Show the hash rate of each implementation

[furuame]

Show the hash rate of each implementation by using BUILD_BENCH option

[furuame]

@marktwtn Please reference the implementation in iotaledger/giota

[marktwtn]

The current implementation of getting the hash times would be calling an API
getHashTimes(ImplContext *impl_ctx, void *pow_ctx)

It would call the corresponding hardware-related function.
Take CPU - SSE as an example:

static uint64_t PoWSSE_getHashTimes(void *pow_ctx)
{
    uint64_t count = 0;
    PoW_SSE_Context *ctx = (PoW_SSE_Context *) pow_ctx;
    Pwork_struct *pitem = ctx->pitem;

    for (int i = 0; i < ctx->num_threads; i++) {
        if (pitem[i].ret >= 0 )
            count += (uint64_t)pitem[i].ret;
        else
            count += (uint64_t)(-(pitem[i].ret) + 1);
    }

    return count;
}

The GPU implementation is still under investigation.
The FPGA implementation would be completed by @ajblane .

[furuame]

@marktwtn About hash times of the OpenCL, you can take a look from giota. It seems trivial.

https://github.com/iotaledger/giota/blob/e34a409290248e6ff2205b06062d702690910f04/pow_cl.go#L94

[marktwtn]

hash rate = hash count / PoW execution time

There are two ways to get the PoW execution time.

1. Subtract the time before and after the doThePoW function.
2. Create an API like getHashCount to get the PoW execution time.
   The time would be slightly accurate since the time subtraction would be done in the doThePoW function.
   It excludes the function call, the PoW initialization part and other things. It focuses on the PoW execution time.

However, if the time unit is second, I think the PoW execution time may have no difference at all.

I prefer the method 2.
Is there any suggestion?

And I am thinking that the hash rate should be got with the API, too.

[furuame]

Method 2 might be better to get the time.

But I'm not sure do we really need "three" APIs to achieve the goal of this enhancement?

[marktwtn]

But I'm not sure do we really need "three" APIs to achieve the goal of this issue?

How about

1. Create the getHashRate API and remove the others.

or

2. Create an API like getPoWContext which returns the pointer of PoW-related information. A structure contains hash count and PoW execution time.

[furuame]

2. Create an API like `getPoWContext` which returns the pointer of PoW-related information. A structure contains `hash count` and `PoW execution time`.

Let's do this! Create an API getPoWInfo to record the time and hashcount.

[marktwtn]

Here are the modifications I am going to do:

The return pointer of getPoWInfo function should be type cast to get the hash count and PoW execution time value.

I assume the PoW information structure of each implementation is the same.
Then I would like to extract it to the header file named common.h.
The concept is the same as the common.h header file in tests directory.

[furuame]

The return pointer of getPoWInfo function should be type cast to get the hash count and PoW execution time value.

It's no need to return a pointer to a struct. Just return the value of the struct.

I assume the PoW information structure of each implementation is the same.
Then I would like to extract it to the header file named common.h.
The concept is the same as the common.h header file in tests directory.

Good idea. After merging this PR, I will refactor the code to follow the OOP style.

[marktwtn]

This is what I am going to do.

1. Get the hash rate n times -> [ h1, h2, ......, hn ]

2. Calculate the standard deviation -> sqrt{ [ (h1 - avg)^2 + ... + (hn - avg)^2 ] / n }

3. Show the result:

average hash rate: XXX kH/sec
with 95% confidence interval and error value (2 * standard deviation).

4. Show the hash rate with histogram.

---

@jserv is this what you expect to see?

[furuame]

This is what I am going to do.

1. Get the hash rate `n` times -> [ h`1`, h`2`, ......, h`n` ]

2. Calculate the `standard deviation` -> sqrt{ [ (h`1` - avg)^2 + ... + (h`n` - avg)^2 ] / n }

3. Show the result:

average hash rate: XXX kH/sec
with 95% confidence interval and error value (2 * standard deviation).

4. Show the hash rate with histogram.

The original goal of this issue is achieved.

Let's create another issue for these features.

Thanks for @marktwtn and @ajblane's contribution. Close.