[FC] Add dequantize action to FC layer

[EunjuYang]

• This commit is temporary one to realize a fc layer with QINT weight.
• This should be fixed in the way to move the dequantize operations under the tensor op.
• dequantization overhead :

Without SWAP		With SWAP (LA=2)
FP32-FP32	QINT8-FP32	FP32-FP32	QINT8-FP32
0.162787	0.166505	0.0487955	0.127012

[baek2sm]

how about adding 'not supported yet' message till implementation?

[EunjuYang]

It intends to make a quantizerfor all QINT4, QINT8, QINT16 cases. AFIK, QINT4, 8, 16 all work with a quantizer. Please let me know if I missed something!

[baek2sm]

LGTM

[EunjuYang]

FYI)

I tested the case where QINT8-FP32 is used (weight dtype-tensor dtype).
Please refer to the result below:

• dequantization overhead (tested with EunjuYang@cfe0f2f):

$  ./nntrainer_simplefc 100 false 0 FP32 FP32
$  ./nntrainer_simplefc 100 false 0 QINT8 FP32
$  ./nntrainer_simplefc 100 true 2 FP32 FP32
$  ./nntrainer_simplefc 100 true 2 QINT8 FP32

Without SWAP		With SWAP (LA=2)
FP32-FP32	QINT8-FP32	FP32-FP32	QINT8-FP32
0.162787	0.166505	0.0487955	0.127012

bin file size is 400 MB vs. 100 MB.
Result was not tested (no suitable initialization method for scale is implemented yet)

[skykongkong8]

Just curious, I wonder how should it be expected to work at Tensor::dot().
Maybe something like:

Tensor::dot() {
...
int8_t data = getData();
int8_t input_data = input;
Tensor qinput = quantizer->dequantize(...);
GEMM(data, qinput.getData(), output);
...
}

?