From e0649632ff53c883b684fd7f4f7591174695d301 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sat, 6 Jul 2024 15:30:14 +0400 Subject: [PATCH 01/11] Add SIMD headers --- src/libImaging/ImPlatform.h | 2 ++ src/libImaging/ImagingSIMD.h | 41 ++++++++++++++++++++++++++++++++++++ 2 files changed, 43 insertions(+) create mode 100644 src/libImaging/ImagingSIMD.h diff --git a/src/libImaging/ImPlatform.h b/src/libImaging/ImPlatform.h index f6e7fb6b921..e5fb4595e39 100644 --- a/src/libImaging/ImPlatform.h +++ b/src/libImaging/ImPlatform.h @@ -96,3 +96,5 @@ typedef signed __int64 int64_t; #ifdef __GNUC__ #define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) #endif + +#include "ImagingSIMD.h" diff --git a/src/libImaging/ImagingSIMD.h b/src/libImaging/ImagingSIMD.h new file mode 100644 index 00000000000..3d80c79d8ec --- /dev/null +++ b/src/libImaging/ImagingSIMD.h @@ -0,0 +1,41 @@ +/* Microsoft compiler doesn't limit intrinsics for an architecture. + This macro is set only on x86 and means SSE2 and above including AVX2. */ +#if defined(_M_X64) || _M_IX86_FP == 2 + #define __SSE2__ +#endif + +#ifdef __SSE4_2__ + #define __SSE4__ +#endif + +#ifdef __SSE2__ + #include // MMX + #include // SSE + #include // SSE2 +#endif +#ifdef __SSE4__ + #include // SSE3 + #include // SSSE3 + #include // SSE4.1 + #include // SSE4.2 +#endif +#ifdef __AVX2__ + #include // AVX, AVX2 +#endif +#ifdef __aarch64__ + #include // ARM NEON +#endif + +#ifdef __SSE4__ +static __m128i inline +mm_cvtepu8_epi32(void *ptr) { + return _mm_cvtepu8_epi32(_mm_cvtsi32_si128(*(INT32 *) ptr)); +} +#endif + +#ifdef __AVX2__ +static __m256i inline +mm256_cvtepu8_epi32(void *ptr) { + return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i *) ptr)); +} +#endif From 73f76e55c36d13f16802d07e86c4906be91c5fe6 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sat, 6 Jul 2024 15:30:22 +0400 Subject: [PATCH 02/11] Add SIMD example --- src/libImaging/Bands.c | 41 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 41 insertions(+) diff --git a/src/libImaging/Bands.c b/src/libImaging/Bands.c index e1b16b34ac0..1cfeb04bf79 100644 --- a/src/libImaging/Bands.c +++ b/src/libImaging/Bands.c @@ -21,6 +21,9 @@ Imaging ImagingGetBand(Imaging imIn, int band) { Imaging imOut; int x, y; +#ifdef __SSE4__ + __m128i shuffle_mask; +#endif /* Check arguments */ if (!imIn || imIn->type != IMAGING_TYPE_UINT8) { @@ -46,14 +49,25 @@ ImagingGetBand(Imaging imIn, int band) { return NULL; } +#ifdef __SSE4__ + shuffle_mask = _mm_set_epi8( + -1,-1,-1,-1, -1,-1,-1,-1, -1,-1,-1,-1, 12+band,8+band,4+band,0+band); +#endif + /* Extract band from image */ for (y = 0; y < imIn->ysize; y++) { UINT8 *in = (UINT8 *)imIn->image[y] + band; UINT8 *out = imOut->image8[y]; x = 0; for (; x < imIn->xsize - 3; x += 4) { +#ifdef __SSE4__ + __m128i source = _mm_loadu_si128((__m128i *)(in - band)); + *((UINT32 *)(out + x)) = _mm_cvtsi128_si32( + _mm_shuffle_epi8(source, shuffle_mask)); +#else UINT32 v = MAKE_UINT32(in[0], in[4], in[8], in[12]); memcpy(out + x, &v, sizeof(v)); +#endif in += 16; } for (; x < imIn->xsize; x++) { @@ -99,10 +113,18 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { UINT8 *out1 = bands[1]->image8[y]; x = 0; for (; x < imIn->xsize - 3; x += 4) { +#ifdef __SSE4__ + __m128i source = _mm_loadu_si128((__m128i *)in); + source = _mm_shuffle_epi8(source, _mm_set_epi8( + 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); + *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 12)); +#else UINT32 v = MAKE_UINT32(in[0], in[4], in[8], in[12]); memcpy(out0 + x, &v, sizeof(v)); v = MAKE_UINT32(in[0 + 3], in[4 + 3], in[8 + 3], in[12 + 3]); memcpy(out1 + x, &v, sizeof(v)); +#endif in += 16; } for (; x < imIn->xsize; x++) { @@ -119,12 +141,21 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { UINT8 *out2 = bands[2]->image8[y]; x = 0; for (; x < imIn->xsize - 3; x += 4) { +#ifdef __SSE4__ + __m128i source = _mm_loadu_si128((__m128i *)in); + source = _mm_shuffle_epi8(source, _mm_set_epi8( + 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); + *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); + *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); +#else UINT32 v = MAKE_UINT32(in[0], in[4], in[8], in[12]); memcpy(out0 + x, &v, sizeof(v)); v = MAKE_UINT32(in[0 + 1], in[4 + 1], in[8 + 1], in[12 + 1]); memcpy(out1 + x, &v, sizeof(v)); v = MAKE_UINT32(in[0 + 2], in[4 + 2], in[8 + 2], in[12 + 2]); memcpy(out2 + x, &v, sizeof(v)); +#endif in += 16; } for (; x < imIn->xsize; x++) { @@ -143,6 +174,15 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { UINT8 *out3 = bands[3]->image8[y]; x = 0; for (; x < imIn->xsize - 3; x += 4) { +#ifdef __SSE4__ + __m128i source = _mm_loadu_si128((__m128i *)in); + source = _mm_shuffle_epi8(source, _mm_set_epi8( + 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); + *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); + *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); + *((UINT32 *)(out3 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 12)); +#else UINT32 v = MAKE_UINT32(in[0], in[4], in[8], in[12]); memcpy(out0 + x, &v, sizeof(v)); v = MAKE_UINT32(in[0 + 1], in[4 + 1], in[8 + 1], in[12 + 1]); @@ -151,6 +191,7 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { memcpy(out2 + x, &v, sizeof(v)); v = MAKE_UINT32(in[0 + 3], in[4 + 3], in[8 + 3], in[12 + 3]); memcpy(out3 + x, &v, sizeof(v)); +#endif in += 16; } for (; x < imIn->xsize; x++) { From 910911061fda520ba3a9d346cc83d9fa0bfea2b4 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sat, 6 Jul 2024 15:40:30 +0400 Subject: [PATCH 03/11] Add __SSE4_2__ for MSCC --- src/libImaging/ImagingSIMD.h | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/libImaging/ImagingSIMD.h b/src/libImaging/ImagingSIMD.h index 3d80c79d8ec..ed259cebea2 100644 --- a/src/libImaging/ImagingSIMD.h +++ b/src/libImaging/ImagingSIMD.h @@ -2,8 +2,13 @@ This macro is set only on x86 and means SSE2 and above including AVX2. */ #if defined(_M_X64) || _M_IX86_FP == 2 #define __SSE2__ + /* However, Microsoft compiler set __AVX2__ if /arch:AVX2 option is set */ + #ifdef __AVX2__ + #define __SSE4_2__ + #endif #endif +/* For better readability */ #ifdef __SSE4_2__ #define __SSE4__ #endif From 2db9cd34991a06ec01943a5a78c7dba9e7d67d9b Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sun, 7 Jul 2024 17:13:44 +0400 Subject: [PATCH 04/11] Add core.acceleration attribute --- src/PIL/features.py | 4 +++- src/_imaging.c | 13 +++++++++++++ 2 files changed, 16 insertions(+), 1 deletion(-) diff --git a/src/PIL/features.py b/src/PIL/features.py index 13908c4eb78..9f4735aee06 