fix: compiler built-in function calling convention workaround

Signed-off-by: usamoi <[email protected]>

Author: usamoi

Cargo.lock

@@ -1,2 +1,3 @@
 *.s
-*.o
+*.o
+*.out

crates/c/.gitignore

@@ -3,5 +3,8 @@ name = "c"
 version.workspace = true
 edition.workspace = true
 
+[dependencies]
+half = { version = "~2.3", features = ["use-intrinsics"] }
+
 [build-dependencies]
 cc = "1.0"

crates/c/Cargo.toml

@@ -1,9 +1,10 @@
 fn main() {
-    println!("rerun-if-changed:src/c.h");
-    println!("rerun-if-changed:src/c.c");
+    println!("cargo:rerun-if-changed=src/c.h");
+    println!("cargo:rerun-if-changed=src/c.c");
     cc::Build::new()
         .compiler("/usr/bin/clang-16")
         .file("./src/c.c")
         .opt_level(3)
+        .debug(true)
         .compile("pgvectorsc");
 }

crates/c/build.rs

@@ -8,8 +8,7 @@
 #if defined(__x86_64__)
 
 __attribute__((target("avx512fp16,bmi2"))) extern float
-v_f16_cosine_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
-                    size_t n) {
+v_f16_cosine_axv512(_Float16 *a, _Float16 *b, size_t n) {
   __m512h xy = _mm512_set1_ph(0);
   __m512h xx = _mm512_set1_ph(0);
   __m512h yy = _mm512_set1_ph(0);
@@ -35,8 +34,7 @@ v_f16_cosine_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
 }
 
 __attribute__((target("avx512fp16,bmi2"))) extern float
-v_f16_dot_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
-                 size_t n) {
+v_f16_dot_axv512(_Float16 *a, _Float16 *b, size_t n) {
   __m512h xy = _mm512_set1_ph(0);
 
   while (n >= 32) {
@@ -55,8 +53,7 @@ v_f16_dot_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
 }
 
 __attribute__((target("avx512fp16,bmi2"))) extern float
-v_f16_sl2_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
-                 size_t n) {
+v_f16_sl2_axv512(_Float16 *a, _Float16 *b, size_t n) {
   __m512h dd = _mm512_set1_ph(0);
 
   while (n >= 32) {
@@ -78,8 +75,7 @@ v_f16_sl2_axv512(_Float16 const *restrict a, _Float16 const *restrict b,
 }
 
 __attribute__((target("avx2"))) extern float
-v_f16_cosine_axv2(_Float16 const *restrict a, _Float16 const *restrict b,
-                  size_t n) {
+v_f16_cosine_axv2(_Float16 *a, _Float16 *b, size_t n) {
   float xy = 0;
   float xx = 0;
   float yy = 0;
@@ -95,8 +91,7 @@ v_f16_cosine_axv2(_Float16 const *restrict a, _Float16 const *restrict b,
 }
 
 __attribute__((target("avx2"))) extern float
-v_f16_dot_axv2(_Float16 const *restrict a, _Float16 const *restrict b,
-               size_t n) {
+v_f16_dot_axv2(_Float16 *a, _Float16 *b, size_t n) {
   float xy = 0;
 #pragma clang loop vectorize_width(8)
   for (size_t i = 0; i < n; i++) {
@@ -108,8 +103,7 @@ v_f16_dot_axv2(_Float16 const *restrict a, _Float16 const *restrict b,
 }
 
 __attribute__((target("avx2"))) extern float
-v_f16_sl2_axv2(_Float16 const *restrict a, _Float16 const *restrict b,
-               size_t n) {
+v_f16_sl2_axv2(_Float16 *a, _Float16 *b, size_t n) {
   float dd = 0;
 #pragma clang loop vectorize_width(8)
   for (size_t i = 0; i < n; i++) {

crates/c/src/c.c

@@ -3,11 +3,11 @@
 
 #if defined(__x86_64__)
 
-extern float v_f16_cosine_axv512(_Float16 const *, _Float16 const *, size_t n);
-extern float v_f16_dot_axv512(_Float16 const *, _Float16 const *, size_t n);
-extern float v_f16_sl2_axv512(_Float16 const *, _Float16 const *, size_t n);
-extern float v_f16_cosine_axv2(_Float16 const *, _Float16 const *, size_t n);
-extern float v_f16_dot_axv2(_Float16 const *, _Float16 const *, size_t n);
-extern float v_f16_sl2_axv2(_Float16 const *, _Float16 const *, size_t n);
+extern float v_f16_cosine_axv512(_Float16 *, _Float16 *, size_t n);
+extern float v_f16_dot_axv512(_Float16 *, _Float16 *, size_t n);
+extern float v_f16_sl2_axv512(_Float16 *, _Float16 *, size_t n);
+extern float v_f16_cosine_axv2(_Float16 *, _Float16 *, size_t n);
+extern float v_f16_dot_axv2(_Float16 *, _Float16 *, size_t n);
+extern float v_f16_sl2_axv2(_Float16 *, _Float16 *, size_t n);
 
 #endif

crates/c/src/c.h

@@ -1,3 +1,4 @@
+#[cfg(target_arch = "x86_64")]
 #[link(name = "pgvectorsc", kind = "static")]
 extern "C" {
     pub fn v_f16_cosine_axv512(a: *const u16, b: *const u16, n: usize) -> f32;
@@ -7,3 +8,16 @@ extern "C" {
     pub fn v_f16_dot_axv2(a: *const u16, b: *const u16, n: usize) -> f32;
     pub fn v_f16_sl2_axv2(a: *const u16, b: *const u16, n: usize) -> f32;
 }
+
+// `compiler_builtin` defines `__extendhfsf2` with integer calling convention.
+// However C compilers links `__extendhfsf2` with floating calling convention.
+// The code should be removed once Rust offically supports `f16`.
+
+#[no_mangle]
+#[linkage = "external"]
+extern "C" fn __extendhfsf2(f: f64) -> f32 {
+    unsafe {
+        let f: half::f16 = std::mem::transmute_copy(&f);
+        f.to_f32()
+    }
+}

crates/c/src/c.rs

@@ -1,3 +1,5 @@
+#![feature(linkage)]
+
 mod c;
 
 pub use self::c::*;