Merge mainline llama.cpp (#3)

* Merging mainline - WIP

* Merging mainline - WIP

AVX2 and CUDA appear to work.
CUDA performance seems slightly (~1-2%) lower as it is so often
the case with llama.cpp/ggml after some "improvements" have been made.

* Merging mainline - fix Metal

* Remove check

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 46 insertions, 0 deletions

diff --git a/examples/gguf-hash/deps/rotate-bits/rotate-bits.h b/examples/gguf-hash/deps/rotate-bits/rotate-bits.h
new file mode 100644
index 00000000..75c4881f
--- /dev/null
+++ b/examples/gguf-hash/deps/rotate-bits/rotate-bits.h
@@ -0,0 +1,46 @@
+
+
+#ifndef __ROTATE_DEFS_H
+#define __ROTATE_DEFS_H
+
+#ifdef _MSC_VER
+
+#include <stdlib.h>
+
+#define ROTL32(v, n) _rotl((v), (n))
+#define ROTL64(v, n) _rotl64((v), (n))
+
+#define ROTR32(v, n) _rotr((v), (n))
+#define ROTR64(v, n) _rotr64((v), (n))
+
+#else
+
+#include <stdint.h>
+
+#define U8V(v) ((uint8_t)(v) & 0xFFU)
+#define U16V(v) ((uint16_t)(v) & 0xFFFFU)
+#define U32V(v) ((uint32_t)(v) & 0xFFFFFFFFU)
+#define U64V(v) ((uint64_t)(v) & 0xFFFFFFFFFFFFFFFFU)
+
+#define ROTL32(v, n) \
+  (U32V((uint32_t)(v) << (n)) | ((uint32_t)(v) >> (32 - (n))))
+
+// tests fail if we don't have this cast...
+#define ROTL64(v, n) \
+  (U64V((uint64_t)(v) << (n)) | ((uint64_t)(v) >> (64 - (n))))
+
+#define ROTR32(v, n) ROTL32(v, 32 - (n))
+#define ROTR64(v, n) ROTL64(v, 64 - (n))
+
+#endif
+
+#define ROTL8(v, n) \
+  (U8V((uint8_t)(v) << (n)) | ((uint8_t)(v) >> (8 - (n))))
+
+#define ROTL16(v, n) \
+  (U16V((uint16_t)(v) << (n)) | ((uint16_t)(v) >> (16 - (n))))
+
+#define ROTR8(v, n) ROTL8(v, 8 - (n))
+#define ROTR16(v, n) ROTL16(v, 16 - (n))
+
+#endif