[CPU][ARM64] Implemented JIT Emitter for Eltwise SoftPlus Operation

[srinjoydutta03]

Details:

• Implemented and added jit_softplus_emitter derived class for element wise softplus operation
• Added entry Algorithm::EltwiseSoftRelu, in executors/aarch64 as one of the supported algorithms
• Added entry in the get_supported_precisions and create_eltwise_emitters in kernel/aarch64
• Added utils::ActivationTypes::SoftPlus in jit kernel check in the tests in activation.cpp

Tests:

Passed all local tests using ./bin/arm64/Release/ov_cpu_func_tests --gtest_filter="*smoke*Activation*SoftPlus*"

Tickets:

• Closes [Good First Issue] [ARM]: Implement CPU plugin just-in-time emitter for SoftPlus operation #24109

CC: @a-sidorova

[a-sidorova]

Good job 👍🏼

Left some comments

[a-sidorova]

There might be situation when in_vec_idxs[0] == out_vec_idxs[0] - the same src and dst register.
In this case after this line vmm_src will contain not original values and we won't be able to use this register (for example, on L2782 the register may have incorrect original data).

May I ask you to handle such possible cases?

[srinjoydutta03]

Oh, I see. Will first moving to an auxillary register from the source and then applying the fmaxnm operation be enough in this case?

[a-sidorova]

Since part of aux vec registers are passed to exp_emitter, I think we need to write some function with exp_emitter->get_aux_vecs_count() calling here? If exp_emitter requires (or will require) more than 8 registers, some vector registers will be spilled (saved on stack) during exp_emitter->emit_code() call.

What do you think?

[srinjoydutta03]

I initially thought to put exp_emitter->get_aux_vecs_count()+4, as exp_emitter->get_aux_vecs_count() is returning 4. I tried with using more than 9 aux registers (few solely for exp_emitter for its own manipulation), but it won't let me allocate that many. I noticed the same in the elu_emitter where the get get_aux_vecs_count() function for it returns max(exp_emitter->get_aux_vecs_count()+1ull, 2ull). I'm wondering if I can apply the same logic while allocating my auxillary registers.

Like below,

    const TReg vmm_aux0(aux_vec_idxs[0]);
    const TReg vmm_aux1(aux_vec_idxs[1]);
    const TReg vmm_aux2(aux_vec_idxs[2]);
    const TReg vmm_aux3(aux_vec_idxs[3]);
    const TReg vmm_aux4(aux_vec_idxs[exp_aux_count]);
    const TReg vmm_aux5(aux_vec_idxs[exp_aux_count + 1]);
    const TReg vmm_mask(aux_vec_idxs[exp_aux_count + 2]);
    const TReg vmm_neg_mask(aux_vec_idxs[exp_aux_count + 3]);

Is this approach correct?

[a-sidorova]

Could we use FMA instruction here (fmla)?

[srinjoydutta03]

Sure, I will modify it. Thanks

[a-sidorova]

Can we have only one mask register? I see that vmm_neg_mask is free (after L2778) when we initialize vmm_mask on L2782. Looks like we can reuse vmm_neg_mask in the code part which handle big values.

[srinjoydutta03]

Sure, will update that

[a-sidorova]

May I ask you to add please brief description of the implemented logic? What is the formula?

[srinjoydutta03]

The implemented logic manipulates the base formula softplus(x) = ln(1+e^x) = ln(1+e^r*2^n) into two separate formula for positive x and negative x,

• For positive we use softplus(x) = nln2 + ln((2^-(n-1)+2e^r)/2) to compute the log approximation
• For negative we use softplus(x) = ln(1+e^x) = ln(1+e^r*2^n) = ln2 + ln(1 + (e^r*2^(n-1) - 0.5)), to compute the log approximation

[a-sidorova]

Could you please elaborate why we have to calculate two log approximation?

[srinjoydutta03]

The base formula, softplus(x) = ln(1+e^x) = ln(1+e^r*2^n), where n is the quotient and r is the remainder when divided by ln2. We have two cases one for positive value of n, i.e., x >= 0 and for negative value of n, i.e. for x < 0.

• For positive values we finally arrive at the formula; softplus(x) = nln2 + ln((2^-(n-1)+2e^r)/2), here we can calculate the log approximation since for n > 0, 2^-(n-1) will be small and will be in desirable range. The problem occurs when n is negative, which makes the 2^-(n-1) value very large and cannot be approximated with 8th order polynomial

• For negative values we can inherently use this formula softplus(x) = ln(1+e^x) = ln(1+e^r*2^n), i have done these manipulations as shown in the comment in the code to get y in ln(1+y) in between [-0.5, 0] where we can accurately approximate for negative n