!22008 [MS][LITE][Develop] optimize script

Merge pull request !22008 from sunsuodong/optimize_build

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/infer_register.c

@@ -15,7 +15,7 @@
  */
 #include "nnacl/infer/infer_register.h"
 
-#ifdef SUPPORT_MSVC
+#ifdef _MSC_VER
 #include "nnacl/infer/adam_infer.h"
 #include "nnacl/infer/add_sub_grad_infer.h"
 #include "nnacl/infer/addn_infer.h"
@@ -77,7 +77,6 @@
 #include "nnacl/infer/matmul_infer.h"
 #include "nnacl/infer/max_min_grad_infer.h"
 #include "nnacl/infer/mean_infer.h"
-#include "nnacl/infer/control/merge_infer.h"
 #include "nnacl/infer/mfcc_infer.h"
 #include "nnacl/infer/non_max_suppression_infer.h"
 #include "nnacl/infer/one_hot_infer.h"
@@ -117,7 +116,6 @@
 #include "nnacl/infer/stack_infer.h"
 #include "nnacl/infer/strided_slice_grad_infer.h"
 #include "nnacl/infer/strided_slice_infer.h"
-#include "nnacl/infer/control/switch_infer.h"
 #include "nnacl/infer/control/tensorlist_fromtensor_infer.h"
 #include "nnacl/infer/control/tensorlist_getitem_infer.h"
 #include "nnacl/infer/control/tensorlist_reserve_infer.h"
@@ -133,6 +131,15 @@
 #include "nnacl/infer/unstack_infer.h"
 #include "nnacl/infer/where_infer.h"
 #include "nnacl/infer/while_infer.h"
+#include "nnacl/infer/split_with_over_lap_infer.h"
+#include "nnacl/infer/ragged_range_infer.h"
+#include "nnacl/infer/glu_infer.h"
+#include "nnacl/infer/control/tensor_array_read_infer.h"
+#include "nnacl/infer/control/tensor_array_infer.h"
+#include "nnacl/infer/control/tensor_array_write_infer.h"
+#include "nnacl/infer/affine_infer.h"
+#include "nnacl/infer/attention_infer.h"
+#include "nnacl/infer/scatter_nd_update_infer.h"
 
 InferShape g_infer_func[PrimType_MAX * sizeof(InferShape)] = {0};
 void RegAllInferFunc1() {
@@ -230,7 +237,7 @@ void RegAllInferFunc1() {
   g_infer_func[PrimType_MaximumGrad] = MaxMinGradInferShape;
   g_infer_func[PrimType_MaxPoolFusion] = PoolingInferShape;
   g_infer_func[PrimType_MaxPoolGrad] = PoolingGradInferShape;
-  g_infer_func[PrimType_Merge] = MergeInferShape;
+  g_infer_func[PrimType_Merge] = NULL;
   g_infer_func[PrimType_Mfcc] = MfccInferShape;
   g_infer_func[PrimType_Minimum] = ArithmeticInferShape;
   g_infer_func[PrimType_MinimumGrad] = MaxMinGradInferShape;
@@ -293,7 +300,7 @@ void RegAllInferFunc2() {
   g_infer_func[PrimType_StridedSlice] = StridedSliceInferShape;
   g_infer_func[PrimType_SubFusion] = ArithmeticInferShape;
   g_infer_func[PrimType_SubGrad] = AddSubGradInferShape;
-  g_infer_func[PrimType_Switch] = SwitchInferShape;
+  g_infer_func[PrimType_Switch] = NULL;
   g_infer_func[PrimType_TensorListFromTensor] = TensorListFromTensorInferShape;
   g_infer_func[PrimType_TensorListGetItem] = TensorListGetItemInferShape;
   g_infer_func[PrimType_TensorListReserve] = TensorListReserveInferShape;
@@ -334,16 +341,32 @@ void RegAllInferFunc2() {
   g_infer_func[PrimType_CumSum] = CumsumInferShape;
 }
 
-typedef void RegFunc();
-#pragma data_seg(".CRT$XIU")
-static RegFunc *before[] = {RegAllInferFunc1, RegAllInferFunc2};
-#pragma data_seg()
+void RegAllInferFunc3() {
+  g_infer_func[PrimType_SplitWithOverlap] = SplitWithOverlapInferShape;
+  g_infer_func[PrimType_GenOP] = NULL;
+  g_infer_func[PrimType_RaggedRange] = RaggedRangeInferShape;
+  g_infer_func[PrimType_GLU] = GluInferShape;
+  g_infer_func[PrimType_TensorArray] = TensorArrayInferShape;
+  g_infer_func[PrimType_TensorArrayRead] = TensorArrayReadInferShape;
+  g_infer_func[PrimType_TensorArrayWrite] = TensorArrayWriteInferShape;
+  g_infer_func[PrimType_Affine] = AffineInferShape;
+  g_infer_func[PrimType_Attention] = AttentionInferShape;
+  g_infer_func[PrimType_LSTMGrad] = NULL;
+  g_infer_func[PrimType_ScatterNdUpdate] = ScatterNdUpdateInferShape;
+}
 
 #else
 __attribute__((init_priority(101))) InferShape g_infer_func[PrimType_MAX * sizeof(InferShape)] = {0};
-#endif  // SUPPORT_MSVC
+#endif  // _MSC_VER
 
