Fix scan issues (#2729)

intel · Apr 7, 2024 · 7bc3869 · 7bc3869
1 parent e1095f7
commit 7bc3869
Show file tree

Hide file tree

Showing 18 changed files with 212 additions and 139 deletions.
diff --git a/csrc/cpu/aten/Linear.cpp b/csrc/cpu/aten/Linear.cpp
@@ -392,10 +392,10 @@ at::Tensor woq_linear_pack_weight(
       int64_t weight_int4_size_bytes = weight_int4.numel();
       int64_t pad_size_bytes = weight_int4_size_bytes - weight_size_bytes;
       std::memcpy(weight_int4.data_ptr(), weight.data_ptr(), weight_size_bytes);
-      std::memset(
+      std::fill_n(
           (uint8_t*)weight_int4.data_ptr() + weight_size_bytes,
-          0,
-          pad_size_bytes);
+          pad_size_bytes,
+          0);
       return woq_tpp_gemm_packB_stub(
           kCPU, weight_int4, weight_dtype, block_n, block_k, lowp_mode);
     }

diff --git a/csrc/cpu/aten/MergedEmbeddingBag.h b/csrc/cpu/aten/MergedEmbeddingBag.h
@@ -24,7 +24,7 @@ class EMBROW {
     length = len;
     arr.resize(length);
     data = &arr[0];
-    memset(data, 0, len * sizeof(T));
+    std::fill_n(data, len, T(0));
   }
 };
 
@@ -36,7 +36,7 @@ class EMBROWFixLen {
 
   EMBROWFixLen(int32_t len) {
     data = &arr[0];
-    memset(data, 0, emb_dim * sizeof(T));
+    std::fill_n(data, emb_dim, T(0));
   }
 };
 

diff --git a/csrc/cpu/aten/kernels/WoqLinearKrnl.cpp b/csrc/cpu/aten/kernels/WoqLinearKrnl.cpp
@@ -1507,7 +1507,7 @@ void dot_update(
 template <typename T>
 void zero_fill(T* C, int M, int N, int stride) {
   for (int m = 0; m < M; m++) {
-    memset(C + m * stride, 0, sizeof(T) * N);
+    std::fill_n(C + m * stride, N, T(0));
   }
 }
 

diff --git a/csrc/cpu/aten/kernels/WoqTppKrnl.cpp b/csrc/cpu/aten/kernels/WoqTppKrnl.cpp
@@ -2016,10 +2016,8 @@ void qlinear_woq_affine_impl(
                     64, num_threads * M * N * sizeof(TGemmOut));
                 y_private_valid = (bool*)std::aligned_alloc(
                     64, num_threads * (M / BLOCK_M) * Nc * sizeof(bool));
-                memset(
-                    y_private_valid,
-                    0,
-                    sizeof(bool) * num_threads * (M / BLOCK_M) * Nc);
+                std::fill_n(
+                    y_private_valid, num_threads * (M / BLOCK_M) * Nc, false);
               }
               auto y_private_ptr = GetVLAPtr<TGemmOut>(y_private, {M, Nc, Nb});
               auto y_private_valid_ptr =

diff --git a/csrc/cpu/aten/utils/woq.h b/csrc/cpu/aten/utils/woq.h
@@ -74,12 +74,16 @@ class DotMicroKernel {
         (trans_a ? LIBXSMM_GEMM_FLAG_TRANS_A : LIBXSMM_GEMM_FLAG_NONE) |
         (trans_b ? LIBXSMM_GEMM_FLAG_TRANS_B : LIBXSMM_GEMM_FLAG_NONE);
     libxsmm_gemm_batch_reduce_config brconfig;
-    memset(&brconfig, 0, sizeof(libxsmm_gemm_batch_reduce_config));
+    std::fill_n(
+        reinterpret_cast<char*>(&brconfig),
+        sizeof(libxsmm_gemm_batch_reduce_config),
+        0);
     brconfig.br_type = LIBXSMM_GEMM_BATCH_REDUCE_NONE;
 
     kernel_func_ = libxsmm_dispatch_brgemm_v2(
         brshape, brflags, /*prefetch_flags=*/0, brconfig);
-    memset(&gemm_param_, 0, sizeof(libxsmm_gemm_param));
+    std::fill_n(
+        reinterpret_cast<char*>(&gemm_param_), sizeof(libxsmm_gemm_param), 0);
   }
 
   void operator()(void* A, void* B, void* C) {

diff --git a/csrc/cpu/comm/shm_reduction.h b/csrc/cpu/comm/shm_reduction.h
@@ -149,8 +149,8 @@ class ShmReduction {
     shmCtx_.nblocks = MAX_SHM_BLOCK_COUNT;
     if (rank_ == 0) {
       torch_ipex::cpu::create_shm(&shmCtx_);
-      memset(shmCtx_.state, 0, shmCtx_.nstates * sizeof(int));
-      memset((void*)shmCtx_.blockState, 0, shmCtx_.nstates * shmCtx_.nblocks);
+      std::fill_n(shmCtx_.state, shmCtx_.nstates, 0);
+      std::fill_n(shmCtx_.blockState, shmCtx_.nstates * shmCtx_.nblocks, 0);
     }
 
     callback(shmCtx_.pid_fd, 2);

diff --git a/csrc/cpu/tpp/jit_compile.cpp b/csrc/cpu/tpp/jit_compile.cpp
@@ -18,7 +18,7 @@ void* jit_compile_and_load(
   int fd = mkstemp(libname);
   unlink(libname);
   char fdname[50];
-  sprintf(fdname, "/proc/self/fd/%d", fd);
+  snprintf(fdname, sizeof(fdname), "/proc/self/fd/%d", fd);
   auto cmd = std::string("g++ -shared -fPIC -x c++ ") + flags;
   cmd = cmd + " -o " + fdname + " " + filename;
   printf("JIT COMPILE: %s\n", cmd.c_str());
@@ -66,7 +66,7 @@ void* jit_from_str(
   int fd = mkstemp(filename);
   unlink(filename);
   char fdname[50];
-  sprintf(fdname, "/proc/self/fd/%d", fd);
+  snprintf(fdname, sizeof(fdname), "/proc/self/fd/%d", fd);
   write(fd, src.c_str(), src.length());
   return jit_from_file(fdname, flags, func_name);
 #else