shithub: openh264

--- a/codec/common/meson.build

+++ b/codec/common/meson.build

@@ -17,21 +17,41 @@

   'src/WelsThreadPool.cpp',

-asm_sources = [

-  'x86/cpuid.asm',

-  'x86/dct.asm',

-  'x86/deblock.asm',

-  'x86/expand_picture.asm',

-  'x86/intra_pred_com.asm',

-  'x86/mb_copy.asm',

-  'x86/mc_chroma.asm',

-  'x86/mc_luma.asm',

-  'x86/satd_sad.asm',

-  'x86/vaa.asm',

-]

+objs_asm = []

+if ['x86', 'x86_64'].contains(cpu_family)

+  asm_sources = [

+    'x86/cpuid.asm',

+    'x86/dct.asm',

+    'x86/deblock.asm',

+    'x86/expand_picture.asm',

+    'x86/intra_pred_com.asm',

+    'x86/mb_copy.asm',

+    'x86/mc_chroma.asm',

+    'x86/mc_luma.asm',

+    'x86/satd_sad.asm',

+    'x86/vaa.asm',

+  ]

+  objs_asm += asm_gen.process(asm_sources)

+elif cpu_family == 'arm'

+  cpp_sources += [

+    'arm/copy_mb_neon.S',

+    'arm/deblocking_neon.S',

+    'arm/expand_picture_neon.S',

+    'arm/intra_pred_common_neon.S',

+    'arm/mc_neon.S',

+  ]

+elif cpu_family == 'aarch64'

+  cpp_sources += [

+    'arm64/copy_mb_aarch64_neon.S',

+    'arm64/deblocking_aarch64_neon.S',

+    'arm64/expand_picture_aarch64_neon.S',

+    'arm64/intra_pred_common_aarch64_neon.S',

+    'arm64/mc_aarch64_neon.S',

+  ]

+else

+  error('Unsupported cpu_family @0@'.format(cpu_family))

+endif

-objs_asm = asm_gen.process(asm_sources)

 libcommon = static_library('common', cpp_sources, objs_asm,

-  include_directories: inc,

+  include_directories: [inc, casm_inc],

   dependencies: deps)

--- a/codec/decoder/meson.build

+++ b/codec/decoder/meson.build

@@ -22,13 +22,27 @@

   'core/src/wels_decoder_thread.cpp',

-asm_sources = [

-  'core/x86/dct.asm',

-  'core/x86/intra_pred.asm',

-]

+objs_asm = []

+if ['x86', 'x86_64'].contains(cpu_family)

+  asm_sources = [

+    'core/x86/dct.asm',

+    'core/x86/intra_pred.asm',

+  ]

+  objs_asm = asm_gen.process(asm_sources)

+elif cpu_family == 'arm'

+  cpp_sources += [

+    'core/arm/block_add_neon.S',

+    'core/arm/intra_pred_neon.S',

+  ]

+elif cpu_family == 'aarch64'

+  cpp_sources += [

+    'core/arm64/block_add_aarch64_neon.S',

+    'core/arm64/intra_pred_aarch64_neon.S',

+  ]

+else

+  error('Unsupported cpu family @0@'.format(cpu_family))

+endif

-objs_asm = asm_gen.process(asm_sources)

 libdecoder = static_library('decoder', cpp_sources, objs_asm,

-  include_directories: [inc, decoder_inc],

+  include_directories: [inc, decoder_inc, casm_inc],

   dependencies: deps)

--- a/codec/encoder/meson.build

+++ b/codec/encoder/meson.build

@@ -33,19 +33,41 @@

   'plus/src/welsEncoderExt.cpp',

-asm_sources = [

-  'core/x86/coeff.asm',

-  'core/x86/dct.asm',

-  'core/x86/intra_pred.asm',

-  'core/x86/matrix_transpose.asm',

-  'core/x86/memzero.asm',

-  'core/x86/quant.asm',

-  'core/x86/sample_sc.asm',

-  'core/x86/score.asm',

-]

+objs_asm = []

+if ['x86', 'x86_64'].contains(cpu_family)

+  asm_sources = [

+    'core/x86/coeff.asm',

+    'core/x86/dct.asm',

+    'core/x86/intra_pred.asm',

+    'core/x86/matrix_transpose.asm',

+    'core/x86/memzero.asm',

+    'core/x86/quant.asm',

+    'core/x86/sample_sc.asm',

+    'core/x86/score.asm',

+  ]

