ggml : move CPU backend to a separate file

[slaren]

Moves the ggml code specific to the CPU backend to a separate file.

This is an initial step to separate the core ggml library from the CPU backend. In the future, this will allow:

• Building other backends as a shared library, without having to link them to the CPU backend
• Building the core ggml library with only the base instruction set for the ABI, and load an optimized version of the CPU backend dynamically

Additionally:

• Removes the optimization interface, since it has dependencies to the CPU backend, and would be removed in ggml: new optimization interface ggml#988 regardless
• Removes the baby-llama example since it depends on the opt interface

[JohannesGaessler]

Are there also plans to split ggml-cpu.c into multiple smaller files like was done for CUDA?

(I did not really look at ggml.c and ggml-cpu.c since I think it's not feasible.)

[JohannesGaessler]

I didn't get around to it, but this PR reminds me that I also want to at some point remove the YAML log code again. It has become pretty outdated and nowadays there are better solutions for the things that I was originally using it for.

[slaren]

Are there also plans to split ggml-cpu.c into multiple smaller files like was done for CUDA?

Yes, I think that would be great. We should also adapt it to C++ and use templates to avoid duplicating the code of the operations for each type.

[ggerganov]

Looking into this now.

[ggerganov]

Isn't this going to produce thread sanitizer data race warnings on the is_first_call var?

llama.cpp/ggml/src/ggml.c

Lines 1424 to 1443 in bf95fff

	struct ggml_context * ggml_init(struct ggml_init_params params) {
	static volatile bool is_first_call = false;
	if (!is_first_call) {
	ggml_critical_section_start();
	if (!is_first_call) {
	// initialize time system (required on Windows)
	ggml_time_init();
	for (int i = 0; i < (1 << 16); ++i) {
	union {
	uint16_t u16;
	ggml_fp16_t fp16;
	} u = {i};
	ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
	}
	is_first_call = true;
	}
	ggml_critical_section_end();
	}

[chaxu01]

@slaren this commit 9f40989 breaks q4_0_4_8 on Arm CPUs, likely related to #10165.

The following command triggers the issue:
./bin/llama-cli -m llama-2-7b-chat.Q4_0_4_8.gguf -p "Write a code in C for bubble sorting" -n 32 -t 4 -ngl 0

The error output is:
Assertion failed: (!isnan(wp[i])Assertion failed: (!isnan(wp[i])Assertion failed: (!isnan(wp[i])Assertion failed: (!isnan(wp[i])), function ggml_compute_forward), function ggml_compute_forward), function ggml_compute_forward_soft_max_f32, file ggml-cpu.c, _soft_max_f32, file ggml-cpu.c, _soft_max_f32, file ggml-cpu.c, ), function ggml_compute_forwardline 8904.

This issue does not occur on commit 08828a6.

[QingtaoLi1]

@slaren Why is the variable is_first_call in ggml_init() in ggml.c set reversely, i.e. is false when first called, while true later?