ggml : implement GEGLU_ERF and GEGLU_QUICK ops

[CISC]

Complimentary to the other GLU ops, used in mtmd.

Implemented for all currently GLU supported backends, except GEGLU_ERF in Vulkan due to missing erf.

[qnixsynapse]

SYCL code LGTM :)

[jeffbolznv]

Vulkan code looks fine. I wonder how much precision erf needs and if we could just handcode something in the Vulkan shader.

[CISC]

Vulkan code looks fine. I wonder how much precision erf needs and if we could just handcode something in the Vulkan shader.

The following seems to work for Metal, so should be good enough for Vulkan:

llama.cpp/ggml/src/ggml-metal/ggml-metal.metal

Lines 1099 to 1115 in 38593bc

	// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
	// ref: https://www.johndcook.com/blog/python_erf/
	constant float p_erf = 0.3275911f;
	constant float a1_erf = 0.254829592f;
	constant float a2_erf = -0.284496736f;
	constant float a3_erf = 1.421413741f;
	constant float a4_erf = -1.453152027f;
	constant float a5_erf = 1.061405429f;
	template<typename T>
	T erf_approx(T x) {
	T sign_x = sign(x);
	x = fabs(x);
	T t = 1.0f / (1.0f + p_erf * x);
	T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
	return sign_x * y;
	}

Edit: I'll give it a try.

[0cc4m]

The Vulkan code works on all of my devices.

[CISC]

@lhez I'll add OpenCL too, pending results of #14476

[lhez]

@CISC OpenCL looks good and works for me.