Build only the SIMD code with SIMD compiler flags

glfw/glfw.py

@@ -6,7 +6,8 @@ import json
 import os
 import re
 import sys
-from typing import Callable, Dict, List, NamedTuple, Optional, Sequence, Tuple
+from enum import Enum
+from typing import Any, Callable, Dict, List, NamedTuple, Optional, Sequence, Tuple
 
 _plat = sys.platform.lower()
 is_linux = 'linux' in _plat
@@ -32,6 +33,19 @@ class Command(NamedTuple):
     keyfile: Optional[str] = None
 
 
+class ISA(Enum):
+    X86 = 0x03
+    AMD64 = 0x3e
+    ARM64 = 0xb7
+    Other = 0x0
+
+
+class BinaryArch(NamedTuple):
+    bits: int = 64
+    isa: ISA = ISA.AMD64
+
+
+
 class Env:
 
     cc: List[str] = []
@@ -42,6 +56,9 @@ class Env:
     ldpaths: List[str] = []
     ccver: Tuple[int, int]
     vcs_rev: str = ''
+    build_universal_binary: bool = False
+    binary_arch: BinaryArch = BinaryArch()
+    native_optimizations: bool = False
 
     # glfw stuff
     all_headers: List[str] = []
@@ -54,12 +71,16 @@ class Env:
     def __init__(
         self, cc: List[str] = [], cppflags: List[str] = [], cflags: List[str] = [], ldflags: List[str] = [],
         library_paths: Dict[str, List[str]] = {}, ldpaths: Optional[List[str]] = None, ccver: Tuple[int, int] = (0, 0),
-        vcs_rev: str = ''
+        vcs_rev: str = '', build_universal_binary: bool = False, binary_arch: BinaryArch = BinaryArch(),
+        native_optimizations: bool = False,
     ):
         self.cc, self.cppflags, self.cflags, self.ldflags, self.library_paths = cc, cppflags, cflags, ldflags, library_paths
         self.ldpaths = ldpaths or []
         self.ccver = ccver
         self.vcs_rev = vcs_rev
+        self.build_universal_binary = build_universal_binary
+        self.binary_arch = binary_arch
+        self.native_optimizations = native_optimizations
 
     def copy(self) -> 'Env':
         ans = Env(self.cc, list(self.cppflags), list(self.cflags), list(self.ldflags), dict(self.library_paths), list(self.ldpaths), self.ccver)
@@ -70,6 +91,9 @@ class Env:
         ans.wayland_scanner_code = self.wayland_scanner_code
         ans.wayland_protocols = self.wayland_protocols
         ans.vcs_rev = self.vcs_rev
+        ans.build_universal_binary = self.build_universal_binary
+        ans.binary_arch = self.binary_arch
+        ans.native_optimizations = self.native_optimizations
         return ans
 
 
@@ -83,7 +107,7 @@ def init_env(
     pkg_config: Callable[..., List[str]],
     pkg_version: Callable[[str], Tuple[int, int]],
     at_least_version: Callable[..., None],
-    test_compile: Callable[..., bool],
+    test_compile: Callable[..., Any],
     module: str = 'x11'
 ) -> Env:
     ans = env.copy()

kitty/arches.h

@@ -0,0 +1,29 @@
+/*
+ * Copyright (C) 2024 Kovid Goyal <kovid at kovidgoyal.net>
+ *
+ * Distributed under terms of the GPL3 license.
+ */
+
+#pragma once
+
+
+#ifdef __aarch64__
+#define KITTY_TARGET_CPU_IS_ARM64
+#define KITTY_128BIT_ALLOWED
+#define KITTY_256BIT_ALLOWED
+#elif defined(__i386__) || defined(__i486__) || defined(__i586__) || defined(__i686__)
+#define KITTY_TARGET_CPU_IS_X86
+#define KITTY_128BIT_ALLOWED
+#elif defined(__amd64__)
+#define KITTY_TARGET_CPU_IS_AMD64
+#define KITTY_128BIT_ALLOWED
+#define KITTY_256BIT_ALLOWED
+#endif
+
+#if defined(__clang__) && defined(KITTY_128BIT_ALLOWED)
+#define KITTY_START_128BIT_CODE
+#elif defined(KITTY_128BIT_ALLOWED)
+#define KITTY_START_128BIT_CODE
+#else
+#define KITTY_START_128BIT_CODE
+#endif

