Rework model/batch size configuration again

[jeremybobbin]

• If GPU_INDICIES contains GPUs that are of different models, the script exits. This assumption makes (future)logging simpler.
• Batch size is calculated based on a model's batch size multiplier(seen below), precision, and GPU memory:

  resnet50='5  + 1/3'
 resnet152='2  + 2/3'
inception3='5  + 1/3'
inception4='1  + 1/3'
     vgg16='5  + 1/3'
   alexnet='42 + 2/3'
    ssd300='2  + 2/3'

• benchmark.sh's options are now position independent:

./benchmark.sh -l 2 -h 4     # low/high - benchmarks 2 GPUs, 3 GPUs and 4GPUs at a time.

Options:

i - GPU_INDEX
l - MIN_NUM_GPU
h - MAX_NUM_GPU
n - ITERATIONS
b - NUM_BATCHES
c - SETTING
v - GPU_VENDOR
t - THERMAL_INTERVAL

• -n has implemented the functionality of batch_benchmark.sh

This:

./batch_benchmark.sh 1 1 1 100 2 config_resnet50_replicated_fp32_train_syn

Is now this:

./benchmark.sh -h 1 -n 1 -b 100 -t 2 -c config_resnet50_replicated_fp32_train_syn

[jeremybobbin]

Issues:

• requires bc, fails mysteriously otherwise
• CPU_NAME is not set