- oneDNN update 1.3 -> 1.5 (#25202)

* - introducing oneDNN 1.6

test=develop

* - onednn 1.5

test=develop

* - oneDNN 1.5

test=develop

* - oneDNN 1.5 bugfix

test=develop

cmake/external/mkldnn.cmake

@@ -20,7 +20,7 @@ SET(MKLDNN_SOURCE_DIR     ${THIRD_PARTY_PATH}/mkldnn/src/extern_mkldnn)
 SET(MKLDNN_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/mkldnn)
 SET(MKLDNN_INC_DIR        "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
 SET(MKLDNN_REPOSITORY     https://github.com/intel/mkl-dnn.git)
-SET(MKLDNN_TAG            fb95345126ade4c54f5507e580a5f5da8d30a515)
+SET(MKLDNN_TAG            1ea812f4f5aa1bd989372a23ab50d0f0f81ee677)
 
 # Introduce variables:
 # * CMAKE_INSTALL_LIBDIR

python/paddle/fluid/tests/unittests/mkldnn/test_matmul_mkldnn_op.py

@@ -105,8 +105,11 @@ class TestDnnlMatMulOpInt8NoScales(TestDnnlMatMulOp):
 
 
 class TestDnnlMatMulOpInt8(TestDnnlMatMulOp):
+    # Due to limitation in int8 matmul implementation
+    # on older platforms (BDW, SKX) we needed to reduce
+    # range from [-127, 127] to [-63, 63]
     def quantize(self, tensor):
-        scale = 127. / np.abs(np.amax(tensor))
+        scale = 63. / np.abs(np.amax(tensor))
         quantized = np.round(scale * tensor).astype("int8")
         return scale, quantized