Refactor cpu_get_num_math() with cores.h header

jart · jart · commit 484a1b266c11 · 2024-09-18T03:31:19.000-07:00
diff --git a/llama.cpp/common.cpp b/llama.cpp/common.cpp
@@ -90,7 +90,6 @@ using json = nlohmann::ordered_json;
 //
 
 int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
 
 //
 // CLI argument parsing
diff --git a/llama.cpp/common.h b/llama.cpp/common.h
@@ -6,6 +6,7 @@
 #pragma once
 
 #include "llamafile/log.h"
+#include "llama.cpp/cores.h"
 #include "llama.h"
 
 #include "sampling.h"
@@ -61,7 +62,6 @@ struct llama_control_vector_load_info;
 //
 
 int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
 
 //
 // CLI argument parsing
diff --git a/llama.cpp/cores.cpp b/llama.cpp/cores.cpp
@@ -1,3 +1,5 @@
+#include "cores.h"
+
 #include <cosmo.h>
 #include <fstream>
 #include <thread>
@@ -87,10 +89,7 @@ static int count_math_cpus(int cpu_count) {
 
 #endif // __x86_64__ && __linux__
 
-/**
- * Returns number of CPUs on system that are useful for math.
- */
-int cpu_get_num_math() {
+static int cpu_get_num_math_impl() {
 #if defined(__x86_64__) && (defined(__linux__) || defined(__COSMOPOLITAN__)) && !defined(__ANDROID__)
     int cpu_count = sysconf(_SC_NPROCESSORS_ONLN);
     if (cpu_count < 1) {
@@ -109,3 +108,16 @@ int cpu_get_num_math() {
 #endif
     return get_num_physical_cores();
 }
+
+/**
+ * Returns number of CPUs on system that are useful for math.
+ */
+int cpu_get_num_math() {
+    static bool once;
+    static int result;
+    if (!once) {
+        result = cpu_get_num_math_impl();
+        once = true;
+    }
+    return result;
+}
diff --git a/llama.cpp/cores.h b/llama.cpp/cores.h
@@ -0,0 +1,10 @@
+#pragma once
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int cpu_get_num_math(void);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/llama.cpp/llama-bench/llama-bench.cpp b/llama.cpp/llama-bench/llama-bench.cpp
@@ -24,6 +24,7 @@
 #include <sys/stat.h>
 #include <sys/auxv.h>
 #include <libc/intrin/x86.h>
+#include "llama.cpp/cores.h"
 #include <libc/sysv/consts/hwcap.h>
 
 #include "llama.cpp/ggml.h"
diff --git a/llamafile/ansiblas.h b/llamafile/ansiblas.h
@@ -20,7 +20,7 @@
 #include <cmath>
 #include <unistd.h>
 
-int cpu_get_num_math();
+#include "llama.cpp/cores.h"
 
 namespace {
 namespace ansiBLAS {
diff --git a/llamafile/core_manager.cpp b/llamafile/core_manager.cpp
@@ -19,7 +19,7 @@
 
 #include <assert.h>
 
-int cpu_get_num_math();
+#include "llama.cpp/cores.h"
 
 CoreManager g_core_manager;
 
diff --git a/llamafile/flags.cpp b/llamafile/flags.cpp
@@ -16,6 +16,7 @@
 // limitations under the License.
 
 #include "debug.h"
+#include "llama.cpp/cores.h"
 #include "llamafile.h"
 #include "trust.h"
 
@@ -67,8 +68,6 @@ int FLAG_verbose = 0;
 int FLAG_warmup = true;
 int FLAG_workers;
 
-int cpu_get_num_math();
-
 static wontreturn void usage(int rc, int fd) {
     tinyprint(fd, "usage: ", program_invocation_name, " -m MODEL -l [HOST:]PORT\n", NULL);
     exit(rc);
diff --git a/llamafile/sgemm_matmul_test.cpp b/llamafile/sgemm_matmul_test.cpp
@@ -29,8 +29,6 @@
 #define ITERATIONS 30
 #define ALLOC(n) (float *)memalign(4096, sizeof(float) * (n))
 
-int cpu_get_num_math();
-
 void llamafile_sgemm_openmp(long m, long n, long k, const void *A, long lda, const void *B,
                             long ldb, void *C, long ldc, int Atype, int Btype, int Ctype) {
     static int nth = cpu_get_num_math();
diff --git a/llamafile/sgemm_sss_test.cpp b/llamafile/sgemm_sss_test.cpp
@@ -28,8 +28,6 @@
 #define ITERATIONS 5
 #define ALLOC(n) (float *)memalign(4096, sizeof(float) * (n))
 
