Merge pull request #2421 from pygame-community/simd-transform-setup

Add SIMD functionality to the transform submodule (Attempt 2)

buildconfig/Setup.Android.SDL2.in

@@ -57,7 +57,7 @@ time src_c/time.c $(SDL) $(DEBUG)
 joystick src_c/joystick.c $(SDL) $(DEBUG)
 draw src_c/draw.c $(SDL) $(DEBUG)
 image src_c/image.c $(SDL) $(DEBUG)
-transform src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG) -D_NO_MMX_FOR_X86_64
+transform src_c/simd_transform_sse2.c src_c/simd_transform_avx2.c src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG) -D_NO_MMX_FOR_X86_64
 mask src_c/mask.c src_c/bitmask.c $(SDL) $(DEBUG)
 bufferproxy src_c/bufferproxy.c $(SDL) $(DEBUG)
 pixelarray src_c/pixelarray.c $(SDL) $(DEBUG)

buildconfig/Setup.Emscripten.SDL2.in

@@ -69,7 +69,7 @@ _sdl2.controller_old src_c/void.c
 #_sdl2.touch src_c/_sdl2/touch.c $(SDL) $(DEBUG) -Isrc_c
 _sdl2.touch src_c/void.c
 
-#transform src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG) -D_NO_MMX_FOR_X86_64
+#transform src_c/simd_transform_sse2.c src_c/simd_transform_avx2.c src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG) -D_NO_MMX_FOR_X86_64
 transform src_c/void.c
 
 

buildconfig/Setup.SDL2.in

@@ -67,7 +67,7 @@ time src_c/time.c $(SDL) $(DEBUG)
 joystick src_c/joystick.c $(SDL) $(DEBUG)
 draw src_c/draw.c $(SDL) $(DEBUG)
 image src_c/image.c $(SDL) $(DEBUG)
-transform src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG)
+transform src_c/simd_transform_sse2.c src_c/simd_transform_avx2.c src_c/transform.c src_c/rotozoom.c src_c/scale2x.c src_c/scale_mmx.c $(SDL) $(DEBUG)
 mask src_c/mask.c src_c/bitmask.c $(SDL) $(DEBUG)
 bufferproxy src_c/bufferproxy.c $(SDL) $(DEBUG)
 pixelarray src_c/pixelarray.c $(SDL) $(DEBUG)

setup.py

@@ -74,7 +74,7 @@
 
 import distutils.ccompiler
 
-avx2_filenames = ['simd_blitters_avx2']
+avx2_filenames = ['simd_blitters_avx2', 'simd_transform_avx2']
 
 compiler_options = {
     'unix': ('-mavx2',),

src_c/alphablit.c

@@ -24,59 +24,7 @@
 
 #define NO_PYGAME_C_API
 #include "_surface.h"
-
-#if !defined(PG_ENABLE_ARM_NEON) && defined(__aarch64__)
-// arm64 has neon optimisations enabled by default, even when fpu=neon is not
-// passed
-#define PG_ENABLE_ARM_NEON 1
-#endif
-
-/* See if we are compiled 64 bit on GCC or MSVC */
-#if _WIN32 || _WIN64
-#if _WIN64
-#define ENV64BIT
-#endif
-#endif
-
-// Check GCC
-#if __GNUC__
-#if __x86_64__ || __ppc64__ || __aarch64__
-#define ENV64BIT
-#endif
-#endif
-
-#if PG_ENABLE_ARM_NEON
-// sse2neon.h is from here: https://github.com/DLTcollab/sse2neon
-#include "include/sse2neon.h"
-#endif /* PG_ENABLE_ARM_NEON */
-
-/* This defines PG_ENABLE_SSE_NEON as True if either SSE or NEON is available
- * at compile time. Since we do compile time translation of SSE2->NEON, they
- * have the same code paths, so this reduces code duplication of those paths.
- */
-#if defined(__SSE2__)
-#define PG_ENABLE_SSE_NEON 1
-#elif PG_ENABLE_ARM_NEON
-#define PG_ENABLE_SSE_NEON 1
-#else
-#define PG_ENABLE_SSE_NEON 0
-#endif
-
-/* This returns True if either SSE2 or NEON is present at runtime.
- * Relevant because they use the same codepaths. Only the relevant runtime
- * SDL cpu feature check is compiled in.*/
-int
-pg_HasSSE_NEON()
-{
-#if defined(__SSE2__)
-    return SDL_HasSSE2();
-#elif PG_ENABLE_ARM_NEON
-    return SDL_HasNEON();
-#else
-    return 0;
-#endif
-}
-
+#include "simd_shared.h"
 #include "simd_blitters.h"
 
