diff --git a/src/cpu/cpu_reducer.hpp b/src/cpu/cpu_reducer.hpp
index 6c364193e08..d370775b3e5 100644
--- a/src/cpu/cpu_reducer.hpp
+++ b/src/cpu/cpu_reducer.hpp
@@ -198,6 +198,10 @@ struct cpu_reducer_t {
     void reduce_nolock(int ithr, data_t *dst);
 };
 
+// Explicit instantiations in cpu_reducer.cpp.
+extern template struct cpu_reducer_t<data_type::f32>;
+extern template struct cpu_reducer_t<data_type::s32>;
+
 template <impl::data_type_t data_type>
 struct cpu_reducer_2d_t {
     typedef typename prec_traits<data_type>::type data_t;
@@ -254,6 +258,10 @@ struct cpu_reducer_2d_t {
     void reduce_nolock(int ithr, data_t *dst);
 };
 
+// Explicit instantiations in cpu_reducer.cpp.
+extern template struct cpu_reducer_2d_t<data_type::f32>;
+extern template struct cpu_reducer_2d_t<data_type::s32>;
+
 /** simple 1d accumulator: y[:] += x[:] */
 template <impl::data_type_t data_type>
 struct cpu_accumulator_1d_t {
@@ -266,6 +274,10 @@ struct cpu_accumulator_1d_t {
     reducer_2d_driver_t<data_type> *drv_;
 };
 
+// Explicit instantiations in cpu_reducer.cpp.
+extern template struct cpu_accumulator_1d_t<data_type::f32>;
+extern template struct cpu_accumulator_1d_t<data_type::s32>;
+
 }
 }
 }
diff --git a/src/cpu/gemm/gemm.hpp b/src/cpu/gemm/gemm.hpp
index 3f33a37132c..5eb2f0df3b6 100644
--- a/src/cpu/gemm/gemm.hpp
+++ b/src/cpu/gemm/gemm.hpp
@@ -36,6 +36,18 @@ void ref_gemm(const char *transa, const char *transb, const int *M,
         const int *N, const int *K, const data_t *alpha, const data_t *A,
         const int *lda, const data_t *B, const int *ldb, const data_t *beta,
         data_t *C, const int *ldc, const data_t *bias);
+
+// Explicit instantiations in ref_gemm.cpp.
+extern template
+void ref_gemm<float>(const char *transa_, const char *transb_,
+        const int *M_, const int *N_, const int *K_, const float *alpha_,
+        const float *A, const int *lda_, const float *B, const int *ldb_,
+        const float *beta_, float *C, const int *ldc_, const float *bias);
+extern template
+void ref_gemm<double>(const char *transa_, const char *transb_,
+        const int *M_, const int *N_, const int *K_, const double *alpha_,
+        const double *A, const int *lda_, const double *B, const int *ldb_,
+        const double *beta_, double *C, const int *ldc_, const double *bias);
 #ifdef USE_CBLAS
 #define GEMM_IMPL_STR "gemm:blas"
 #else
diff --git a/src/cpu/gemm/gemm_utils.hpp b/src/cpu/gemm/gemm_utils.hpp
index 0888787b9ce..ca6e59993b9 100644
--- a/src/cpu/gemm/gemm_utils.hpp
+++ b/src/cpu/gemm/gemm_utils.hpp
@@ -51,6 +51,14 @@ template <typename data_type>
 void sum_two_matrices(
         int m, int n, data_type *p_src, int ld_src, data_type *p_dst, int ld_dst);
 
+// Explicit instantiations are provided in gemm_utils.cpp.
+extern template
+void sum_two_matrices<float>(
+        int m, int n, float *p_src, int ld_src, float *p_dst, int ld_dst);
+extern template
+void sum_two_matrices<double>(
+        int m, int n, double *p_src, int ld_src, double *p_dst, int ld_dst);
+
 void calc_nthr_nocopy_avx512_common(int m,
         int n, int k, int nthrs, int *nthrs_m, int *nthrs_n, int *nthrs_k,
         int *BM, int *BN, int *BK);
diff --git a/src/cpu/jit_avx512_common_convolution_winograd.hpp b/src/cpu/jit_avx512_common_convolution_winograd.hpp
index bba30ea8944..1e98a88af10 100644
--- a/src/cpu/jit_avx512_common_convolution_winograd.hpp
+++ b/src/cpu/jit_avx512_common_convolution_winograd.hpp
@@ -203,6 +203,20 @@ struct _jit_avx512_common_convolution_winograd_t {
         const primitive_attr_t *attr_;
 };
 
