Skip to content

Distribute and run LLMs with a single file.

Upgrade to Cosmopolitan v3.7.1

Justine Tunneyd0b5e8f2ac1d
diff --git a/build/config.mk b/build/config.mk
index ad69bbd4..5b9f9ce6 100644
--- a/build/config.mk
+++ b/build/config.mk
@@ -2,7 +2,7 @@
#── vi: set noet ft=make ts=8 sw=8 fenc=utf-8 :vi ────────────────────┘
PREFIX = /usr/local
-COSMOCC = .cosmocc/3.6.2
+COSMOCC = .cosmocc/3.7.1
TOOLCHAIN = $(COSMOCC)/bin/cosmo
CC = $(TOOLCHAIN)cc
@@ -52,5 +52,5 @@ clean:; rm -rf o
.PHONY: distclean
distclean:; rm -rf o .cosmocc
-.cosmocc/3.6.2:
- build/download-cosmocc.sh $@ 3.6.2 268aa82d9bfd774f76951b250f87b8edcefd5c754b8b409e1639641e8bd8d5bc
+.cosmocc/3.7.1:
+ build/download-cosmocc.sh $@ 3.7.1 13b65b0e659b493bd82f3d0a319d0265d66f849839e484aa2a54191024711e85
diff --git a/llama.cpp/ggml.c b/llama.cpp/ggml.c
index 7a3ea652..70a427c3 100644
--- a/llama.cpp/ggml.c
+++ b/llama.cpp/ggml.c
@@ -1810,60 +1810,26 @@ inline static void ggml_critical_section_start(void) {
}
}
-#define GGML_BARRIER_IMPL \
- if (params->shared->n_threads == 1) \
- return; \
- int n = params->shared->n_threads; \
- atomic_int * count = &params->shared->n_barrier; \
- atomic_uint * phase = &params->shared->n_barrier_passed[params->ith].i; \
- unsigned i = atomic_load_explicit(phase, memory_order_relaxed); \
- if (atomic_fetch_add_explicit(count, 1, memory_order_acq_rel) == n - 1) { \
- atomic_store_explicit(count, 0, memory_order_relaxed); \
- for (int j = 0; j < n; ++j) { \
- atomic_store_explicit(&params->shared->n_barrier_passed[j].i, \
- i + 1, memory_order_relaxed); \
- } \
- atomic_thread_fence(memory_order_release); \
- } else { \
- while (atomic_load_explicit(phase, memory_order_relaxed) == i) \
- pthread_pause_np(); \
- atomic_thread_fence(memory_order_acquire); \
- }
-
-#ifndef __aarch64__
-
void ggml_barrier(const struct ggml_compute_params * params) {
- GGML_BARRIER_IMPL
-}
-
-#else
-
-static void ggml_barrier_default(const struct ggml_compute_params * params) {
- GGML_BARRIER_IMPL
-}
-
-#pragma GCC push_options
-#pragma GCC target ("arch=armv8.1-a") // shaves microseconds per call on m2
-static void ggml_barrier_armv81a(const struct ggml_compute_params * params) {
- GGML_BARRIER_IMPL
-}
-#pragma GCC pop_options
-
-void ggml_barrier(const struct ggml_compute_params * params) {
- static void (*impl)(const struct ggml_compute_params *);
- if (!impl) {
- long hwcap = getauxval(AT_HWCAP);
- if (hwcap & HWCAP_ATOMICS) {
- impl = ggml_barrier_armv81a;
- } else {
- impl = ggml_barrier_default;
- }
+ if (params->shared->n_threads == 1)
+ return;
+ int n = params->shared->n_threads;
+ atomic_int * count = &params->shared->n_barrier;
+ atomic_uint * phase = &params->shared->n_barrier_passed[params->ith].i;
+ unsigned i = atomic_load_explicit(phase, memory_order_relaxed);
+ if (atomic_fetch_add_explicit(count, 1, memory_order_acq_rel) == n - 1) {
+ atomic_store_explicit(count, 0, memory_order_relaxed);
+ for (int j = 0; j < n; ++j)
+ atomic_store_explicit(&params->shared->n_barrier_passed[j].i,
+ i + 1, memory_order_relaxed);
+ atomic_thread_fence(memory_order_release);
+ } else {
+ while (atomic_load_explicit(phase, memory_order_relaxed) == i)
+ pthread_pause_np();
+ atomic_thread_fence(memory_order_acquire);
}
- return impl(params);
}
-#endif // __aarch64__
-
// TODO: make this somehow automatically executed
// some sort of "sentry" mechanism
inline static void ggml_critical_section_end(void) {