[NNAdapter] part of qnn (#9106)

PaddlePaddle · Jun 9, 2022 · b3e73d4 · b3e73d4
1 parent ef8d53c
commit b3e73d4
Show file tree

Hide file tree

Showing 5 changed files with 71 additions and 56 deletions.
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
@@ -238,6 +238,9 @@ if (LITE_WITH_NNADAPTER)
     if (NNADAPTER_WITH_NVIDIA_TENSORRT)
       add_definitions("-DNNADAPTER_WITH_NVIDIA_TENSORRT")
     endif()
+    if (NNADAPTER_WITH_QUALCOMM_QNN)
+      add_definitions("-DNNADAPTER_WITH_QUALCOMM_QNN")
+    endif()
     if (NNADAPTER_WITH_KUNLUNXIN_XTCL)
       add_definitions("-DNNADAPTER_WITH_KUNLUNXIN_XTCL")
     endif()

diff --git a/lite/backends/nnadapter/nnadapter/src/driver/CMakeLists.txt b/lite/backends/nnadapter/nnadapter/src/driver/CMakeLists.txt
@@ -50,6 +50,10 @@ if(NNADAPTER_WITH_NVIDIA_TENSORRT)
   add_subdirectory(nvidia_tensorrt)
 endif()
 
+if(NNADAPTER_WITH_QUALCOMM_QNN)
+  add_subdirectory(qualcomm_qnn)
+endif()
+
 if(NNADAPTER_WITH_KUNLUNXIN_XTCL)
   add_subdirectory(kunlunxin_xtcl)
 endif()

diff --git a/lite/core/test/arena/framework.cc b/lite/core/test/arena/framework.cc
@@ -56,6 +56,8 @@ std::shared_ptr<lite::OpLite> TestCase::CreateSubgraphOp() {
 #elif defined(NNADAPTER_WITH_NVIDIA_TENSORRT)
   ctx_->As<NNAdapterContext>().SetNNAdapterDeviceNames(scope,
                                                        {"nvidia_tensorrt"});
+#elif defined(DNNADAPTER_WITH_QUALCOMM_QNN)
+  ctx_->As<NNAdapterContext>().SetNNAdapterDeviceNames(scope, {"qualcomm_qnn"});
 #elif defined(NNADAPTER_WITH_KUNLUNXIN_XTCL)
   ctx_->As<NNAdapterContext>().SetNNAdapterDeviceNames(scope,
                                                        {"kunlunxin_xtcl"});

diff --git a/lite/kernels/nnadapter/converter/all.h b/lite/kernels/nnadapter/converter/all.h
@@ -15,12 +15,10 @@
 #ifndef __NNADAPTER_CONVERTER_ALL_H__  // NOLINT
 #define __NNADAPTER_CONVERTER_ALL_H__
 