100644 --- a/src/PIL/features.py +++ b/src/PIL/features.py @@ -128,6 +128,7 @@ def get_supported_codecs() -> list[str]: "libjpeg_turbo": ("PIL._imaging", "HAVE_LIBJPEGTURBO", "libjpeg_turbo_version"), "libimagequant": ("PIL._imaging", "HAVE_LIBIMAGEQUANT", "imagequant_version"), "xcb": ("PIL._imaging", "HAVE_XCB", None), + "acceleration": ("PIL._imaging", "acceleration", "acceleration"), } @@ -267,6 +268,7 @@ def pilinfo(out: IO[str] | None = None, supported_formats: bool = True) -> None: for name, feature in [ ("pil", "PIL CORE"), + ("acceleration", "Acceleration"), ("tkinter", "TKINTER"), ("freetype2", "FREETYPE2"), ("littlecms2", "LITTLECMS2"), @@ -291,7 +293,7 @@ def pilinfo(out: IO[str] | None = None, supported_formats: bool = True) -> None: if v is None: v = version(name) if v is not None: - version_static = name in ("pil", "jpg") + version_static = name in ("pil", "jpg", "acceleration") if name == "littlecms2": # this check is also in src/_imagingcms.c:setup_module() version_static = tuple(int(x) for x in v.split(".")) < (2, 7) diff --git a/src/_imaging.c b/src/_imaging.c index 07d9a64ccff..84c1d55c39d 100644 --- a/src/_imaging.c +++ b/src/_imaging.c @@ -4407,6 +4407,19 @@ setup_module(PyObject *m) { Py_INCREF(have_xcb); PyModule_AddObject(m, "HAVE_XCB", have_xcb); +#ifdef __AVX2__ + PyModule_AddStringConstant(m, "acceleration", "avx2"); +#elif defined(__SSE4__) + PyModule_AddStringConstant(m, "acceleration", "sse4"); +#elif defined(__SSE2__) + PyModule_AddStringConstant(m, "acceleration", "sse2"); +#elif defined(__NEON__) + PyModule_AddStringConstant(m, "acceleration", "neon"); +#else + Py_INCREF(Py_False); + PyModule_AddObject(m, "acceleration", Py_False); +#endif + PyObject *pillow_version = PyUnicode_FromString(version); PyDict_SetItemString( d, "PILLOW_VERSION", pillow_version ? pillow_version : Py_None From 5df34a24f88e9e68f18011ddeb6cdca2f435612b Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sun, 7 Jul 2024 19:09:35 +0400 Subject: [PATCH 05/11] clang-format --- src/libImaging/Bands.c | 36 +++++++++++++++++++++++++--------- src/libImaging/ImagingSIMD.h | 38 ++++++++++++++++++------------------ 2 files changed, 46 insertions(+), 28 deletions(-) diff --git a/src/libImaging/Bands.c b/src/libImaging/Bands.c index 1cfeb04bf79..6b3f8c03af3 100644 --- a/src/libImaging/Bands.c +++ b/src/libImaging/Bands.c @@ -51,7 +51,22 @@ ImagingGetBand(Imaging imIn, int band) { #ifdef __SSE4__ shuffle_mask = _mm_set_epi8( - -1,-1,-1,-1, -1,-1,-1,-1, -1,-1,-1,-1, 12+band,8+band,4+band,0+band); + -1, + -1, + -1, + -1, + -1, + -1, + -1, + -1, + -1, + -1, + -1, + -1, + 12 + band, + 8 + band, + 4 + band, + 0 + band); #endif /* Extract band from image */ @@ -62,8 +77,8 @@ ImagingGetBand(Imaging imIn, int band) { for (; x < imIn->xsize - 3; x += 4) { #ifdef __SSE4__ __m128i source = _mm_loadu_si128((__m128i *)(in - band)); - *((UINT32 *)(out + x)) = _mm_cvtsi128_si32( - _mm_shuffle_epi8(source, shuffle_mask)); + *((UINT32 *)(out + x)) = + _mm_cvtsi128_si32(_mm_shuffle_epi8(source, shuffle_mask)); #else UINT32 v = MAKE_UINT32(in[0], in[4], in[8], in[12]); memcpy(out + x, &v, sizeof(v)); @@ -115,8 +130,9 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { for (; x < imIn->xsize - 3; x += 4) { #ifdef __SSE4__ __m128i source = _mm_loadu_si128((__m128i *)in); - source = _mm_shuffle_epi8(source, _mm_set_epi8( - 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + source = _mm_shuffle_epi8( + source, + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 12)); #else @@ -143,8 +159,9 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { for (; x < imIn->xsize - 3; x += 4) { #ifdef __SSE4__ __m128i source = _mm_loadu_si128((__m128i *)in); - source = _mm_shuffle_epi8(source, _mm_set_epi8( - 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + source = _mm_shuffle_epi8( + source, + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); @@ -176,8 +193,9 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { for (; x < imIn->xsize - 3; x += 4) { #ifdef __SSE4__ __m128i source = _mm_loadu_si128((__m128i *)in); - source = _mm_shuffle_epi8(source, _mm_set_epi8( - 15,11,7,3, 14,10,6,2, 13,9,5,1, 12,8,4,0)); + source = _mm_shuffle_epi8( + source, + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); diff --git a/src/libImaging/ImagingSIMD.h b/src/libImaging/ImagingSIMD.h index ed259cebea2..a63f3b0aa02 100644 --- a/src/libImaging/ImagingSIMD.h +++ b/src/libImaging/ImagingSIMD.h @@ -1,46 +1,46 @@ /* Microsoft compiler doesn't limit intrinsics for an architecture. This macro is set only on x86 and means SSE2 and above including AVX2. */ #if defined(_M_X64) || _M_IX86_FP == 2 - #define __SSE2__ - /* However, Microsoft compiler set __AVX2__ if /arch:AVX2 option is set */ - #ifdef __AVX2__ - #define __SSE4_2__ - #endif +#define __SSE2__ +/* However, Microsoft compiler set __AVX2__ if /arch:AVX2 option is set */ +#ifdef __AVX2__ +#define __SSE4_2__ +#endif #endif /* For better readability */ #ifdef __SSE4_2__ - #define __SSE4__ +#define __SSE4__ #endif #ifdef __SSE2__ - #include // MMX - #include // SSE - #include // SSE2 +#include // MMX +#include // SSE +#include // SSE2 #endif #ifdef __SSE4__ - #include // SSE3 - #include // SSSE3 - #include // SSE4.1 - #include // SSE4.2 +#include // SSE3 +#include // SSSE3 +#include // SSE4.1 +#include // SSE4.2 #endif #ifdef __AVX2__ - #include // AVX, AVX2 +#include // AVX, AVX2 #endif #ifdef __aarch64__ - #include // ARM NEON +#include // ARM NEON #endif #ifdef __SSE4__ -static __m128i inline +static inline __m128i mm_cvtepu8_epi32(void *ptr) { - return _mm_cvtepu8_epi32(_mm_cvtsi32_si128(*(INT32 *) ptr)); + return _mm_cvtepu8_epi32(_mm_cvtsi32_si128(*(INT32 *)ptr)); } #endif #ifdef __AVX2__ -static __m256i inline +static inline __m256i mm256_cvtepu8_epi32(void *ptr) { - return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i *) ptr)); + return _mm256_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)ptr)); } #endif From f891043101f8f890776b079cf42b1b1411047575 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Mon, 8 Jul 2024 11:13:21 +0400 Subject: [PATCH 06/11] Define better name __NEON__ --- src/libImaging/ImagingSIMD.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/src/libImaging/ImagingSIMD.h b/src/libImaging/ImagingSIMD.h index a63f3b0aa02..fbf612868a4 100644 --- a/src/libImaging/ImagingSIMD.h +++ b/src/libImaging/ImagingSIMD.h @@ -12,6 +12,9 @@ #ifdef __SSE4_2__ #define __SSE4__ #endif +#ifdef __aarch64__ +#define __NEON__ +#endif #ifdef __SSE2__ #include // MMX @@ -27,7 +30,7 @@ #ifdef __AVX2__ #include // AVX, AVX2 #endif -#ifdef __aarch64__ +#ifdef __NEON__ #include // ARM NEON #endif From 4d10b3c0ab9e5b1f8173ecf89896fdee98417dd3 Mon Sep 17 00:00:00 2001 From: Andrew Murray Date: Mon, 8 Jul 2024 18:32:03 +1000 Subject: [PATCH 07/11] Updated test --- Tests/test_features.