 static void

src_c/simd_blitters.h

@@ -8,13 +8,6 @@
 #define PG_ENABLE_ARM_NEON 1
 #endif
 
-int
-pg_sse2_at_runtime_but_uncompiled();
-int
-pg_neon_at_runtime_but_uncompiled();
-int
-pg_avx2_at_runtime_but_uncompiled();
-
 #if (defined(__SSE2__) || defined(PG_ENABLE_ARM_NEON))
 void
 alphablit_alpha_sse2_argb_surf_alpha(SDL_BlitInfo *info);
@@ -60,8 +53,6 @@ premul_surf_color_by_alpha_non_simd(SDL_Surface *src, SDL_Surface *dst);
 void
 premul_surf_color_by_alpha_sse2(SDL_Surface *src, SDL_Surface *dst);
 
-int
-pg_has_avx2();
 void
 alphablit_alpha_avx2_argb_no_surf_alpha_opaque_dst(SDL_BlitInfo *info);
 void

src_c/simd_shared.h

@@ -0,0 +1,68 @@
+#define NO_PYGAME_C_API
+#ifndef SIMD_SHARED_H
+#define SIMD_SHARED_H
+
+#include "_surface.h"
+
+int
+pg_sse2_at_runtime_but_uncompiled();
+int
+pg_neon_at_runtime_but_uncompiled();
+int
+pg_avx2_at_runtime_but_uncompiled();
+int
+pg_has_avx2();
+
+#if !defined(PG_ENABLE_ARM_NEON) && defined(__aarch64__)
+// arm64 has neon optimisations enabled by default, even when fpu=neon is not
+// passed
+#define PG_ENABLE_ARM_NEON 1
+#endif
+
+/* See if we are compiled 64 bit on GCC or MSVC */
+#if _WIN32 || _WIN64
+#if _WIN64
+#define ENV64BIT
+#endif
+#endif
+
+// Check GCC
+#if __GNUC__
+#if __x86_64__ || __ppc64__ || __aarch64__
+#define ENV64BIT
+#endif
+#endif
+
+#if PG_ENABLE_ARM_NEON
+// sse2neon.h is from here: https://github.com/DLTcollab/sse2neon
+#include "include/sse2neon.h"
+#endif /* PG_ENABLE_ARM_NEON */
+
+/* This defines PG_ENABLE_SSE_NEON as True if either SSE or NEON is available
+ * at compile time. Since we do compile time translation of SSE2->NEON, they
+ * have the same code paths, so this reduces code duplication of those paths.
+ */
+#if defined(__SSE2__)
+#define PG_ENABLE_SSE_NEON 1
+#elif PG_ENABLE_ARM_NEON
+#define PG_ENABLE_SSE_NEON 1
+#else
+#define PG_ENABLE_SSE_NEON 0
+#endif
+
+/* This returns True if either SSE2 or NEON is present at runtime.
+ * Relevant because they use the same codepaths. Only the relevant runtime
+ * SDL cpu feature check is compiled in.*/
+int
+pg_HasSSE_NEON()
+{
+#if defined(__SSE2__)
+    return SDL_HasSSE2();
+#elif PG_ENABLE_ARM_NEON
+    return SDL_HasNEON();
+#else
+    return 0;
+#endif
+}
+
+#endif  // SIMD_SHARED_H