-REGISTER_CONVERTER(
-    batch_norm,
-    ConvertBatchNorm,
-    "huawei_ascend_npu,verisilicon_"
-    "timvx,cambricon_mlu,huawei_kirin_npu,intel_openvino,nvidia_tensorrt,"
-    "eeasytech_npu");
+REGISTER_CONVERTER(batch_norm,
+                   ConvertBatchNorm,
+                   "huawei_ascend_npu,verisilicon_timvx,cambricon_mlu,huawei_"
+                   "kirin_npu,intel_openvino,nvidia_tensorrt,eeasytech_npu");
 REGISTER_CONVERTER(cast,
                    ConvertCast,
                    "huawei_ascend_npu,cambricon_mlu,huawei_kirin_npu,nvidia_"
@@ -29,34 +27,31 @@ REGISTER_CONVERTER(clip,
                    ConvertClip,
                    "huawei_ascend_npu,cambricon_mlu,verisilicon_timvx,huawei_"
                    "kirin_npu,nvidia_tensorrt,intel_openvino");
-REGISTER_CONVERTER(
-    conv2d,
-    ConvertConv2D,
-    "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
-    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_openvino,"
-    "eeasytech_npu");
-REGISTER_CONVERTER(
-    depthwise_conv2d,
-    ConvertConv2D,
-    "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,verisilicon_timvx,"
-    "kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_openvino,eeasytech_"
-    "npu");
+REGISTER_CONVERTER(conv2d,
+                   ConvertConv2D,
+                   "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,"
+                   "huawei_ascend_npu,amlogic_npu,imagination_nna,cambricon_"
+                   "mlu,verisilicon_timvx,kunlunxin_xtcl,android_nnapi,nvidia_"
+                   "tensorrt,intel_openvino,qualcomm_qnn,eeasytech_npu");
+REGISTER_CONVERTER(depthwise_conv2d,
+                   ConvertConv2D,
+                   "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,"
+                   "huawei_ascend_npu,amlogic_npu,imagination_nna,verisilicon_"
+                   "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_"
+                   "openvino,qualcomm_qnn,eeasytech_npu");
 REGISTER_CONVERTER(deformable_conv,
                    ConvertDeformableConv,
                    "huawei_ascend_npu,cambricon_mlu,intel_openvino");
 REGISTER_CONVERTER(dropout,
                    ConvertDropout,
                    "huawei_ascend_npu,huawei_kirin_npu,verisilicon_timvx,"
                    "nvidia_tensorrt,cambricon_mlu,intel_openvino");
-REGISTER_CONVERTER(
-    pool2d,
-    ConvertPool,
-    "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
-    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_openvino,"
-    "eeasytech_npu");
+REGISTER_CONVERTER(pool2d,
+                   ConvertPool,
+                   "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,"
+                   "huawei_ascend_npu,amlogic_npu,imagination_nna,cambricon_"
+                   "mlu,verisilicon_timvx,kunlunxin_xtcl,android_nnapi,nvidia_"
+                   "tensorrt,intel_openvino,qualcomm_qnn,eeasytech_npu");
 REGISTER_CONVERTER(matmul,
                    ConvertMatmul,
                    "huawei_ascend_npu,huawei_kirin_npu,imagination_nna,"
@@ -65,13 +60,12 @@ REGISTER_CONVERTER(matmul_v2,
                    ConvertMatmulV2,
                    "huawei_ascend_npu,huawei_kirin_npu,imagination_nna,intel_"
                    "openvino,android_nnapi,nvidia_tensorrt,intel_openvino");
-REGISTER_CONVERTER(
-    softmax,
-    ConvertSoftmax,
-    "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
-    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_"
-    "openvino,google_xnnpack");
+REGISTER_CONVERTER(softmax,
+                   ConvertSoftmax,
+                   "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,"
+                   "huawei_ascend_npu,amlogic_npu,imagination_nna,cambricon_"
+                   "mlu,verisilicon_timvx,kunlunxin_xtcl,android_nnapi,nvidia_"
+                   "tensorrt,intel_openvino,qualcomm_qnn,google_xnnpack");
 REGISTER_CONVERTER(cumsum, ConvertCumsum, "huawei_ascend_npu,intel_openvino");
 REGISTER_CONVERTER(conv2d_transpose,
                    ConvertConv2dTranspose,
@@ -83,13 +77,13 @@ REGISTER_CONVERTER(reshape,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,imagination_nna,verisilicon_timvx,"
                    "kunlunxin_xtcl,cambricon_mlu,android_nnapi,nvidia_tensorrt,"
-                   "intel_openvino,google_xnnpack");
+                   "intel_openvino,qualcomm_qnn,google_xnnpack");
 REGISTER_CONVERTER(reshape2,
                    ConvertReshape,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,imagination_nna,verisilicon_timvx,"
                    "kunlunxin_xtcl,cambricon_mlu,android_nnapi,nvidia_tensorrt,"
-                   "intel_openvino,google_xnnpack");
+                   "intel_openvino,qualcomm_qnn,google_xnnpack");
 REGISTER_CONVERTER(
     unsqueeze,
     ConvertUnsqueeze,
@@ -107,7 +101,7 @@ REGISTER_CONVERTER(elementwise_add,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
                    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_"
-                   "openvino,google_xnnpack,eeasytech_npu");
+                   "openvino,qualcomm_qnn,google_xnnpack,eeasytech_npu");
 REGISTER_CONVERTER(elementwise_sub,
                    ConvertElementwise,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
@@ -180,13 +174,12 @@ REGISTER_CONVERTER(sigmoid,
                    "npu,amlogic_npu,cambricon_mlu,verisilicon_timvx,kunlunxin_"
                    "xtcl,android_nnapi,nvidia_tensorrt,intel_openvino,"
                    "eeasytech_npu");
-REGISTER_CONVERTER(
-    relu,
-    ConvertUnaryActivations,
-    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
-    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_openvino,"
-    "eeasytech_npu");
+REGISTER_CONVERTER(relu,
+                   ConvertUnaryActivations,
+                   "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
+                   "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
+                   "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_"
+                   "openvino,qualcomm_qnn,eeasytech_npu");
 REGISTER_CONVERTER(
     relu6,
     ConvertUnaryActivations,
@@ -295,12 +288,12 @@ REGISTER_CONVERTER(transpose,
                    ConvertTranspose,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,verisilicon_timvx,kunlunxin_xtcl,android_"
-                   "nnapi,nvidia_tensorrt,intel_openvino");
+                   "nnapi,nvidia_tensorrt,intel_openvino,qualcomm_qnn");
 REGISTER_CONVERTER(transpose2,
                    ConvertTranspose,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,verisilicon_timvx,kunlunxin_xtcl,android_"
-                   "nnapi,nvidia_tensorrt,intel_openvino");
+                   "nnapi,nvidia_tensorrt,intel_openvino,qualcomm_qnn");
 REGISTER_CONVERTER(
     shape,
     ConvertShape,
@@ -375,17 +368,18 @@ REGISTER_CONVERTER(flatten2,
                    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
                    "npu,amlogic_npu,verisilicon_timvx,kunlunxin_xtcl,cambricon_"
                    "mlu,android_nnapi,nvidia_tensorrt,intel_openvino");
-REGISTER_CONVERTER(flatten_contiguous_range,
-                   ConvertFlattenContiguousRange,
-                   "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-                   "npu,amlogic_npu,verisilicon_timvx,kunlunxin_xtcl,cambricon_"
-                   "mlu,android_nnapi,nvidia_tensorrt,intel_openvino");
 REGISTER_CONVERTER(
-    fc,
-    ConvertFC,
-    "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
-    "npu,amlogic_npu,imagination_nna,cambricon_mlu,verisilicon_"
-    "timvx,kunlunxin_xtcl,android_nnapi,nvidia_tensorrt,intel_openvino");
+    flatten_contiguous_range,
+    ConvertFlattenContiguousRange,
+    "rockchip_npu,mediatek_apu,huawei_kirin_npu,huawei_ascend_"
+    "npu,amlogic_npu,verisilicon_timvx,kunlunxin_xtcl,cambricon_"
+    "mlu,android_nnapi,nvidia_tensorrt,intel_openvino,qualcomm_qnn");
+REGISTER_CONVERTER(fc,
+                   ConvertFC,
+                   "builtin_device,rockchip_npu,mediatek_apu,huawei_kirin_npu,"
+                   "huawei_ascend_npu,amlogic_npu,imagination_nna,cambricon_"
+                   "mlu,verisilicon_timvx,kunlunxin_xtcl,android_nnapi,nvidia_"
+                   "tensorrt,intel_openvino,qualcomm_qnn");
 REGISTER_CONVERTER(norm,
                    ConvertNorm,
                    "huawei_ascend_npu,cambricon_mlu,huawei_kirin_npu");