kitty/simd-string-128.c

@@ -0,0 +1,9 @@
+/*
+ * simd-string-128.c
+ * Copyright (C) 2024 Kovid Goyal <kovid at kovidgoyal.net>
+ *
+ * Distributed under terms of the GPL3 license.
+ */
+
+#define KITTY_SIMD_LEVEL 128
+#include "simd-string-impl.h"

kitty/simd-string-256.c

@@ -0,0 +1,9 @@
+/*
+ * simd-string-128.c
+ * Copyright (C) 2024 Kovid Goyal <kovid at kovidgoyal.net>
+ *
+ * Distributed under terms of the GPL3 license.
+ */
+
+#define KITTY_SIMD_LEVEL 256
+#include "simd-string-impl.h"

kitty/simd-string-impl.h

@@ -4,8 +4,8 @@
  * Distributed under terms of the GPL3 license.
  */
 
-#ifndef BITS
+#ifndef KITTY_SIMD_LEVEL
-#define BITS 128
+#define KITTY_SIMD_LEVEL 128
 #endif
 
 #include "simd-string.h"
@@ -26,13 +26,13 @@ _Pragma("clang diagnostic pop")
 #endif
 #define CONCAT(A, B) A##B
 #define CONCAT_EXPAND(A, B) CONCAT(A,B)
-#define FUNC(name) CONCAT_EXPAND(name##_, BITS)
+#define FUNC(name) CONCAT_EXPAND(name##_, KITTY_SIMD_LEVEL)
-#define integer_t CONCAT_EXPAND(CONCAT_EXPAND(simde__m, BITS), i)
+#define integer_t CONCAT_EXPAND(CONCAT_EXPAND(simde__m, KITTY_SIMD_LEVEL), i)
 #define shift_right_by_bytes128 simde_mm_srli_si128
 #define zero_last_n_bytes FUNC(zero_last_n_bytes)
 #define is_zero FUNC(is_zero)
 
-#if BITS == 128
+#if KITTY_SIMD_LEVEL == 128
 #define set1_epi8(x) simde_mm_set1_epi8((char)(x))
 #define set_epi8 simde_mm_set_epi8
 #define add_epi8 simde_mm_add_epi8
@@ -199,7 +199,7 @@ FUNC(zero_last_n_bytes)(integer_t vec, char n) {
     return andnot_si(mask, vec);
 }
 
-static inline const uint8_t*
+const uint8_t*
 FUNC(find_either_of_two_bytes)(const uint8_t *haystack, const size_t sz, const uint8_t a, const uint8_t b) {
     const integer_t a_vec = set1_epi8(a), b_vec = set1_epi8(b);
     for (const uint8_t* limit = haystack + sz; haystack < limit; haystack += sizeof(integer_t)) {
@@ -220,7 +220,7 @@ FUNC(find_either_of_two_bytes)(const uint8_t *haystack, const size_t sz, const u
 
 static inline void
 FUNC(output_plain_ascii)(UTF8Decoder *d, integer_t vec, size_t src_sz) {
-#if BITS == 128
+#if KITTY_SIMD_LEVEL == 128
     for (const uint32_t *limit = d->output + src_sz, *p = d->output; p < limit; p += output_increment) {
         const integer_t unpacked = extract_lower_quarter_as_chars(vec);
         store_aligned((integer_t*)p, unpacked);
@@ -252,7 +252,7 @@ FUNC(output_plain_ascii)(UTF8Decoder *d, integer_t vec, size_t src_sz) {
 
