ggml : fix missing cpu_set_t on emscripten

[ngxson]

cpu_set_t doesn't exist on emscripten, we need to add #ifdef __gnu_linux__. The feature was added in #8672

The same method was already used in the other part of the code (for numa):

llama.cpp/ggml/src/ggml.c

Lines 3134 to 3136 in 409dc4f

	#if defined(__gnu_linux__)
	cpu_set_t cpuset; // cpuset from numactl
	#else

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[slaren]

@max-krasnyansky @fmz please take a look.

[fmz]

@ngxson this should be handled by adding __GNU_SOURCE before including <sched.h> (pretty finicky, I know)

https://github.com/ggerganov/llama.cpp/pull/8672/files#diff-d4bb67eec77d05ad8d4056c50492de104bd17dc6e0b0f9919854aeeaa5002842L1250

Did something change?

[ngxson]

@fmz for android: yes, it's still there:

llama.cpp/ggml/src/CMakeLists.txt

Lines 1254 to 1256 in 409dc4f

	if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
	add_compile_definitions(_GNU_SOURCE)
	endif()

But since I can't test the android build, I think it's better to reduce the scope of this PR to only target emscripten. We can continue the discussion for android on #9324

For emscripten, even with #define __GNU_SOURCE at the beginning of the file, it cannot build.

To reproduce, follow this guile: https://github.com/ngxson/wllama?tab=readme-ov-file#how-to-compile-the-binary-yourself

Any idea though?

[ngxson]

Also, in other part of code:

llama.cpp/ggml/src/ggml.c

Lines 19179 to 19181 in 409dc4f

	// Android's libc implementation "bionic" does not support setting affinity
	#if defined(__gnu_linux__)
	static void set_numa_thread_affinity(int thread_n) {

So I think it will make more sense if we have the same check for ggml_thread_apply_affinity and ggml_thread_apply_priority? It's strange to see that we have different checks for the same feature.

[fmz]

Also, in other part of code:

llama.cpp/ggml/src/ggml.c

Lines 19179 to 19181 in 409dc4f

	// Android's libc implementation "bionic" does not support setting affinity
	#if defined(__gnu_linux__)
	static void set_numa_thread_affinity(int thread_n) {

So I think it will make more sense if we have the same check for ggml_thread_apply_affinity and ggml_thread_apply_priority? It's strange to see that we have different checks for the same feature.

yeah I would suggest you do that

@fmz for android: yes, it's still there:

llama.cpp/ggml/src/CMakeLists.txt

Lines 1254 to 1256 in 409dc4f

	if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
	add_compile_definitions(_GNU_SOURCE)
	endif()

But since I can't test the android build, I think it's better to reduce the scope of this PR to only target emscripten. We can continue the discussion for android on #9324

For emscripten, even with #define __GNU_SOURCE at the beginning of the file, it cannot build.

To reproduce, follow this guile: https://github.com/ngxson/wllama?tab=readme-ov-file#how-to-compile-the-binary-yourself

Any idea though?

Interesting...
Would it be possible to specifically check for emscripten?

[fmz]

Also, in other part of code:

llama.cpp/ggml/src/ggml.c

Lines 19179 to 19181 in 409dc4f

	// Android's libc implementation "bionic" does not support setting affinity
	#if defined(__gnu_linux__)
	static void set_numa_thread_affinity(int thread_n) {

So I think it will make more sense if we have the same check for ggml_thread_apply_affinity and ggml_thread_apply_priority? It's strange to see that we have different checks for the same feature.

I'm not sure I follow. They're both under the same #if/else branches

[ngxson]

what I mean is:

• ggml_get_numa_affinity is only enabled with #if defined(__gnu_linux__)
• ggml_thread_apply_affinity is currently enabled for ALL platforms (except Apple which uses a different implementation)

See the code:

llama.cpp/ggml/src/ggml.c

Lines 19525 to 19527 in 134bc38

	#elif defined(__APPLE__)
	#include <sys/types.h>
	#include <sys/resource.h>

Then below that:

llama.cpp/ggml/src/ggml.c

Lines 19558 to 19560 in 134bc38

	#else // posix?

My question is: can we simply apply the same #if defined(__gnu_linux__) used by ggml_get_numa_affinity?

[fmz]

llama.cpp/ggml/src/ggml.c

Ah yes
I see no harm in that

[ngxson]

OK thanks for the confirmation.

So the current PR is exactly that: the code branch using cpu_set_t is only enabled with #if defined(__gnu_linux__). Otherwise, it will be no-op.

This should fix compilation on both emscripten & android (issue #9324), cc @slaren can you give a review? Thanks!