-int cpu_get_num_math();
-
 void llamafile_sgemm_openmp(long m, long n, long k, const void *A, long lda, const void *B,
                             long ldb, void *C, long ldc, int Atype, int Btype, int Ctype) {
     static int nth = cpu_get_num_math();
diff --git a/llamafile/sgemm_vecdot_test.cpp b/llamafile/sgemm_vecdot_test.cpp
@@ -27,8 +27,6 @@
 #define ITERATIONS 30
 #define ALLOC(n) (float *)memalign(4096, sizeof(float) * (n))
 
-int cpu_get_num_math();
-
 void llamafile_sgemm_openmp(long m, long n, long k, const void *A, long lda, const void *B,
                             long ldb, void *C, long ldc, int Atype, int Btype, int Ctype) {
     static int nth = cpu_get_num_math();
diff --git a/stable-diffusion.cpp/BUILD.mk b/stable-diffusion.cpp/BUILD.mk
@@ -24,7 +24,6 @@ $(STABLE_DIFFUSION_CPP_OBJS): private					\
 $(STABLE_DIFFUSION_CPP_OBJS): private					\
 		CXXFLAGS +=						\
 			-frtti						\
-			-Wno-alloc-size-larger-than			\
 			-Wno-deprecated-declarations
 
 o/$(MODE)/stable-diffusion.cpp/main:					\
diff --git a/stable-diffusion.cpp/darts.h b/stable-diffusion.cpp/darts.h
@@ -1459,7 +1459,7 @@ class DoubleArrayBuilder {
  private:
   enum { BLOCK_SIZE = 256 };
   enum { NUM_EXTRA_BLOCKS = 16 };
-  enum { NUM_EXTRAS = BLOCK_SIZE * NUM_EXTRA_BLOCKS };
+  enum { NUM_EXTRAS = (int)BLOCK_SIZE * (int)NUM_EXTRA_BLOCKS };
 
   enum { UPPER_MASK = 0xFF << 21 };
   enum { LOWER_MASK = 0xFF };
diff --git a/stable-diffusion.cpp/main.cpp b/stable-diffusion.cpp/main.cpp
@@ -5,6 +5,7 @@
 #include <random>
 #include <string>
 #include <vector>
+#include "llama.cpp/cores.h"
 #include <cosmo.h>
 
 // #include "preprocessing.hpp"
@@ -527,7 +528,6 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         exit(1);
     }
     if (params.n_threads <= 0) {
-        int cpu_get_num_math(void);
         params.n_threads = cpu_get_num_math();
     }
 
diff --git a/stable-diffusion.cpp/tae.hpp b/stable-diffusion.cpp/tae.hpp
@@ -191,8 +191,8 @@ struct TinyAutoEncoder : public GGMLRunner {
                     ggml_type wtype,
                     bool decoder_only = true)
         : decode_only(decoder_only),
-          taesd(decode_only),
           GGMLRunner(backend, wtype) {
+        taesd = TAESD(decode_only); // [jart] fix ub
         taesd.init(params_ctx, wtype);
     }
 
@@ -248,4 +248,4 @@ struct TinyAutoEncoder : public GGMLRunner {
     }
 };
 
-#endif  // __TAE_HPP__
+#endif  // __TAE_HPP__
diff --git a/whisper.cpp/main.cpp b/whisper.cpp/main.cpp
@@ -1,7 +1,9 @@
 // -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-
 // vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
+
 #include "llamafile/version.h"
 #include "llamafile/llamafile.h"
+#include "llama.cpp/cores.h"
 #include "common.h"
 
 #include "whisper.h"
@@ -25,11 +27,9 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-int cpu_get_num_math();
-
 // command-line parameters
 struct whisper_params {
-    int32_t n_threads     = cpu_get_num_math();
+    int32_t n_threads     = std::min(16, cpu_get_num_math());
     int32_t n_processors  = 1;
     int32_t offset_t_ms   = 0;
     int32_t offset_n      = 0;
diff --git a/whisper.cpp/whisper.cpp b/whisper.cpp/whisper.cpp
@@ -1,6 +1,7 @@
 // -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-
 // vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
 #include "llama.cpp/ggml-vector.h"
+#include "llama.cpp/cores.h"
 #include "whisper.h"
 
 #define GGML_USE_CUDA
@@ -4777,7 +4778,7 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
     struct whisper_full_params result = {
         /*.strategy          =*/ strategy,
 
-        /*.n_threads         =*/ std::min(4, (int32_t) std::thread::hardware_concurrency()),
+        /*.n_threads         =*/ std::min(8, (int32_t) cpu_get_num_math()),
         /*.n_max_text_ctx    =*/ 16384,
         /*.offset_ms         =*/ 0,
         /*.duration_ms       =*/ 0,