 static inline void
 FUNC(output_unicode)(UTF8Decoder *d, integer_t output1, integer_t output2, integer_t output3, const size_t num_codepoints) {
-#if BITS == 128
+#if KITTY_SIMD_LEVEL == 128
     for (const uint32_t *limit = d->output + num_codepoints, *p = d->output; p < limit; p += output_increment) {
         const integer_t unpacked1 = extract_lower_quarter_as_chars(output1);
         const integer_t unpacked2 = shift_right_by_one_byte(extract_lower_quarter_as_chars(output2));
@@ -294,7 +294,6 @@ FUNC(output_unicode)(UTF8Decoder *d, integer_t output1, integer_t output2, integ
 }
 #undef output_increment
 
-#ifndef SIMD_STRING_IMPL_INCLUDED_ONCE
 static inline unsigned
 sum_bytes_128(simde__m128i v) {
     // Use _mm_sad_epu8 to perform a sum of absolute differences against zero
@@ -345,9 +344,8 @@ scalar_decode_all(UTF8Decoder *d, const uint8_t *src, size_t src_sz) {
     return pos;
 }
 #undef do_one_byte
-#endif
 
-static inline bool
+bool
 FUNC(utf8_decode_to_esc)(UTF8Decoder *d, const uint8_t *src, size_t src_sz) {
     // Based on the algorithm described in: https://woboq.com/blog/utf-8-processing-using-simd.html
 
@@ -489,7 +487,7 @@ start_classification:
     shifts = add_epi8(shifts, shift_right_by_two_bytes(shifts));
     shifts = add_epi8(shifts, shift_right_by_four_bytes(shifts));
     shifts = add_epi8(shifts, shift_right_by_eight_bytes(shifts));
-#if BITS == 256
+#if KITTY_SIMD_LEVEL == 256
     shifts = add_epi8(shifts, shift_right_by_sixteen_bytes(shifts));
 #endif
     // zero the shifts for discarded continuation bytes
@@ -505,7 +503,7 @@ start_classification:
     shifts = move(shifts, two_bytes, 2);
     shifts = move(shifts, four_bytes, 3);
     shifts = move(shifts, eight_bytes, 4);
-#if BITS == 256
+#if KITTY_SIMD_LEVEL == 256
     shifts = move(shifts, sixteen_bytes, 5);
 #endif
 #undef move
@@ -547,7 +545,7 @@ invalid_utf8:
 #undef movemask_epi8
 #undef CONCAT
 #undef CONCAT_EXPAND
-#undef BITS
+#undef KITTY_SIMD_LEVEL
 #undef shift_right_by_one_byte
 #undef shift_right_by_two_bytes
 #undef shift_right_by_four_bytes
@@ -575,6 +573,3 @@ invalid_utf8:
 #undef sum_bytes
 #undef is_zero
 #undef print_register_as_bytes
-#ifndef SIMD_STRING_IMPL_INCLUDED_ONCE
-#define SIMD_STRING_IMPL_INCLUDED_ONCE
-#endif

kitty/simd-string.c

@@ -8,12 +8,6 @@
 #include "data-types.h"
 #include "charsets.h"
 #include "simd-string.h"
-#undef BITS
-#define BITS 128
-#include "simd-string-impl.h"
-#define BITS 256
-#include "simd-string-impl.h"
-#undef BITS
 static bool has_sse4_2 = false, has_avx2 = false;
 