diff --git a/lite/tools/build_linux.sh b/lite/tools/build_linux.sh
@@ -56,6 +56,8 @@ NNADAPTER_VERISILICON_TIMVX_VIV_SDK_URL="http://paddlelite-demo.bj.bcebos.com/de
 NNADAPTER_WITH_NVIDIA_TENSORRT=OFF
 NNADAPTER_NVIDIA_CUDA_ROOT="/usr/local/cuda"
 NNADAPTER_NVIDIA_TENSORRT_ROOT="/usr/local/tensorrt"
+NNADAPTER_WITH_QUALCOMM_QNN=OFF
+NNADAPTER_QUALCOMM_QNN_SDK_ROOT="/usr/local/qnn"
 NNADAPTER_WITH_KUNLUNXIN_XTCL=OFF
 NNADAPTER_KUNLUNXIN_XTCL_SDK_ROOT=""
 NNADAPTER_KUNLUNXIN_XTCL_SDK_URL=""
@@ -232,6 +234,8 @@ function init_cmake_mutable_options {
                         -DNNADAPTER_WITH_NVIDIA_TENSORRT=$NNADAPTER_WITH_NVIDIA_TENSORRT \
                         -DNNADAPTER_NVIDIA_CUDA_ROOT=$NNADAPTER_NVIDIA_CUDA_ROOT \
                         -DNNADAPTER_NVIDIA_TENSORRT_ROOT=$NNADAPTER_NVIDIA_TENSORRT_ROOT \
+                        -DNNADAPTER_WITH_QUALCOMM_QNN=$NNADAPTER_WITH_QUALCOMM_QNN \
+                        -DNNADAPTER_QUALCOMM_QNN_SDK_ROOT=$NNADAPTER_QUALCOMM_QNN_SDK_ROOT \
                         -DNNADAPTER_WITH_KUNLUNXIN_XTCL=$NNADAPTER_WITH_KUNLUNXIN_XTCL \
                         -DNNADAPTER_KUNLUNXIN_XTCL_SDK_ROOT=$NNADAPTER_KUNLUNXIN_XTCL_SDK_ROOT \
                         -DNNADAPTER_KUNLUNXIN_XTCL_SDK_URL=$NNADAPTER_KUNLUNXIN_XTCL_SDK_URL \
@@ -618,6 +622,14 @@ function main {
                 NNADAPTER_NVIDIA_TENSORRT_ROOT="${i#*=}"
                 shift
                 ;;
+            --nnadapter_with_qualcomm_qnn=*)
+                NNADAPTER_WITH_QUALCOMM_QNN="${i#*=}"
+                shift
+                ;;
+            --nnadapter_qualcomm_qnn_sdk_root=*)
+                NNADAPTER_QUALCOMM_QNN_SDK_ROOT="${i#*=}"
+                shift
+                ;;
             --nnadapter_with_kunlunxin_xtcl=*)
                 NNADAPTER_WITH_KUNLUNXIN_XTCL="${i#*=}"
                 shift