 InferShape GetInferFunc(int prim_type) {
+#ifdef _MSC_VER
+  if (g_infer_func[PrimType_Abs] == NULL) {
+    RegAllInferFunc1();
+    RegAllInferFunc2();
+    RegAllInferFunc3();
+  }
+#endif
   if (prim_type < PrimType_MAX) {
     return g_infer_func[prim_type];
   }

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/infer_register.h

@@ -234,7 +234,7 @@ enum PrimType {
 
 void RegInfer(int prim_type, InferShape func);
 
-#ifdef SUPPORT_MSVC
+#ifdef _MSC_VER
 #define REG_INFER(op, type, func)
 #else
 #define REG_INFER(op, type, func) \

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/intrinsics/ms_simd_instructions.h

@@ -147,10 +147,10 @@ static inline float32x4_t vrecp(float32x4_t v) {
 #if defined(ENABLE_ARM) || defined(ENABLE_SSE)
 static inline MS_FLOAT32X4 MS_SQRTFX4_F32(MS_FLOAT32X4 src) {
   MS_FLOAT32X4 dst;
-  dst[0] = sqrtf(src[0]);
-  dst[1] = sqrtf(src[1]);
-  dst[2] = sqrtf(src[2]);
-  dst[3] = sqrtf(src[3]);
+  MS_F32X4_GETI(dst, 0) = sqrtf(MS_F32X4_GETI(src, 0));
+  MS_F32X4_GETI(dst, 1) = sqrtf(MS_F32X4_GETI(src, 1));
+  MS_F32X4_GETI(dst, 2) = sqrtf(MS_F32X4_GETI(src, 2));
+  MS_F32X4_GETI(dst, 3) = sqrtf(MS_F32X4_GETI(src, 3));
   return dst;
 }
 

mindspore/lite/build_lite.sh

@@ -221,6 +221,14 @@ build_lite() {
         else
           mv ${BASEPATH}/output/tmp/*.tar.gz* ${BASEPATH}/output/
         fi
+
+        if [[ "${local_lite_platform}" == "x86_64" ]]; then
+          mkdir -pv ${BASEPATH}/mindspore/lite/test/do_test || true
+          cp ${BASEPATH}/output/tmp/mindspore-lite*/tools/converter/lib/*.so* ${BASEPATH}/mindspore/lite/test/do_test || true
+          cp ${BASEPATH}/output/tmp/mindspore-lite*/runtime/lib/*.so* ${BASEPATH}/mindspore/lite/test/do_test || true
+          cp ${BASEPATH}/output/tmp/mindspore-lite*/runtime/third_party/libjpeg-turbo/lib/*.so* ${BASEPATH}/mindspore/lite/test/do_test || true
+        fi
+
         [ -n "${BASEPATH}" ] && rm -rf ${BASEPATH}/output/tmp/
         if [[ "${MSLITE_ENABLE_NNIE}" == "on" ]]; then
           compile_nnie_script=${BASEPATH}/mindspore/lite/tools/providers/NNIE/Hi3516D/compile_nnie.sh

mindspore/lite/test/runtest.sh

@@ -13,9 +13,6 @@ mkdir -pv ${CUR_DIR}/do_test
 cd ${CUR_DIR}/do_test
 cp ${BUILD_DIR}/test/lite-test ./
 cp ${BUILD_DIR}/googletest/googlemock/gtest/libgtest.so ./
-tar -xzf ../../../../output/mindspore-lite-*.tar.gz --strip-components=3 --wildcards *runtime/lib/*.so* || true
-tar -xzf ../../../../output/mindspore-lite-*.tar.gz --strip-components=4 --wildcards *converter/lib/*.so* || true
-tar -xzf ../../../../output/mindspore-lite-*.tar.gz --strip-components=5 --wildcards *libjpeg-turbo/lib/*.so* || true
 ls -l *.so*
 export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:./
 

mindspore/lite/tools/converter/quantizer/fse_decoder.cc

@@ -52,8 +52,18 @@ int FSEDecoder::FSECreateStatesForDecoding(const uint16_t *symbol_frequency, int
     uint16_t sym = symbol_table[i];
     uint16_t x = frequency[sym];
     frequency[sym] += 1;
-
+#ifdef _MSC_VER
+    int num = 0;
+    uint32_t tmp = x;
+    tmp |= 1;
+    while (!(tmp & 0x80000000)) {
+      num += 1;
+      tmp <<= 1;
+    }
+    bit_count[i] = table_log - (num ^ 31);
+#else
     bit_count[i] = table_log - (__builtin_clz(x) ^ 31);
+#endif
     new_state[i] = (x << bit_count[i]) - table_size;
   }
   return RET_OK;

mindspore/lite/tools/providers/NNIE/Hi3516D/compile_nnie.sh

@@ -14,8 +14,6 @@ function Run_Build_x86() {
   rm -rf ${nnie_code_path}/mindspore/mindspore/lite/tools/converter/nnie/third_party/ms_lite/
   mkdir -p ${nnie_code_path}/mindspore/mindspore/lite/tools/converter/nnie/third_party/ms_lite/ || exit 1
   cp -r ./tools/ ${nnie_code_path}/mindspore/mindspore/lite/tools/converter/nnie/third_party/ms_lite/ || exit 1
-  mkdir -pv ${open_source_ms_path}/mindspore/lite/test/do_test
-  cp ./tools/converter/lib/*.so* ${open_source_ms_path}/mindspore/lite/test/do_test
 
   # compile nnie converter so
   export MSLITE_ENABLE_NNIE=on