oneDNN new features for next major version. (#439)

* oneDNN: add fpmath_mode requirements

* oneDNN: move quantization model description to int8 section

* oneDNN: update the int8 quantization model with asymetric and dynamic quant support

* oneDNN: add matmul to primitives supporting the `any` format.

* oneDNN: add support of broadcast semantic

* oneDNN: split scale and shift flags/parameters in batchnorm

* oneDNN: split scale and shift flags/parameters in layernorm

* oneDNN: split src/dst descriptor for normalization, eltwise, shuffle

* oneDNN: remove logsoftmax and handle it through softmax

* oneDNN: add new eltwise functions and some parametrizations

* oneDNN: add prelu primitive

* oneDNN: add reduction primitive

* oneDNN: new quantization mecanism

- split quantization parameter as per tensor properties
- now inputs qparams are used to compute compensations, and dequantize
  after op, and output qparams are applied before write to requantize.
- all scales/shift are passed at execution time, removes overhead of
  transfering host scales to device for implementions. Also allow
  better reuse of primitives for dynamic quantization.

* oneDNN: add binary post-op

* oneDNN: add AUGRU and LBR-AUGRU cells to RNN primitive

* oneDNN: remove op_desc

* oneDNN: fix outdated doc in header

* fixup! oneDNN: new quantization mecanism

* fixup! oneDNN: update the int8 quantization model with asymetric and dynamic quant support

* oneDNN: various warning/rendering fixes

- Use relative paths for includes
- fix missing replacements
- fix missing references
- fix itemizations
- fix latex rendering

* oneDNN: int8: fix notations

* oneDNN: modify header to match spec changes

* oneDNN: update post-ops support and naming

Author: mgouicem

source/elements/oneDNN/Doxyfile

@@ -267,6 +267,41 @@ TAB_SIZE               = 4
 
 ALIASES                =
 
+ALIASES += src="\f$\src\f$"
+ALIASES += srclayer="\f$\srclayer\f$"
+ALIASES += srclayerattention="\f$\srclayerattention\f$"
+ALIASES += srciter="\f$\srciter\f$"
+ALIASES += srciterc="\f$\srciterc\f$"
+ALIASES += weights="\f$\weights\f$"
+ALIASES += weightslayer="\f$\weightslayer\f$"
+ALIASES += weightsiter="\f$\weightsiter\f$"
+ALIASES += weightspeephole="\f$\weightspeephole\f$"
+ALIASES += weightsprojection="\f$\weightsprojection\f$"
+ALIASES += bias="\f$\bias\f$"
+ALIASES += dst="\f$\dst\f$"
+ALIASES += dstlayer="\f$\dstlayer\f$"
+ALIASES += dstiter="\f$\dstiter\f$"
+ALIASES += dstiterc="\f$\dstiterc\f$"
+ALIASES += diffsrc="\f$\diffsrc\f$"
+ALIASES += diffsrclayer="\f$\diffsrclayer\f$"
+ALIASES += diffsrclayerattention="\f$\diffsrclayerattention\f$"
+ALIASES += diffsrciter="\f$\diffsrciter\f$"
+ALIASES += diffsrciterc="\f$\diffsrciterc\f$"
+ALIASES += diffweights="\f$\diffweights\f$"
+ALIASES += diffweightslayer="\f$\diffweightslayer\f$"
+ALIASES += diffweightsiter="\f$\diffweightsiter\f$"
+ALIASES += diffweightspeephole="\f$\diffweightspeephole\f$"
+ALIASES += diffweightsprojection="\f$\diffweightsprojection\f$"
+ALIASES += diffbias="\f$\diffbias\f$"
+ALIASES += diffdst="\f$\diffdst\f$"
+ALIASES += diffdstlayer="\f$\diffdstlayer\f$"
+ALIASES += diffdstiter="\f$\diffdstiter\f$"
+ALIASES += diffdstiterc="\f$\diffdstiterc\f$"
+ALIASES += diffgamma="\f$\diffgamma\f$"
+ALIASES += diffbeta="\f$\diffbeta\f$"
+ALIASES += workspace="\f$\workspace\f$"
+
+
 # Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C sources
 # only. Doxygen will then generate output that is more tailored for C. For
 # instance, some of the names that are used will be different. The list of all