Merge pull request #22 from gartung/v2.12.0-stack-unwind

V2.12.0 with  stack unwind patch and memory debug

Author: gartung

third_party/mkl_dnn/mkldnn_v1.BUILD

@@ -101,7 +101,9 @@ _COPTS_LIST = select({
     "-UUSE_MKL",
     "-UUSE_CBLAS",
     "-DDNNL_ENABLE_MAX_CPU_ISA",
+    "-DDNNL_ENABLE_MEM_DEBUG",
     "-fno-omit-frame-pointer",
+    "-g",
 ] + tf_openmp_copts()
 
 _INCLUDES_LIST = [

third_party/mkl_dnn/onednn_unwind_stack.patch

@@ -0,0 +1,50 @@
+diff --git a/src/cpu/x64/jit_generator.hpp b/src/cpu/x64/jit_generator.hpp
+index a6b1c3397e..c8ebe0f33b 100644
+--- a/src/cpu/x64/jit_generator.hpp
++++ b/src/cpu/x64/jit_generator.hpp
+@@ -103,8 +103,8 @@ static inline void tc_configure_tile(
+ 
+ #ifdef XBYAK64
+ constexpr Xbyak::Operand::Code abi_save_gpr_regs[] = {
+-        Xbyak::Operand::RBX,
+         Xbyak::Operand::RBP,
++        Xbyak::Operand::RBX,
+         Xbyak::Operand::R12,
+         Xbyak::Operand::R13,
+         Xbyak::Operand::R14,
+@@ -194,11 +194,17 @@ public:
+                 uni_vmovdqu(ptr[rsp + i * xmm_len],
+                         Xbyak::Xmm(xmm_to_preserve_start + i));
+         }
+-        for (size_t i = 0; i < num_abi_save_gpr_regs; ++i)
++        for (size_t i = 0; i < num_abi_save_gpr_regs; ++i) {
+             push(Xbyak::Reg64(abi_save_gpr_regs[i]));
++            // Stack magic: save rsp into rbp state to be able to unwind stack.
++            if (i == 0) mov(rbp, rsp);
++        }
++#ifndef DNNL_ENABLE_MEM_DEBUG
++        // do not use RBP in mem debug mode to enable backtracing from jit code
+         if (is_valid_isa(avx512_core)) {
+             mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt);
+         }
++#endif
+ 
+ #ifdef DNNL_ENABLE_MEM_DEBUG
+         // This section poisons vector registers with NaNs to catch situations
+@@ -277,6 +283,8 @@ public:
+ 
+         int scale = 0;
+ 
++#ifndef DNNL_ENABLE_MEM_DEBUG
++        // do not use RBP in mem debug mode to enable backtracing from jit code
+         if (EVEX_max_8b_offt <= offt && offt < 3 * EVEX_max_8b_offt) {
+             offt = offt - 2 * EVEX_max_8b_offt;
+             scale = 1;
+@@ -285,6 +293,7 @@ public:
+             offt = offt - 4 * EVEX_max_8b_offt;
+             scale = 2;
+         }
++#endif
+ 
+         auto re = RegExp() + base + offt;
+         if (scale) re = re + reg_EVEX_max_8b_offt * scale;