Merge pull request #2542 from alibaba/feature/bugfix

fix bug for vulkan tensorconvert error for some case of nc4hw4

CMakeLists.txt

@@ -59,11 +59,7 @@ option(MNN_INTERNAL "Build with MNN internal features, such as model authenticat
 option(MNN_JNI "Build MNN Jni for java to use" OFF)
 
 IF (NOT DEFINED MNN_USE_SPARSE_COMPUTE)
-  IF (TARGET_OS_IPHONE OR TARGET_OS_SIMULATOR)
-    set(MNN_USE_SPARSE_COMPUTE OFF)
-  ELSE()
-    set(MNN_USE_SPARSE_COMPUTE ON)
-  ENDIF()
+   set(MNN_USE_SPARSE_COMPUTE ON)
 ENDIF()
 
 IF(NOT MNN_BUILD_SHARED_LIBS AND MNN_SEP_BUILD)

source/backend/cpu/compute/ConvolutionFloatFactory.cpp

@@ -86,9 +86,9 @@ Execution* ConvolutionFloatFactory::create(const std::vector<Tensor*>& inputs, c
     std::shared_ptr<ConvolutionCommon::Int8Common> quanCommon;
     std::unique_ptr<Tensor> externalWeightTensor, externalBiasTensor;
     bool supportSparse = false;
-#ifdef MNN_USE_SPARSE_COMPUTE
     auto core = static_cast<CPUBackend*>(backend)->functions();
     int bytes = core->bytes;
+#ifdef MNN_USE_SPARSE_COMPUTE
 #ifdef MNN_USE_SSE
     const bool onlySSENotAVX = core->pack == 4; // no backend of only sse without avx2 or avx512
 #else

source/backend/vulkan/image/backend/VulkanTensor.cpp

@@ -54,7 +54,6 @@ std::array<int, 4> VulkanTensor::tensorShapeFormat(const Tensor *input) {
         iW = 1;
         iC = input->buffer().dim[0].extent;
     }
-
 #ifdef LOG_VERBOSE
     MNN_PRINT("tensorShapeFormat : [%d, %d, %d, %d] \n", iN, iH, iW, iC);
 #endif

source/backend/vulkan/image/execution/VulkanImageConverter.cpp

@@ -126,10 +126,6 @@ void VulkanImageConverter::encodeTensorToBuffer(const Tensor* srcTensor, VkBuffe
                                                 const VulkanCommandPool::Buffer* cmdBuffer) {
     auto sourceFormat = TensorUtils::getDescribe(srcTensor)->dimensionFormat;
     auto destFormat   = destBufferFormat;
-    if (sourceFormat == MNN_DATA_FORMAT_NC4HW4 && 1 >= srcTensor->width() && 1 >= srcTensor->height() &&
-        srcTensor->channel() % 4 == 0) {
-        destFormat = MNN_DATA_FORMAT_NC4HW4;
-    }
 
     auto vkTensor = (VulkanTensor*)(srcTensor->deviceId());
     auto tensor = srcTensor;

source/backend/vulkan/image/execution/VulkanRaster.cpp

@@ -156,8 +156,6 @@ ErrorCode VulkanRaster::onEncode(const std::vector<Tensor *> &___inputs, const s
         for (int i=0; i<origin->dimensions(); ++i) {
             bufferSize *= origin->length(i);
         }
-        ivec4 dims;
-        writeNCHW(dims, origin);
         ConvertInfo info;
         info.buffer.reset(new VulkanBuffer(vkBn->getDynamicMemoryPool(),
                                            false, bufferSize, nullptr, VK_BUFFER_USAGE_STORAGE_BUFFER_BIT));
@@ -212,6 +210,7 @@ ErrorCode VulkanRaster::onEncode(const std::vector<Tensor *> &___inputs, const s
     for (auto& iter : mInputBuffers) {
         auto& info = iter.second;
         info.convert->encodeTensorToBuffer(iter.first, info.buffer->buffer(), info.buffer->size(), 0, VulkanImageConverter::getTensorLinearFormat(iter.first), cmdBuffer);
+        cmdBuffer->barrierSource(info.buffer->buffer(), 0, info.buffer->size());
     }
     //Blit
     if (needZero) {
@@ -223,7 +222,6 @@ ErrorCode VulkanRaster::onEncode(const std::vector<Tensor *> &___inputs, const s
         info.describe->writeBuffer(info.srcBuffer, 1, info.srcBufferSize);
         info.describe->writeBuffer(info.uniform->buffer(), 2, info.uniform->size());
         info.pipeline->bind(cmdBuffer->get(), info.describe->get());
-        cmdBuffer->barrierSource(info.srcBuffer, 0, info.srcBufferSize);
         vkCmdDispatch(cmdBuffer->get(), info.workGroup[0], info.workGroup[1], info.workGroup[2]);
     }