+  objs_asm = asm_gen.process(asm_sources)

+elif cpu_family == 'arm'

+  cpp_sources += [

+    'core/arm/intra_pred_neon.S',

+    'core/arm/intra_pred_sad_3_opt_neon.S',

+    'core/arm/memory_neon.S',

+    'core/arm/pixel_neon.S',

+    'core/arm/reconstruct_neon.S',

+    'core/arm/svc_motion_estimation.S',

+  ]

+elif cpu_family == 'aarch64'

+  cpp_sources += [

+    'core/arm64/intra_pred_aarch64_neon.S',

+    'core/arm64/intra_pred_sad_3_opt_aarch64_neon.S',

+    'core/arm64/memory_aarch64_neon.S',

+    'core/arm64/pixel_aarch64_neon.S',

+    'core/arm64/reconstruct_aarch64_neon.S',

+    'core/arm64/svc_motion_estimation_aarch64_neon.S',

+  ]

+else

+  error('Unsupported cpu family @0@'.format(cpu_family))

+endif

-objs_asm = asm_gen.process(asm_sources)

 libencoder = static_library('encoder', cpp_sources, objs_asm,

-  include_directories: [inc, processing_inc, encoder_inc],

+  include_directories: [inc, processing_inc, encoder_inc, casm_inc],

   dependencies: deps)

--- a/codec/meson.build

+++ b/codec/meson.build

@@ -2,5 +2,8 @@

 subdir('decoder')

 subdir('encoder')

 subdir('processing')

-subdir('console')

+if not ['android', 'ios'].contains(system)

+  # also disabled in the Makefile for these platforms

+  subdir('console')

+endif

 subdir('api')

--- a/codec/processing/meson.build

+++ b/codec/processing/meson.build

@@ -18,14 +18,32 @@

   'src/vaacalc/vaacalculation.cpp',

-asm_sources = [

-  'src/x86/denoisefilter.asm',

-  'src/x86/downsample_bilinear.asm',

-  'src/x86/vaa.asm',

-]

+objs_asm = []

+if ['x86', 'x86_64'].contains(cpu_family)

+  asm_sources = [

+    'src/x86/denoisefilter.asm',

+    'src/x86/downsample_bilinear.asm',

+    'src/x86/vaa.asm',

+  ]

+  objs_asm = asm_gen.process(asm_sources)

+elif cpu_family == 'arm'

+  cpp_sources += [

+    'src/arm/adaptive_quantization.S',

+    'src/arm/down_sample_neon.S',

+    'src/arm/pixel_sad_neon.S',

+    'src/arm/vaa_calc_neon.S',

+  ]

+elif cpu_family == 'aarch64'

+  cpp_sources += [

+	'src/arm64/adaptive_quantization_aarch64_neon.S',

+	'src/arm64/down_sample_aarch64_neon.S',

+	'src/arm64/pixel_sad_aarch64_neon.S',

+	'src/arm64/vaa_calc_aarch64_neon.S',

+  ]

+else

+  error('Unsupported cpu family @0@'.format(cpu_family))

+endif

-objs_asm = asm_gen.process(asm_sources)

 libprocessing = static_library('processing', cpp_sources, objs_asm,

-  include_directories: [inc, processing_inc],

+  include_directories: [inc, processing_inc, casm_inc],

   dependencies: deps)

--- a/meson.build

+++ b/meson.build

@@ -36,8 +36,6 @@

   join_paths('codec', 'encoder', 'plus', 'inc'),

])

-asm_inc = join_paths(meson.current_source_dir(), 'codec', 'common', 'x86', '')

 nasm = find_program('nasm', 'nasm.exe')

 system = host_machine.system()

@@ -54,24 +52,52 @@

 c_args = []

 cpp_args = []

 asm_args = []

+asm_inc = []

+casm_inc = []

+cpp_lib = '-lstdc++'

-if system == 'linux'

+# TODO: should rely on dependency('threads') instead and change the pkg-config

+# generator below

+pthread_dep = cpp.find_library('pthread', required : false)

+libm_dep = cpp.find_library('libm', required : false)

+deps += [libm_dep]

+if ['linux', 'android', 'ios', 'darwin'].contains(system)

+  asm_format32 = 'elf'

+  asm_format64 = 'elf64'

+  if ['ios', 'darwin'].contains(system)

+    asm_format32 = 'macho32'

+    asm_format64 = 'macho64'

+  endif

   if cpu_family == 'x86'

