ref: 7cd946937ae499ba3a966f19c7c4d4544154623a
parent: 8e1295201e96a01cde0752418ae460833db3b5af
author: Janne Grunau <[email protected]>
date: Thu Mar 26 17:44:45 EDT 2020
meson/x86: add option to disable AVX-512 asm Allows building with nasm < 2.14.
--- a/meson.build
+++ b/meson.build
@@ -362,20 +362,11 @@
cdata_asm.set10('PREFIX', true)endif
-# Generate config.h
-config_h_target = configure_file(output: 'config.h', configuration: cdata)
-
-
-
#
# ASM specific stuff
#
if is_asm_enabled and host_machine.cpu_family().startswith('x86')- # Generate config.asm
- config_asm_target = configure_file(output: 'config.asm', output_format: 'nasm', configuration: cdata_asm)
-
-
# NASM compiler support
nasm = find_program('nasm')@@ -390,14 +381,22 @@
out = nasm_r.stdout().strip().split()
if out[1].to_lower() == 'version'
- if out[2].version_compare('<2.14')- error('nasm 2.14 or later is required, found nasm @0@'.format(out[2]))+ if out[2].version_compare('<2.13.02')+ error('nasm 2.13.02 or later is required, found nasm @0@'.format(out[2]))+ elif out[2].version_compare('<2.14') and get_option('enable_avx512')+ error('nasm 2.14 or later is required for AVX-512 asm.\n' ++ 'AVX-512 asm can be disabled with \'-Denable_avx512=false\'')
endif
+ cdata.set10('HAVE_AVX512ICL', get_option('enable_avx512'))+ cdata_asm.set10('HAVE_AVX512ICL', get_option('enable_avx512'))else
error('unexpected nasm version string: @0@'.format(nasm_r.stdout()))endif
endif
+ # Generate config.asm
+ config_asm_target = configure_file(output: 'config.asm', output_format: 'nasm', configuration: cdata_asm)
+
if host_machine.system() == 'windows'
nasm_format = 'win'
elif host_machine.system() == 'darwin'
@@ -424,6 +423,10 @@
'-o', '@OUTPUT@'
])
endif
+
+
+# Generate config.h
+config_h_target = configure_file(output: 'config.h', configuration: cdata)
--- a/meson_options.txt
+++ b/meson_options.txt
@@ -10,6 +10,11 @@
value: true,
description: 'Build asm files, if available')
+option('enable_avx512',+ type: 'boolean',
+ value: true,
+ description: 'Build AVX-512 asm files, requires nasm 2.14')
+
option('enable_tools',type: 'boolean',
value: true,
--- a/src/x86/cdef.asm
+++ b/src/x86/cdef.asm
@@ -1849,6 +1849,8 @@
; 4e 4f 40 41 42 43 44 45
; 5e 5f 50 51 52 53 54 55
+%if HAVE_AVX512ICL
+
INIT_ZMM avx512icl
cglobal cdef_filter_4x4, 4, 8, 13, dst, stride, left, top, pri, sec, dir, damping, edge
%define base r7-edge_mask
@@ -1997,5 +1999,7 @@
pminub m5, m12
vpdpbusd m0, m5, m9
ret
+
+%endif ; HAVE_AVX512ICL
%endif ; ARCH_X86_64
--- a/src/x86/cdef_init_tmpl.c
+++ b/src/x86/cdef_init_tmpl.c
@@ -84,7 +84,7 @@
if (!(flags & DAV1D_X86_CPU_FLAG_AVX512ICL)) return;
-#if BITDEPTH == 8
+#if HAVE_AVX512ICL && BITDEPTH == 8
c->fb[2] = dav1d_cdef_filter_4x4_avx512icl;
#endif
--- a/src/x86/mc.asm
+++ b/src/x86/mc.asm
@@ -240,6 +240,7 @@
BIDIR_JMP_TABLE blend_v_avx2, 2, 4, 8, 16, 32
BIDIR_JMP_TABLE blend_h_avx2, 2, 4, 8, 16, 32, 32, 32
+%if HAVE_AVX512ICL
BASE_JMP_TABLE prep, avx512icl, 4, 8, 16, 32, 64, 128
HV_JMP_TABLE prep, bilin, avx512icl, 7, 4, 8, 16, 32, 64, 128
HV_JMP_TABLE prep, 8tap, avx512icl, 7, 4, 8, 16, 32, 64, 128
@@ -249,6 +250,7 @@
BIDIR_JMP_TABLE w_mask_420_avx512icl, 4, 8, 16, 32, 64, 128
BIDIR_JMP_TABLE w_mask_422_avx512icl, 4, 8, 16, 32, 64, 128
BIDIR_JMP_TABLE w_mask_444_avx512icl, 4, 8, 16, 32, 64, 128
+%endif ; HAVE_AVX512ICL
SECTION .text
@@ -5501,6 +5503,7 @@
jg .w128_loop
RET
+%if HAVE_AVX512ICL
INIT_ZMM avx512icl
PREP_BILIN
PREP_8TAP
@@ -6022,5 +6025,7 @@
dec hd
jg .w128_loop
RET
+
+%endif ; HAVE_AVX512ICL
%endif ; ARCH_X86_64
--- a/src/x86/mc_init_tmpl.c
+++ b/src/x86/mc_init_tmpl.c
@@ -224,7 +224,7 @@
if (!(flags & DAV1D_X86_CPU_FLAG_AVX512ICL))
return;
-#if BITDEPTH == 8
+#if HAVE_AVX512ICL && BITDEPTH == 8
init_mct_fn(FILTER_2D_8TAP_REGULAR, 8tap_regular, avx512icl);
init_mct_fn(FILTER_2D_8TAP_REGULAR_SMOOTH, 8tap_regular_smooth, avx512icl);
init_mct_fn(FILTER_2D_8TAP_REGULAR_SHARP, 8tap_regular_sharp, avx512icl);