src_c/simd_transform.h

@@ -0,0 +1,14 @@
+#define NO_PYGAME_C_API
+#include "_surface.h"
+
+#if !defined(PG_ENABLE_ARM_NEON) && defined(__aarch64__)
+// arm64 has neon optimisations enabled by default, even when fpu=neon is not
+// passed
+#define PG_ENABLE_ARM_NEON 1
+#endif
+
+// SSE2 functions
+#if defined(__SSE2__) || defined(PG_ENABLE_ARM_NEON)
+#endif /* (defined(__SSE2__) || defined(PG_ENABLE_ARM_NEON)) */
+
+// AVX2 functions

src_c/simd_transform_avx2.c

@@ -0,0 +1,44 @@
+#include "simd_transform.h"
+
+#if defined(HAVE_IMMINTRIN_H) && !defined(SDL_DISABLE_IMMINTRIN_H)
+#include <immintrin.h>
+#endif /* defined(HAVE_IMMINTRIN_H) && !defined(SDL_DISABLE_IMMINTRIN_H) */
+
+#define BAD_AVX2_FUNCTION_CALL                                               \
+    printf(                                                                  \
+        "Fatal Error: Attempted calling an AVX2 function when both compile " \
+        "time and runtime support is missing. If you are seeing this "       \
+        "message, you have stumbled across a pygame bug, please report it "  \
+        "to the devs!");                                                     \
+    PG_EXIT(1)
+
+/* helper function that does a runtime check for AVX2. It has the added
+ * functionality of also returning 0 if compile time support is missing */
+int
+pg_has_avx2()
+{
+#if defined(__AVX2__) && defined(HAVE_IMMINTRIN_H) && \
+    !defined(SDL_DISABLE_IMMINTRIN_H)
+    return SDL_HasAVX2();
+#else
+    return 0;
+#endif /* defined(__AVX2__) && defined(HAVE_IMMINTRIN_H) && \
+          !defined(SDL_DISABLE_IMMINTRIN_H) */
+}
+
+/* This returns 1 when avx2 is available at runtime but support for it isn't
+ * compiled in, 0 in all other cases */
+int
+pg_avx2_at_runtime_but_uncompiled()
+{
+    if (SDL_HasAVX2()) {
+#if defined(__AVX2__) && defined(HAVE_IMMINTRIN_H) && \
+    !defined(SDL_DISABLE_IMMINTRIN_H)
+        return 0;
+#else
+        return 1;
+#endif /* defined(__AVX2__) && defined(HAVE_IMMINTRIN_H) && \
+          !defined(SDL_DISABLE_IMMINTRIN_H) */
+    }
+    return 0;
+}

src_c/simd_transform_sse2.c

@@ -0,0 +1,36 @@
+#include "simd_transform.h"
+
+#if PG_ENABLE_ARM_NEON
+// sse2neon.h is from here: https://github.com/DLTcollab/sse2neon
+#include "include/sse2neon.h"
+#endif /* PG_ENABLE_ARM_NEON */
+
+/* This returns 1 when sse2 is available at runtime but support for it isn't
+ * compiled in, 0 in all other cases */
+int
+pg_sse2_at_runtime_but_uncompiled()
+{
+    if (SDL_HasSSE2()) {
+#ifdef __SSE2__
+        return 0;
+#else
+        return 1;
+#endif /* __SSE2__ */
+    }
+    return 0;
+}
+
+/* This returns 1 when neon is available at runtime but support for it isn't
+ * compiled in, 0 in all other cases */
+int
+pg_neon_at_runtime_but_uncompiled()
+{
+    if (SDL_HasNEON()) {
+#if PG_ENABLE_ARM_NEON
+        return 0;
+#else
+        return 1;
+#endif /* PG_ENABLE_ARM_NEON */
+    }
+    return 0;
+}

src_c/transform.c

@@ -33,6 +33,8 @@
 #include <math.h>
 #include <string.h>
 
+#include "simd_shared.h"
+#include "simd_transform.h"
 #include "scale.h"
 
 typedef void (*SMOOTHSCALE_FILTER_P)(Uint8 *, Uint8 *, int, int, int, int,