+// Explicit instantiations in jit_avx512_common_convolution_winograd.cpp.
+extern template void
+_jit_avx512_common_convolution_winograd_t<true>::_execute_data_W_S_G_D(
+        float *, float *, float *, float *);
+extern template void
+_jit_avx512_common_convolution_winograd_t<false>::_execute_data_W_S_G_D(
+        float *, float *, float *, float *);
+extern template void
+_jit_avx512_common_convolution_winograd_t<true>::_execute_data_W_SGD(
+        float *, float *, float *, float *);
+extern template void
+_jit_avx512_common_convolution_winograd_t<false>::_execute_data_W_SGD(
+        float *, float *, float *, float *);
+
 template <bool with_relu>
 struct _jit_avx512_common_convolution_winograd_fwd_t
      : _jit_avx512_common_convolution_winograd_t<true>
diff --git a/src/cpu/jit_uni_dw_conv_kernel_f32.hpp b/src/cpu/jit_uni_dw_conv_kernel_f32.hpp
index 555bf025910..fd6a7369362 100644
--- a/src/cpu/jit_uni_dw_conv_kernel_f32.hpp
+++ b/src/cpu/jit_uni_dw_conv_kernel_f32.hpp
@@ -300,6 +300,12 @@ struct jit_uni_dw_conv_bwd_weights_kernel_f32 : public jit_generator {
 
     void generate();
 };
+
+// Explicit instantiations in jit_uni_dw_conv_kernel_f32.cpp.
+extern template struct jit_uni_dw_conv_bwd_weights_kernel_f32<avx512_common>;
+extern template struct jit_uni_dw_conv_bwd_weights_kernel_f32<avx2>;
+extern template struct jit_uni_dw_conv_bwd_weights_kernel_f32<sse42>;
+
 }
 }
 }
diff --git a/src/cpu/jit_uni_eltwise.hpp b/src/cpu/jit_uni_eltwise.hpp
index 3f8906b25f4..32493e5bb8f 100644
--- a/src/cpu/jit_uni_eltwise.hpp
+++ b/src/cpu/jit_uni_eltwise.hpp
@@ -29,6 +29,11 @@ namespace mkldnn {
 namespace impl {
 namespace cpu {
 
+// Explicit instantiations are in jit_uni_eltwise.cpp.
+extern template struct jit_uni_eltwise_injector_f32<avx512_common>;
+extern template struct jit_uni_eltwise_injector_f32<avx2>;
+extern template struct jit_uni_eltwise_injector_f32<sse42>;
+
 struct jit_uni_eltwise_kernel_f32;
 
 template <cpu_isa_t isa>
diff --git a/src/cpu/jit_uni_lrn_kernel_f32.hpp b/src/cpu/jit_uni_lrn_kernel_f32.hpp
index 68270197d66..611c939beb9 100644
--- a/src/cpu/jit_uni_lrn_kernel_f32.hpp
+++ b/src/cpu/jit_uni_lrn_kernel_f32.hpp
@@ -173,6 +173,60 @@ struct jit_uni_lrn_bwd_kernel_f32 : public jit_generator {
     void(*ker)(jit_args_bwd_t *);
 };
 