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Tests/test_features.py b/Tests/test_features.py index b7eefa09ae1..dbe53ee298c 100644 --- a/Tests/test_features.py +++ b/Tests/test_features.py @@ -36,7 +36,9 @@ def test(name: str, function: Callable[[str], str | None]) -> None: assert version is None else: assert function(name) == version - if name != "PIL": + if name == "acceleration": + assert version in ("avx2", "sse4", "sse2", "neon", None) + elif name != "PIL": if name == "zlib" and version is not None: version = re.sub(".zlib-ng$", "", version) elif name == "libtiff" and version is not None: From a410dcbe9a9c0db5d2e254302876e8e9b0af3a40 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Tue, 16 Jul 2024 22:45:27 +0400 Subject: [PATCH 08/11] Add accelerated test builds --- .github/workflows/test.yml | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml index 6e63333b0da..9e6fcab32e1 100644 --- a/.github/workflows/test.yml +++ b/.github/workflows/test.yml @@ -52,15 +52,19 @@ jobs: include: - { python-version: "3.11", PYTHONOPTIMIZE: 1, REVERSE: "--reverse" } - { python-version: "3.10", PYTHONOPTIMIZE: 2 } + # SIMD-accelerated builds for x86 + - { os: "ubuntu-latest", python-version: "3.9", acceleration: "sse4"} + - { os: "ubuntu-latest", python-version: "3.12", acceleration: "avx2"} # Free-threaded - { os: "ubuntu-latest", python-version: "3.13-dev", disable-gil: true } # M1 only available for 3.10+ - { os: "macos-13", python-version: "3.9" } + - { os: "macos-13", python-version: "3.9", acceleration: "avx2"} exclude: - { os: "macos-14", python-version: "3.9" } runs-on: ${{ matrix.os }} - name: ${{ matrix.os }} Python ${{ matrix.python-version }} ${{ matrix.disable-gil && 'free-threaded' || '' }} + name: ${{ matrix.os }} Python ${{ matrix.python-version }} ${{ matrix.acceleration }} ${{ matrix.disable-gil && 'free-threaded' || '' }} steps: - uses: actions/checkout@v4 @@ -108,7 +112,7 @@ jobs: GHA_LIBIMAGEQUANT_CACHE_HIT: ${{ steps.cache-libimagequant.outputs.cache-hit }} - name: Install macOS dependencies - if: startsWith(matrix.os, 'macOS') + if: startsWith(matrix.os, 'macos') run: | .github/workflows/macos-install.sh env: @@ -118,6 +122,11 @@ jobs: if: "matrix.os == 'ubuntu-latest' && matrix.python-version == '3.12'" run: echo "::add-matcher::.github/problem-matchers/gcc.json" + - name: Set compiler options for optimization + if: ${{ matrix.acceleration }} + run: | + echo "CC=cc -m${{ matrix.acceleration }}" >> $GITHUB_ENV + - name: Build run: | .ci/build.sh From cb55ee64b50327721b630fd6be98e1d170b67083 Mon Sep 17 00:00:00 2001 From: Aleksandr Karpinskii Date: Sun, 11 Aug 2024 21:11:46 +0400 Subject: [PATCH 09/11] update clang formatting --- src/libImaging/Bands.