SVE2: Improves SVE2 8888 swizzling performance and important fixes (#15662)

* SVE2 was actually disabled in fdfbbce, this issue is fixed - The macro __ARM_FEATURE_SVE is only defined when the compilation target is set as -march=armv8-m+sve2 * Improves 8888 alpha-blending performance - Now, in In-Order AArch64 processors, e.g. A520, SVE2 is better than NEON with the 128bit vector width - For Out-of-order processors, NEON is still better than SVE2 (We could improve this in the future), the performance is improved from 3.0 to 3.6. * The 8888 -> RGB565 performance is also improved (from 7.4 to 9.3)
2026-05-28 07:45:22 +00:00 · 2026-05-26 22:57:44 +08:00
parent d5438360d2
commit f2dba2626e
4 changed files with 62 additions and 22 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -934,6 +934,10 @@ if(SDL_ASSEMBLY)
          return 0;
        }]==] COMPILER_SUPPORTS_ARMSVE2)
      if(COMPILER_SUPPORTS_ARMSVE2)
+        # IMPORTANT: As not all AArch64 processors support SVE2, we only
+        #            attach the following compilation option to SVE 
+        #            dedicated source files.
+        set(SVE2_MARCH_FLAG "-march=armv8-a+sve2")
        set(HAVE_ARMSVE2 TRUE)
      endif()
    cmake_pop_check_state()
@@ -947,6 +951,7 @@ if(SDL_ASSEMBLY)
        "${SDL3_SOURCE_DIR}/src/video/arm/SDL_sve2_blit_A.c"
        "${SDL3_SOURCE_DIR}/src/video/arm/SDL_sve2_blit_N.c"
        PROPERTIES
+          COMPILE_FLAGS "${SVE2_MARCH_FLAG}"
          SKIP_PRECOMPILE_HEADERS ON
      )
    endif()
--- a/include/SDL3/SDL_intrin.h
+++ b/include/SDL3/SDL_intrin.h
@@ -88,8 +88,12 @@
 /**
 * Defined if (and only if) the compiler supports ARM SVE2 intrinsics.
 *
- * If this macro is defined, SDL will have already included `<arm_sve.h>` as
- * appropriate.
+ * If this macro is defined, `<arm_sve.h>` (providing SVE intrinsics) will 
+ * only be included if the target architecture supports SVE 
+ * (`__ARM_FEATURE_SVE` feature macro).
+ * Some toolchains do not support `SDL_TARGETING("arch=armv8-a+sve2")`, so 
+ * for best portability you need to write all SVE code in a separate 
+ * translation unit and add appropriate compile flags.
 *
 * \since This macro is available since SDL 3.6.0.
 */
@@ -247,9 +251,11 @@ _m_prefetch(void *__P)
 #  define SDL_NEON_INTRINSICS 1
 #  include <arm_neon.h>
 #endif
-#if defined(__ARM_FEATURE_SVE2) && !defined(SDL_DISABLE_SVE2)
+#if !defined(SDL_DISABLE_SVE2)
 #  define SDL_SVE2_INTRINSICS 1
-#  include <arm_sve.h>
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
 #endif

 #else
@@ -284,16 +290,19 @@ _m_prefetch(void *__P)
 /* Visual Studio doesn't define __ARM_ARCH, but _M_ARM (if set, always 7), and _M_ARM64 (if set, always 1). */
 #    if defined (_M_ARM64) && 0 /* Please only remove this 0 when MSVC releasing support for SVE2 officially. */
 #      define SDL_SVE2_INTRINSICS 1
-#      include <arm_sve.h>
 #      define __ARM_FEATURE_SVE2 1 /* Set __ARM_FEATURE_SVE2 so that it can be used elsewhere, at compile time */
+#      define __ARM_FEATURE_SVE 1 /* Set __ARM_FEATURE_SVE so that it can be used elsewhere, at compile time */
 #      define __ARM_ARCH 8
+#      include <arm_sve.h>
 #    endif
 #  elif defined(SDL_PLATFORM_APPLE)
 /* Apple has no AArch64 device supporting SVE2 */
 #  elif defined(__ARM_ARCH) && (__ARM_ARCH >= 8) && (defined(__aarch64__) || defined(_M_ARM64)) && \
-        defined(__has_include) && __has_include(<arm_sve.h>) && defined(__ARM_FEATURE_SVE)
+        defined(__has_include) && __has_include(<arm_sve.h>)
 #    define SDL_SVE2_INTRINSICS 1
-#    include <arm_sve.h>
+#    if defined(__ARM_FEATURE_SVE)
+#      include <arm_sve.h>
+#    endif
 #  endif
 #endif
 #endif /* compiler version */
--- a/src/video/arm/SDL_sve2_extension.h
+++ b/src/video/arm/SDL_sve2_extension.h
@@ -19,10 +19,33 @@
  3. This notice may not be removed or altered from any source distribution.
 */