-    asm_format = 'elf'

-    asm_args += ['-DX86_32']

-    add_project_arguments('-DX86_32_ASM', language: 'c')

+    asm_format = asm_format32

+    asm_args += ['-DX86_32', '-DHAVE_AVX2']

+    add_project_arguments('-DHAVE_AVX2', language: 'cpp')

+    add_project_arguments('-DHAVE_AVX2', '-DX86_ASM', '-DX86_32_ASM', language: 'c')

+    asm_inc = join_paths(meson.current_source_dir(), 'codec', 'common', 'x86', '')

   elif cpu_family == 'x86_64'

-    asm_format = 'elf64'

-    asm_args += ['-DUNIX64']

+    asm_format = asm_format64

+    asm_args += ['-DUNIX64', '-DHAVE_AVX2']

+    add_project_arguments('-DHAVE_AVX2', language: 'cpp')

+    add_project_arguments('-DHAVE_AVX2', '-DX86_ASM', language: 'c')

+    asm_inc = join_paths(meson.current_source_dir(), 'codec', 'common', 'x86', '')

+  elif cpu_family == 'arm'

+    asm_format = asm_format32

+    add_project_arguments('-DHAVE_NEON', language: 'c')

+    add_project_arguments('-DHAVE_NEON', language: 'c')

+    casm_inc = include_directories(join_paths('codec', 'common', 'arm'))

+  elif cpu_family == 'aarch64'

+    asm_format = asm_format64

+    add_project_arguments('-DHAVE_NEON_ARM64', language: 'c')

+    add_project_arguments('-DHAVE_NEON_ARM64', language: 'cpp')

+    casm_inc = include_directories(join_paths('codec', 'common', 'arm64'))

   else

-    error ('FIXME: unhandled CPU family @0@ for Linux'.format(cpu_family))

+    error ('FIXME: unhandled CPU family @0@ for @1@'.format(cpu_family, system))

   endif

-  deps += [cpp.find_library('libm')]

-  asm_args += ['-DHAVE_AVX2']

-  add_project_arguments('-DHAVE_AVX2', language: 'cpp')

-  add_project_arguments('-DHAVE_AVX2', '-DX86_ASM', language: 'c')

+  if ['ios', 'darwin', 'android'].contains(system)

+    cpp_lib = '-lc++'

+  endif

 elif system == 'windows'

   if cpu_family == 'x86'

     asm_format = 'win32'

@@ -82,17 +108,20 @@

   else

     error ('FIXME: unhandled CPU family @0@ for Windows'.format(cpu_family))

   endif

+  asm_inc = join_paths(meson.current_source_dir(), 'codec', 'common', 'x86', '')

 else

   error ('FIXME: Unhandled system @0@'.format(system))

 endif

-asm_gen = generator(nasm,

-  output : '@BASENAME@.o',

-  arguments : [

-    '-f', asm_format,

-    '-i', asm_inc,

-    '@INPUT@',

-    '-o', '@OUTPUT@'] + asm_args)

+if ['x86', 'x86_64'].contains(cpu_family)

+  asm_gen = generator(nasm,

+    output : '@BASENAME@.o',

+    arguments : [

+      '-f', asm_format,

+      '-i', asm_inc,

+      '@INPUT@',

+      '-o', '@OUTPUT@'] + asm_args)

+endif

 api_headers = []

 api_header_deps = []

@@ -127,17 +156,20 @@

   pkgconf.set('prefix', join_paths(get_option('prefix')))

   pkgconf.set('libdir', '${prefix}/@0@'.format(get_option('libdir')))

   pkgconf.set('VERSION', meson.project_version())

+  pkglibs = cpp_lib

+  if libm_dep.found()

+    pkglibs += ' -lm'

+  endif

+  if pthread_dep.found()

+    pkglibs += ' -lpthread'

+  endif

   if t == '-static'

-    do_install = false

-    pkgconf.set('LIBS', '-lstdc++ -lpthread -lm')

+    pkgconf.set('LIBS', pkglibs)

     pkgconf.set('LIBS_PRIVATE', '')

   else

-    do_install = true

     pkgconf.set('LIBS', '')

-    pkgconf.set('LIBS_PRIVATE', '-lstdc++ -lpthread -lm')

+    pkgconf.set('LIBS_PRIVATE', pkglibs)

   endif

-  message('do_install: @0@'.format(do_install))

   configure_file(

     input: 'openh264.pc.in',