Enable Fused-Multiply-Add (FMA) and F16C/CVT16 vector extensions on MSVC (#375)

* Enable Fused-Multiply-Add (FMA) instructions on MSVC

__FMA__ macro does not exist in MSVC

* Enable F16C/CVT16 vector extensions on MSVC

__F16C__ macro does not exist in MSVC, but is implied with AVX2/AVX512

* MSVC cvt intrinsics

* Add __SSE3__ macro for MSVC too because why not

even though it's not currently used for anything when AVX is defined

diff --git a/ggml.c b/ggml.c
index 222d199be0285ac893c6e8e9abff783a0b8d6624..efe9316bb01ad4b2288399135aae8ee934576303 100644 (file)
--- a/ggml.c
+++ b/ggml.c
@@ -79,6 +79,19 @@ static int sched_yield (void) {
 typedef void* thread_ret_t;
 #endif
 
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
 #ifdef __HAIKU__
 #define static_assert(cond, msg) _Static_assert(cond, msg)
 #endif
@@ -172,8 +185,13 @@ typedef double ggml_float;
 
 #ifdef __F16C__
 
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
 #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
 
 #elif defined(__POWER9_VECTOR__)