+/*
+ * IMPORTANT: Please do NOT include this header file directly or indirectly
+ *            outside the src/video/arm folder.
+ * 
+ */
+
 #if !defined(SDL_SVE2_EXTENSION_H) //&& (defined(__ARM_FEATURE_SVE2) && __ARM_FEATURE_SVE2)
 #define SDL_SVE2_EXTENSION_H

 #include "SDL_sve2_util.h"
+
+/*
+ * NOTE: Some Android builds didn't attach '-march=armv8-a+sve2' to 
+ *       SDL_sve2_*.c and hence the macro __ARM_FEATURE_SVE is not
+ *       defined by the compiler. This might not be a problem as the 
+ *       SDL_TARGETING("arch=armv8-a+sve2") enables the feature for
+ *       individual functions, until some version of compilers
+ *       provides arm_sve.h raising errors then __ARM_FEATURE_SVE 
+ *       is not defined. Although it should be avoided, as a 
+ *       workaround, we have to define the __ARM_FEATURE_SVE here as 
+ *       an ugly hack. 
+ */
+#ifdef SDL_PLATFORM_ANDROID
+#ifndef __ARM_FEATURE_SVE
+#define __ARM_FEATURE_SVE 1
+#endif
+#endif
 #include <arm_sve.h>
 #include <stdint.h>

@@ -907,7 +930,8 @@ static inline svuint16_t sdl_sve_chn_blend_with_mask(svuint16_t vSource,
                                                     svuint16_t vMask)
 {
    // vTarget = vSource * vMask + vTarget * (255 - vMask);
-    svuint16_t vTemp0 = svmul_u16_m(svptrue_b16(), vSource, vMask);
+    svuint16_t vTemp0 = svdup_u16(1);
+    vTemp0 = svmla_u16_m(svptrue_b16(), vTemp0, vSource, vMask);
    vTemp0 = svmla_u16_m(svptrue_b16(),
                         vTemp0,
                         vTarget,
@@ -915,17 +939,13 @@ static inline svuint16_t sdl_sve_chn_blend_with_mask(svuint16_t vSource,
                                     svdup_u16(255),
                                     vMask));

-    vTemp0 = svadd_n_u16_m(svptrue_b16(), vTemp0, 1);
-
-    svuint16_t vTemp1 = svlsr_n_u16_m(svptrue_b16(), vTemp0, 8);
    /* x += x >> 8 */
-    vTemp0 = svadd_u16_m(svptrue_b16(),
-                         vTemp0,
-                         vTemp1);
-
-    return svlsr_n_u16_m(svptrue_b16(), vTemp0, 8); // vTarget >> 8;
+    return svreinterpret_u16_u8(
+        svaddhnb_u16(vTemp0,
+                     svlsr_n_u16_m(svptrue_b16(),
+                                   vTemp0,
+                                   8)));
 }
-
 /*! \note the Element range of vMask is [0, 0xFF]
 */
 SDL_TARGETING("arch=armv8-a+sve2")
@@ -968,15 +988,15 @@ static inline svuint16_t sdl_sve_chn_blend_with_opacity(svuint16_t vSource,
 */
 SDL_TARGETING("arch=armv8-a+sve2")
 static inline svuint16_t sdl_sve_chn_blend_with_opacity_fast(svuint16_t vSource,
-                                                        svuint16_t vTarget,
-                                                        uint16_t hwOpacity)
+                                                             svuint16_t vTarget,
+                                                             uint16_t hwOpacity)
 {
    // vTarget = vSource * vMask + vTarget * (255 - vMask);
    svuint16_t vTemp0 = svmul_n_u16_m(svptrue_b16(), vSource, hwOpacity);
    vTemp0 = svmla_n_u16_m(svptrue_b16(),
-                         vTemp0,
-                         vTarget,
-                         256 - hwOpacity);
+                           vTemp0,
+                           vTarget,
+                           256 - hwOpacity);

    return svlsr_n_u16_m(svptrue_b16(), vTemp0, 8); // vTarget >> 8;
 }
--- a/src/video/arm/SDL_sve2_swizzle.h
+++ b/src/video/arm/SDL_sve2_swizzle.h
@@ -19,6 +19,12 @@
  3. This notice may not be removed or altered from any source distribution.
 */

+/*
+ * IMPORTANT: Please do NOT include this header file directly or indirectly
+ *            outside the src/video/arm folder.
+ * 
+ */
+
 #if !defined(SD_SVE2_SWIZZLE_H) //&& (defined(__ARM_FEATURE_SVE2) && __ARM_FEATURE_SVE2)
 #define SD_SVE2_SWIZZLE_H