Use hardware-accelerated _Float16 type with x86 GCC

Maratyszcza · Jun 16, 2024 · c2c86b8 · c2c86b8
1 parent bc497e2
commit c2c86b8
Show file tree

Hide file tree

Showing 3 changed files with 43 additions and 2 deletions.
diff --git a/.github/workflows/cmake.yml b/.github/workflows/cmake.yml
@@ -26,6 +26,24 @@ jobs:
         run: cmake --build build --parallel
       - name: Test
         run: ctest --test-dir build --parallel --output-on-failure
+  cmake-linux-x86_64-f16c:
+    runs-on: ubuntu-24.04  # required for gcc >= 12
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+      - name: Update apt
+        run: sudo apt update
+      - name: Install ninja
+        run: sudo apt install ninja-build
+      - name: Configure
+        run: cmake -Bbuild -S. -G Ninja -DCMAKE_BUILD_TYPE=Release -DFP16_BUILD_COMPARATIVE_BENCHMARKS=ON
+        env:
+          CFLAGS: "-mf16c"
+          CXXFLAGS: "-mf16c"
+      - name: Build
+        run: cmake --build build --parallel
+      - name: Test
+        run: ctest --test-dir build --parallel --output-on-failure
   cmake-linux-x86:
     runs-on: ubuntu-20.04
     timeout-minutes: 15

diff --git a/include/fp16/fp16.h b/include/fp16/fp16.h
@@ -107,7 +107,13 @@ static inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) {
  * floating-point operations and bitcasts between integer and floating-point variables.
  */
 static inline float fp16_ieee_to_fp32_value(uint16_t h) {
-#if FP16_USE_FP16_TYPE
+#if FP16_USE_FLOAT16_TYPE
+	union {
+		uint16_t as_bits;
+		_Float16 as_value;
+	} fp16 = { h };
+	return (float) fp16.as_value;
+#elif FP16_USE_FP16_TYPE
 	union {
 		uint16_t as_bits;
 		__fp16 as_value;
@@ -230,7 +236,13 @@ static inline float fp16_ieee_to_fp32_value(uint16_t h) {
  * floating-point operations and bitcasts between integer and floating-point variables.
  */
 static inline uint16_t fp16_ieee_from_fp32_value(float f) {
-#if FP16_USE_FP16_TYPE
+#if FP16_USE_FLOAT16_TYPE
+	union {
+		_Float16 as_value;
+		uint16_t as_bits;
+	} fp16 = { (_Float16) f };
+	return fp16.as_bits;
+#elif FP16_USE_FP16_TYPE
 	union {
 		__fp16 as_value;
 		uint16_t as_bits;

diff --git a/include/fp16/macros.h b/include/fp16/macros.h
@@ -3,6 +3,17 @@
 #define FP16_MACROS_H
 
 
+#ifndef FP16_USE_FLOAT16_TYPE
+	#if !defined(__clang__) && !defined(__INTEL_COMPILER) && defined(__GNUC__) && (__GNUC__ >= 12)
+		#if defined(__F16C__)
+			#define FP16_USE_FLOAT16_TYPE 1
+		#endif
+	#endif
+	#if !defined(FP16_USE_FLOAT16_TYPE)
+		#define FP16_USE_FLOAT16_TYPE 0
+	#endif  // !defined(FP16_USE_FLOAT16_TYPE)
+#endif  // !defined(FP16_USE_FLOAT16_TYPE)
+
 #ifndef FP16_USE_FP16_TYPE
 	#if defined(__clang__)
 		#if defined(__F16C__) || defined(__aarch64__)