 // find_either_of_two_bytes {{{

kitty/simd-string.h

@@ -6,11 +6,9 @@
 
 #pragma once
 
-#include <stdint.h>
-#include <stddef.h>
-#include <stdbool.h>
-#include <stdalign.h>
 #include "data-types.h"
+#include <stddef.h>
+#include <stdalign.h>
 
 typedef void (*control_byte_callback)(void *data, uint8_t ch);
 typedef void (*output_chars_callback)(void *data, const uint32_t *chars, unsigned count);
@@ -36,3 +34,9 @@ const uint8_t* find_either_of_two_bytes(const uint8_t *haystack, const size_t sz
 // first position in haystack that contains a char that is not in [a, b].
 // a must be <= b
 const uint8_t* find_byte_not_in_range(const uint8_t *haystack, const size_t sz, const uint8_t a1, const uint8_t b);
+
+// SIMD implementations, internal use
+bool utf8_decode_to_esc_128(UTF8Decoder *d, const uint8_t *src, size_t src_sz);
+bool utf8_decode_to_esc_256(UTF8Decoder *d, const uint8_t *src, size_t src_sz);
+const uint8_t* find_either_of_two_bytes_128(const uint8_t *haystack, const size_t sz, const uint8_t a, const uint8_t b);
+const uint8_t* find_either_of_two_bytes_256(const uint8_t *haystack, const size_t sz, const uint8_t a, const uint8_t b);

setup.py

@@ -10,6 +10,7 @@ import re
 import runpy
 import shlex
 import shutil
+import struct
 import subprocess
 import sys
 import sysconfig
@@ -22,7 +23,7 @@ from pathlib import Path
 from typing import Callable, Dict, FrozenSet, Iterable, Iterator, List, Optional, Sequence, Set, Tuple, Union, cast
 
 from glfw import glfw
-from glfw.glfw import Command, CompileKey
+from glfw.glfw import ISA, BinaryArch, Command, CompileKey
 
 if sys.version_info[:2] < (3, 8):
     raise SystemExit('kitty requires python >= 3.8')
@@ -362,6 +363,24 @@ def get_sanitize_args(cc: List[str], ccver: Tuple[int, int]) -> List[str]:
     return sanitize_args
 
 
+def get_binary_arch(path: str) -> BinaryArch:
+    with open(path, 'rb') as f:
+        sig = f.read(64)
+    if sig.startswith(b'\x7fELF'):  # ELF
+        bits = {1: 32, 2: 64}[sig[4]]
+        endian = {1: '<', 2: '>'}[sig[5]]
+        machine, = struct.unpack_from(endian + 'H', sig, 0x12)
+        isa = {i.value:i for i in ISA}.get(machine, ISA.Other)
+    elif sig[:4] in (b'\xcf\xfa\xed\xfe', b'\xce\xfa\xed\xfe'): # Mach-O
+        s, cpu_type, = struct.unpack_from('<II', sig, 0)
+        bits = {0xfeedface: 32, 0xfeedfacf: 64}[s]
+        cpu_type &= 0xff
+        isa = {0x7: ISA.AMD64, 0xc: ISA.ARM64}[cpu_type]
+    else:
+        raise SystemExit(f'Unknown binary format with signature: {sig[:4]!r}')
+    return BinaryArch(bits=bits, isa=isa)
+
+
 def test_compile(
     cc: List[str], *cflags: str,
     src: str = '',
@@ -370,18 +389,25 @@ def test_compile(
     show_stderr: bool = False,
     libraries: Iterable[str] = (),
     ldflags: Iterable[str] = (),
-) -> bool:
+    get_output_arch: bool = False,
+) -> Union[bool, BinaryArch]:
     src = src or 'int main(void) { return 0; }'
     with tempfile.TemporaryDirectory(prefix='kitty-test-compile-') as tdir:
         with open(os.path.join(tdir, f'source.{source_ext}'), 'w', encoding='utf-8') as srcf:
             print(src, file=srcf)
-        return subprocess.Popen(
+        output = os.path.join(tdir, 'source.output')
+        ret = subprocess.Popen(
             cc + ['-Werror=implicit-function-declaration'] + list(cflags) + ([] if link_also else ['-c']) +
-            ['-o', os.path.join(tdir, 'source.output'), srcf.name] +
+            ['-o', output, srcf.name] +
             [f'-l{x}' for x in libraries] + list(ldflags),
             stdout=subprocess.DEVNULL, stdin=subprocess.DEVNULL,
             stderr=None if show_stderr else subprocess.DEVNULL
-        ).wait() == 0
+        ).wait()
+        if get_output_arch:
+            if ret != 0:
+                raise SystemExit(f'Failed to determine target architecture compiling test program failed with exit code: {ret}')
+            return get_binary_arch(output)
+        return ret == 0
 