+// Explicit specializations and instantiations in jit_uni_lrn_kernel_f32.cpp.
+template <>
+jit_uni_lrn_fwd_kernel_f32<avx2>::jit_uni_lrn_fwd_kernel_f32(
+    const struct nchw8c_across &J, float A, float K, prop_kind_t pk,
+    void *code_ptr, size_t code_size);
+template <>
+jit_uni_lrn_fwd_kernel_f32<sse42>::jit_uni_lrn_fwd_kernel_f32(
+    const struct nchw8c_across &J, float A, float K, prop_kind_t pk,
+    void *code_ptr, size_t code_size);
+template <>
+jit_uni_lrn_fwd_kernel_f32<avx2>::jit_uni_lrn_fwd_kernel_f32(
+    const struct nhwc_across &J, float A, float K, prop_kind_t pk,
+    void *code_ptr, size_t code_size);
+template <>
+jit_uni_lrn_fwd_kernel_f32<sse42>::jit_uni_lrn_fwd_kernel_f32(
+    const struct nhwc_across &J, float A, float K, prop_kind_t pk,
+    void *code_ptr, size_t code_size);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<sse42>::nchw_body(
+    int tail, int HW, prop_kind_t pk, Xbyak::Ymm ymask, Xbyak::Ymm ya,
+    Xbyak::Ymm yb, Xbyak::Ymm yc, Xbyak::Ymm yd, Xbyak::Ymm ye,
+    Xbyak::Ymm ysum);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<avx2>::nchw_body(
+    int tail, int HW, prop_kind_t pk, Xbyak::Ymm ymask, Xbyak::Ymm ya,
+    Xbyak::Ymm yb, Xbyak::Ymm yc, Xbyak::Ymm yd, Xbyak::Ymm ye,
+    Xbyak::Ymm ysum);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<avx2>::nchw_tail_sse42(
+    int tail, Xbyak::Reg64 reg_dst, Xbyak::Xmm xtail_lo, Xbyak::Xmm xtail_hi);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<sse42>::nchw_tail_sse42(
+    int tail, Xbyak::Reg64 reg_dst, Xbyak::Xmm xtail_lo, Xbyak::Xmm xtail_hi);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<sse42>::nchw_body_sse42(
+    int tail, int HW, prop_kind_t pk, Xbyak::Xmm xmask_lo, Xbyak::Xmm xmask_hi,
+    Xbyak::Xmm xe_lo, Xbyak::Xmm xe_hi, Xbyak::Xmm xsum_lo, Xbyak::Xmm xsum_hi);
+template <>
+void jit_uni_lrn_fwd_kernel_f32<avx2>::nchw_body_sse42(
+    int tail, int HW, prop_kind_t pk, Xbyak::Xmm xmask_lo, Xbyak::Xmm xmask_hi,
+    Xbyak::Xmm xe_lo, Xbyak::Xmm xe_hi, Xbyak::Xmm xsum_lo, Xbyak::Xmm xsum_hi);
+template <>
+jit_uni_lrn_fwd_kernel_f32<avx2>::jit_uni_lrn_fwd_kernel_f32(
+    struct nchw_across J, float A, float K, prop_kind_t pk, void *code_ptr,
+    size_t code_size);
+template <>
+jit_uni_lrn_fwd_kernel_f32<sse42>::jit_uni_lrn_fwd_kernel_f32(
+    struct nchw_across J, float A, float K, prop_kind_t pk, void *code_ptr,
+    size_t code_size);
+
+extern template struct jit_uni_lrn_fwd_kernel_f32<sse42>;
+extern template struct jit_uni_lrn_fwd_kernel_f32<avx2>;
+extern template struct jit_uni_lrn_bwd_kernel_f32<avx2>;
+
 }
 }
 }
diff --git a/src/cpu/jit_uni_pool_kernel_f32.hpp b/src/cpu/jit_uni_pool_kernel_f32.hpp
index 79327a85214..60df28e38c5 100644
--- a/src/cpu/jit_uni_pool_kernel_f32.hpp
+++ b/src/cpu/jit_uni_pool_kernel_f32.hpp
@@ -184,6 +184,11 @@ struct jit_uni_pool_kernel_f32: public jit_generator {
     }
 };
 
+// Explicit instantiations in jit_uni_pool_kernel_f32.cpp.
+extern template struct jit_uni_pool_kernel_f32<sse42>;
+extern template struct jit_uni_pool_kernel_f32<avx>;
+extern template struct jit_uni_pool_kernel_f32<avx512_common>;
+
 }
 }
 }