = {kh, kw};
+        std::vector<int64_t> stride = {sh, sw};
+        std::vector<int64_t> pad = {ph, pw};
+        std::vector<int64_t> dilation = {dh, dw};
+
         auto selfD = op->getInputs(0)->getDims();
         auto selfS = op->getInputs(0)->getStride();
         auto outD = op->getOutput()->getDims();
@@ -127,20 +102,12 @@ class MaxPooling : public ASCENDKernelWithoutConfig {
         std::vector<int64_t> outputDim = castTo64(outD);
         std::vector<int64_t> outputStride = castTo64(outS);
 
-        std::vector<int64_t> outputHW(2, 1);
+        aclIntArray *kernelSize = aclCreateIntArray(ksize.data(), ksize.size());
-        outputHW[0] = outputDim[outputDim.size() - 2];
+        aclIntArray *strides = aclCreateIntArray(stride.data(), stride.size());
-        outputHW[1] = outputDim[outputDim.size() - 1];
+        aclIntArray *paddings = aclCreateIntArray(pad.data(), pad.size());
+        aclIntArray *dilations =
+            aclCreateIntArray(dilation.data(), dilation.size());
 
-        int64_t indicesOutSize = 1;
-        for (auto i : outputDim) {
-            indicesOutSize *= i;
-        }
-        void *indicesOutDeviceAddr = nullptr;
-        aclrtMalloc(&indicesOutDeviceAddr, indicesOutSize,
-                    ACL_MEM_MALLOC_HUGE_FIRST);
-
-        aclIntArray *outputsize =
-            aclCreateIntArray(outputHW.data(), outputHW.size());
         auto selfTensor = aclCreateTensor(
             selfDim.data(), selfDim.size(), ACL_FLOAT, selfStride.data(), 0,
             aclFormat::ACL_FORMAT_NCHW, selfDim.data(), selfDim.size(), aData);
@@ -148,16 +115,12 @@ class MaxPooling : public ASCENDKernelWithoutConfig {
             aclCreateTensor(outputDim.data(), outputDim.size(), ACL_FLOAT,
                             outputStride.data(), 0, aclFormat::ACL_FORMAT_NCHW,
                             outputDim.data(), outputDim.size(), cData);
-        auto indicesOutTensor = aclCreateTensor(
-            outputDim.data(), outputDim.size(), ACL_INT64, outputStride.data(),
-            0, aclFormat::ACL_FORMAT_NCHW, outputDim.data(), outputDim.size(),
-            indicesOutDeviceAddr);
 
         uint64_t workspaceSize = 0;
         aclOpExecutor *executor;
-        auto ret = aclnnAdaptiveMaxPool2dGetWorkspaceSize(
+        auto ret = aclnnMaxPoolGetWorkspaceSize(
-            selfTensor, outputsize, outputTensor, indicesOutTensor,
+            selfTensor, kernelSize, strides, 0, paddings, dilations, ceilMode,
-            &workspaceSize, &executor);
+            outputTensor, &workspaceSize, &executor);
         assert(ret == ACL_SUCCESS);
 
         void *workspaceAddr = nullptr;
@@ -165,15 +128,13 @@ class MaxPooling : public ASCENDKernelWithoutConfig {
             workspaceAddr = context->getWorkspace(workspaceSize);
         }
 
-        ret = aclnnAdaptiveMaxPool2d(workspaceAddr, workspaceSize, executor,
+        ret = aclnnMaxPool(workspaceAddr, workspaceSize, executor,
                            context->ASCENDHandle());
         assert(ret == ACL_SUCCESS);
 
         ret = aclrtSynchronizeStream(context->ASCENDHandle());
         assert(ret == ACL_SUCCESS);
 
-        aclDestroyTensor(indicesOutTensor);
-
         return;
     }
 };

test/kernels/ascend/test_ascend_conv.cc

@@ -45,14 +45,16 @@ void testConv(const std::function<void(void *, size_t, DataType)> &generatorA,
     cpuRuntime->run(cpuGraph);
     auto outputCpu = cpuOp->getOutput();
     // Check
-    EXPECT_TRUE(outputCpu->equalData(outputNpu2Cpu));
+    // outputCpu->printData();
+    // outputNpu2Cpu->printData();
+    EXPECT_TRUE(outputCpu->equalData(outputNpu2Cpu, 1e-3));
 }
 
 TEST(ascend_Conv, run) {
-    aclInit(nullptr);
+    // aclInit(nullptr);
     testConv<ConvObj>(IncrementalGenerator(), IncrementalGenerator(),
-                      Shape{1, 3, 32, 32}, Shape{2, 3, 3, 3});
+                      Shape{1, 3, 128, 128}, Shape{2, 3, 3, 3});
-    aclFinalize();
+    // aclFinalize();
 }
 
 } // namespace infini

test/kernels/ascend/test_ascend_element_wise.cc

@@ -53,9 +53,10 @@ TEST(ascend_ElementWise, run) {
     // aclInit(nullptr);
     // testElementWise<PowObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
     // testElementWise<AddObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
-    // testElementWise<SubObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
+    testElementWise<SubObj>(IncrementalGenerator(), Shape{1, 1, 48, 48},
-    testElementWise<DivObj>(IncrementalGenerator(), Shape{1},
+                            Shape{1, 1, 1, 1});
-                            Shape{1, 2, 2, 3});
+    // testElementWise<DivObj>(IncrementalGenerator(), Shape{1}, Shape{1, 2, 2,
+    // 3});
     // testElementWise<MulObj>(IncrementalGenerator(), Shape{1, 2, 2, 3});
     // aclFinalize();
 }

test/kernels/ascend/test_ascend_matmul.cc

@@ -50,10 +50,10 @@ void testMatmul(const std::function<void(void *, size_t, DataType)> &generatorA,
 }
 
 TEST(ascend_Matmul, run) {
-    aclInit(nullptr);
+    // aclInit(nullptr);
     testMatmul<MatmulObj>(IncrementalGenerator(), IncrementalGenerator(), false,
                           false, Shape{1, 2, 3}, Shape{1, 3, 4});
-    aclFinalize();
+    // aclFinalize();
 }
 
 } // namespace infini

test/kernels/ascend/test_ascend_pooling.cc

@@ -26,6 +26,7 @@ void testPooling(const std::function<void(void *, size_t, DataType)> &generator,
     auto inputNpu = npuGraph->cloneTensor(inputCpu);
     auto npuOp =
         npuGraph->addOp<T>(inputNpu, nullptr, 3, 3, 1, 1, 1, 1, 2, 2, 0);
+    // npuGraph->addOp<T>(inputNpu, nullptr, 2, 2, 1, 1, 0, 0, 1, 1, 0);
     npuGraph->dataMalloc();
     inputNpu->setData(generator);
     npuRuntime->run(npuGraph);
@@ -38,10 +39,10 @@ void testPooling(const std::function<void(void *, size_t, DataType)> &generator,
 }
 
 TEST(cnnl_Pooling, run) {
-    aclInit(nullptr);
+    // aclInit(nullptr);
-    testPooling<MaxPoolObj>(IncrementalGenerator(), Shape{1, 2, 5, 5});
+    // testPooling<MaxPoolObj>(IncrementalGenerator(), Shape{1, 2, 5, 5});
     testPooling<AvgPoolObj>(IncrementalGenerator(), Shape{1, 2, 5, 5});
-    aclFinalize();
+    //  aclFinalize();
 }
 
 } // namespace infini