 
 def first_successful_compile(cc: List[str], *cflags: str, src: str = '', source_ext: str = 'c') -> str:
@@ -432,6 +458,7 @@ def init_env(
     vcs_rev: str = '',
 ) -> Env:
     native_optimizations = native_optimizations and not sanitize
+    build_universal_binary = build_universal_binary and is_macos
     cc, ccver = cc_version()
     if verbose:
         print('CC:', cc, ccver)
@@ -463,7 +490,7 @@ def init_env(
     # in https://github.com/kovidgoyal/kitty/issues/6845#issuecomment-1835886938
     arm_control_flow_protection = '-mbranch-protection=standard' if is_macos else ''
     # Universal build fails with -fcf-protection clang is not smart enough to filter it out for the ARM part
-    intel_control_flow_protection = '-fcf-protection=full' if ccver >= (9, 0) and not build_universal_binary else ''
+    intel_control_flow_protection = '-fcf-protection=full' if ccver >= (9, 0) else ''
     control_flow_protection = arm_control_flow_protection if is_arm else intel_control_flow_protection
     env_cflags = shlex.split(os.environ.get('CFLAGS', ''))
     env_cppflags = shlex.split(os.environ.get('CPPFLAGS', ''))
@@ -471,11 +498,10 @@ def init_env(
     if control_flow_protection and not test_compile(cc, control_flow_protection, *env_cppflags, *env_cflags, ldflags=env_ldflags):
         control_flow_protection = ''
     march = ''
-    if not (is_macos and is_arm) and not build_universal_binary:
+    if native_optimizations and not build_universal_binary and not (is_macos and is_arm):
         # see https://github.com/kovidgoyal/kitty/issues/3126
         # -march=native is not supported when targeting Apple Silicon
-        if native_optimizations:
+        march = '-march=native -mtune=native'
-            march = '-march=native -mtune=native'
     cflags_ = os.environ.get(
         'OVERRIDE_CFLAGS', (
             f'-Wextra {float_conversion} -Wno-missing-field-initializers -Wall -Wstrict-prototypes {std}'
@@ -508,13 +534,6 @@ def init_env(
         cflags.append('-g3')
         ldflags.append('-lprofiler')
 
-    # SIMD instructions
-    if is_arm:
-        if not is_macos:
-            cflags.append('-mfpu=neon')
-    else:
-        cflags.append('-msse4.2')
-        cflags.append('-mavx2')
     library_paths: Dict[str, List[str]] = {}
 
     def add_lpath(which: str, name: str, val: Optional[str]) -> None:
@@ -539,11 +558,13 @@ def init_env(
         cflags.insert(0, f'-I{os.environ["DEVELOP_ROOT"]}/include')
         ldpaths.insert(0, f'-L{os.environ["DEVELOP_ROOT"]}/lib')
 
-    if build_universal_binary:
+    ba = test_compile(cc, *(cppflags + cflags), ldflags=ldflags, get_output_arch=True)
-        set_arches(cflags)
+    assert isinstance(ba, BinaryArch)
-        set_arches(ldflags)
 
-    return Env(cc, cppflags, cflags, ldflags, library_paths, ccver=ccver, ldpaths=ldpaths, vcs_rev=vcs_rev)
+    return Env(
+        cc, cppflags, cflags, ldflags, library_paths, binary_arch=ba, native_optimizations=native_optimizations,
+        ccver=ccver, ldpaths=ldpaths, vcs_rev=vcs_rev, build_universal_binary=build_universal_binary
+    )
 
 
 def kitty_env(args: Options) -> Env:
@@ -638,7 +659,7 @@ def get_vcs_rev() -> str:
 