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/src/libImaging/Bands.c b/src/libImaging/Bands.c index 6b3f8c03af3..0ab3156edc1 100644 --- a/src/libImaging/Bands.c +++ b/src/libImaging/Bands.c @@ -66,7 +66,8 @@ ImagingGetBand(Imaging imIn, int band) { 12 + band, 8 + band, 4 + band, - 0 + band); + 0 + band + ); #endif /* Extract band from image */ @@ -132,7 +133,8 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { __m128i source = _mm_loadu_si128((__m128i *)in); source = _mm_shuffle_epi8( source, - _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0) + ); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 12)); #else @@ -161,7 +163,8 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { __m128i source = _mm_loadu_si128((__m128i *)in); source = _mm_shuffle_epi8( source, - _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0) + ); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); @@ -195,7 +198,8 @@ ImagingSplit(Imaging imIn, Imaging bands[4]) { __m128i source = _mm_loadu_si128((__m128i *)in); source = _mm_shuffle_epi8( source, - _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0)); + _mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0) + ); *((UINT32 *)(out0 + x)) = _mm_cvtsi128_si32(source); *((UINT32 *)(out1 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 4)); *((UINT32 *)(out2 + x)) = _mm_cvtsi128_si32(_mm_srli_si128(source, 8)); From 24c0569f8d4e2ec105c721116a2e7c47af35c032 Mon Sep 17 00:00:00 2001 From: Andrew Murray Date: Mon, 12 Aug 2024 18:52:38 +1000 Subject: [PATCH 10/11] Do not return string from check_feature() --- src/PIL/features.py | 2 +- src/_imaging.c | 5 +++++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/src/PIL/features.py b/src/PIL/features.py index 9f4735aee06..ac7dbcb7fac 100644 --- a/src/PIL/features.py +++ b/src/PIL/features.py @@ -128,7 +128,7 @@ def get_supported_codecs() -> list[str]: "libjpeg_turbo": ("PIL._imaging", "HAVE_LIBJPEGTURBO", "libjpeg_turbo_version"), "libimagequant": ("PIL._imaging", "HAVE_LIBIMAGEQUANT", "imagequant_version"), "xcb": ("PIL._imaging", "HAVE_XCB", None), - "acceleration": ("PIL._imaging", "acceleration", "acceleration"), + "acceleration": ("PIL._imaging", "HAVE_ACCELERATION", "acceleration"), } diff --git a/src/_imaging.c b/src/_imaging.c index 84c1d55c39d..b639cd8a84e 100644 --- a/src/_imaging.c +++ b/src/_imaging.c @@ -4407,6 +4407,7 @@ setup_module(PyObject *m) { Py_INCREF(have_xcb); PyModule_AddObject(m, "HAVE_XCB", have_xcb); + PyObject *have_acceleration = Py_True; #ifdef __AVX2__ PyModule_AddStringConstant(m, "acceleration", "avx2"); #elif defined(__SSE4__) @@ -4418,7 +4419,11 @@ setup_module(PyObject *m) { #else Py_INCREF(Py_False); PyModule_AddObject(m, "acceleration", Py_False); + + have_acceleration = Py_False; #endif + Py_INCREF(have_acceleration); + PyModule_AddObject(m, "HAVE_ACCELERATION", have_acceleration); PyObject *pillow_version = PyUnicode_FromString(version); PyDict_SetItemString( From 55a97f9ddfebfa27df3541ddb988c07ba5194574 Mon Sep 17 00:00:00 2001 From: Andrew Murray Date: Mon, 12 Aug 2024 18:27:16 +1000 Subject: [PATCH 11/11] Do not return bool from version_feature() --- src/_imaging.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/_imaging.c b/src/_imaging.c index b639cd8a84e..184f5fb98a6 100644 --- a/src/_imaging.c +++ b/src/_imaging.c @@ -4417,8 +4417,8 @@ setup_module(PyObject *m) { #elif defined(__NEON__) PyModule_AddStringConstant(m, "acceleration", "neon"); #else - Py_INCREF(Py_False); - PyModule_AddObject(m, "acceleration", Py_False); + Py_INCREF(Py_None); + PyModule_AddObject(m, "acceleration", Py_None); have_acceleration = Py_False; #endif