 
 @lru_cache
-def base64_defines() -> List[str]:
+def base64_defines(isa: ISA) -> List[str]:
     defs = {
         'HAVE_AVX512': 0,
         'HAVE_AVX2': 0,
@@ -649,14 +670,18 @@ def base64_defines() -> List[str]:
         'HAVE_SSE42': 0,
         'HAVE_AVX': 0,
     }
-    if is_arm:
+    if isa == ISA.ARM64:
         defs['HAVE_NEON64'] = 1
-    else:
+    elif isa == ISA.AMD64:
         defs['HAVE_AVX2'] = 1
         defs['HAVE_AVX'] = 1
         defs['HAVE_SSE42'] = 1
         defs['HAVE_SSE41'] = 1
         defs['HAVE_SSE3'] = 1
+    elif isa == ISA.X86:
+        defs['HAVE_SSE42'] = 1
+        defs['HAVE_SSE41'] = 1
+        defs['HAVE_SSE3'] = 1
     return [f'{k}={v}' for k, v in defs.items()]
 
 
@@ -668,13 +693,40 @@ def get_source_specific_defines(env: Env, src: str) -> Tuple[str, List[str], Opt
             env.vcs_rev = get_vcs_rev()
         return src, [], [f'KITTY_VCS_REV="{env.vcs_rev}"', f'WRAPPED_KITTENS="{wrapped_kittens()}"']
     if src.startswith('3rdparty/base64/'):
-        return src, ['3rdparty/base64',], base64_defines()
+        return src, ['3rdparty/base64',], base64_defines(env.binary_arch.isa)
     try:
         return src, [], env.library_paths[src]
     except KeyError:
         return src, [], None
 
 
+def get_source_specific_cflags(env: Env, src: str) -> List[str]:
+    ans = list(env.cflags)
+    # SIMD specific flags, ignored for native optimizations as they give slightly better performance
+    if src == 'kitty/simd-string-128.c':
+        if env.binary_arch.isa in (ISA.AMD64, ISA.X86):
+            if not env.native_optimizations:
+                ans.append('-msse4.2')
+    elif src == 'kitty/simd-string-256.c':
+        if env.binary_arch.isa in (ISA.AMD64, ISA.X86):
+            if not env.native_optimizations:
+                ans.append('-mavx2')
+    elif src.startswith('3rdparty/base64/lib/arch/'):
+        if not env.native_optimizations:
+            q = src.split(os.path.sep)
+            if 'sse3' in q:
+                ans.append('-msse3')
+            elif 'sse41' in q:
+                ans.append('-msse4.1')
+            elif 'sse42' in q:
+                ans.append('-msse4.2')
+            elif 'avx' in q:
+                ans.append('-mavx')
+            elif 'avx2' in q:
+                ans.append('-mavx2')
+    return ans
+
+
 def newer(dest: str, *sources: str) -> bool:
     try:
         dtime = os.path.getmtime(dest)
@@ -782,7 +834,8 @@ def compile_c_extension(
         src, include_paths, defines = get_source_specific_defines(kenv, src)
         if defines is not None:
             cppflags.extend(map(define, defines))
-        cmd = kenv.cc + ['-MMD'] + cppflags + [f'-I{x}' for x in include_paths] + kenv.cflags
+        cflags = get_source_specific_cflags(kenv, src)
+        cmd = kenv.cc + ['-MMD'] + cppflags + [f'-I{x}' for x in include_paths] + cflags
         cmd += ['-c', src] + ['-o', dest]
         key = CompileKey(original_src, os.path.basename(dest))
         desc = f'Compiling {emphasis(desc_prefix + src)} ...'