Linux 5.0-ad1: -march=native support
Subsystems (9)
-ad1 patchset adds support for compiling kernel with "-march=native" compiler option optimizing kernel for the specific CPU. "-march=native" has been available in userspace for a long time and is trivial to enable in Gentoo specifically.
"-march=native" can be used for folks like me who compile kernels on their home machines and never share binaries.
See the link for more information:
https://www.shlomifish.org/humour/by-others/funroll-loops/Gentoo-is-Rice.html
Requirements:
Intel CPU x86_64 arch
Usage:
# apply -ad1 patchset
# copy regular kernel .config
# enable "-march=native" support in # "Processor type and features" # "Processor family"
# ensure CONFIG_MARCH_NATIVE is enabled $ grep -e CONFIG_MARCH_NATIVE .config
# workaround kbuild race condition if "-j" is used $ make syncconfig
# build the kernel $ make ...
# install kernel, reboot into new kernel
# verify detected CONFIG_MARCH_NATIVE_* options $ gzip -d </proc/config.gz | grep -e CONFIG_MARCH_NATIVE
Ping me over email of something breaks or horribly slow.
Original announcement:
https://marc.info/?l=linux-kernel&m=151268659328024&w=4
---
Makefile | 16 +++ arch/x86/Kconfig.cpu | 8 + arch/x86/Makefile | 27 +++++- arch/x86/boot/compressed/head_64.S | 4 arch/x86/crypto/des3_ede-asm_64.S | 28 ++++++ arch/x86/crypto/sha1_ssse3_asm.S | 7 + arch/x86/include/asm/arch_hweight.h | 28 ++++++ arch/x86/include/asm/page_64.h | 26 ++++++ arch/x86/include/asm/segment.h | 1 arch/x86/kernel/relocate_kernel_64.S | 15 +++ arch/x86/kernel/verify_cpu.S | 27 ++++++ arch/x86/lib/Makefile | 12 ++ arch/x86/lib/memcpy_64.S | 13 +++ arch/x86/lib/memset_64.S | 15 +++ arch/x86/lib/usercopy_64.c | 16 +++ arch/x86/platform/pvh/head.S | 4 drivers/net/wireless/mediatek/mt76/mac80211.c | 2 include/linux/bitops.h | 2 lib/Makefile | 2 scripts/kconfig/.gitignore | 1 scripts/kconfig/Makefile | 7 + scripts/kconfig/cpuid.c | 108 ++++++++++++++++++++++++++ scripts/march-native.sh | 74 +++++++++++++++++ 23 files changed, 429 insertions(+), 14 deletions(-)
Makefile +13 -3
@@ -2,7 +2,7 @@ VERSION = 5 PATCHLEVEL = 0 SUBLEVEL = 0 -EXTRAVERSION = +EXTRAVERSION = -ad1 NAME = Shy Crocodile # *DOCUMENTATION*
@@ -370,10 +370,10 @@ HOST_LFS_LIBS := $(shell getconf LFS_LIBS 2>/dev/null) HOSTCC = gcc HOSTCXX = g++ -KBUILD_HOSTCFLAGS := -Wall -Wmissing-prototypes -Wstrict-prototypes -O2 \ +KBUILD_HOSTCFLAGS := -Wall -Wmissing-prototypes -Wstrict-prototypes -march=native -O2 \ -fomit-frame-pointer -std=gnu89 $(HOST_LFS_CFLAGS) \ $(HOSTCFLAGS) -KBUILD_HOSTCXXFLAGS := -O2 $(HOST_LFS_CFLAGS) $(HOSTCXXFLAGS) +KBUILD_HOSTCXXFLAGS := -march=native -O2 $(HOST_LFS_CFLAGS) $(HOSTCXXFLAGS) KBUILD_HOSTLDFLAGS := $(HOST_LFS_LDFLAGS) $(HOSTLDFLAGS) KBUILD_HOSTLDLIBS := $(HOST_LFS_LIBS) $(HOSTLDLIBS)
@@ -594,6 +594,16 @@ ifeq ($(dot-config),1) include include/config/auto.conf endif +ifdef CONFIG_MARCH_NATIVE +KBUILD_CFLAGS += -march=native +endif +ifdef CONFIG_MARCH_NATIVE_REP_MOVSB +KBUILD_CFLAGS += -mmemcpy-strategy=rep_byte:-1:align,rep_byte:-1:noalign +endif +ifdef CONFIG_MARCH_NATIVE_REP_STOSB +KBUILD_CFLAGS += -mmemset-strategy=rep_byte:-1:align,rep_byte:-1:noalign +endif + # The all: target is the default when no target is given on the # command line. # This allow a user to issue only 'make' to build a kernel including modules
arch/x86/Kconfig.cpu +8 -0
@@ -287,6 +287,12 @@ config GENERIC_CPU Generic x86-64 CPU. Run equally well on all x86-64 CPUs. +config MARCH_NATIVE + bool "-march=native" + depends on X86_64 + ---help--- + -march=native support. + endchoice config X86_GENERIC
@@ -307,6 +313,7 @@ config X86_INTERNODE_CACHE_SHIFT int default "12" if X86_VSMP default X86_L1_CACHE_SHIFT + depends on !MARCH_NATIVE config X86_L1_CACHE_SHIFT int
@@ -314,6 +321,7 @@ config X86_L1_CACHE_SHIFT default "6" if MK7 || MK8 || MPENTIUMM || MCORE2 || MATOM || MVIAC7 || X86_GENERIC || GENERIC_CPU default "4" if MELAN || M486 || MGEODEGX1 default "5" if MWINCHIP3D || MWINCHIPC6 || MCRUSOE || MEFFICEON || MCYRIXIII || MK6 || MPENTIUMIII || MPENTIUMII || M686 || M586MMX || M586TSC || M586 || MVIAC3_2 || MGEODE_LX + depends on !MARCH_NATIVE config X86_F00F_BUG def_bool y
arch/x86/Makefile +24 -3
@@ -12,6 +12,28 @@ else KBUILD_DEFCONFIG := $(ARCH)_defconfig endif +CFLAGS_NO_FP := +CFLAGS_NO_FP += $(call cc-option,-mno-mmx,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse2,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse3,) +CFLAGS_NO_FP += $(call cc-option,-mno-ssse3,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse4,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse4a,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse4.1,) +CFLAGS_NO_FP += $(call cc-option,-mno-sse4.2,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx2,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512f,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512pf,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512er,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512cd,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512vl,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512bw,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512dq,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512ifma,) +CFLAGS_NO_FP += $(call cc-option,-mno-avx512vbmi,) + # For gcc stack alignment is specified with -mpreferred-stack-boundary, # clang has the option -mstack-alignment for that purpose. ifneq ($(call cc-option, -mpreferred-stack-boundary=4),)
@@ -34,7 +56,7 @@ M16_CFLAGS := $(call cc-option, -m16, $(CODE16GCC_CFLAGS)) REALMODE_CFLAGS := $(M16_CFLAGS) -g -Os -DDISABLE_BRANCH_PROFILING \ -Wall -Wstrict-prototypes -march=i386 -mregparm=3 \ -fno-strict-aliasing -fomit-frame-pointer -fno-pic \ - -mno-mmx -mno-sse + $(CFLAGS_NO_FP) REALMODE_CFLAGS += $(call __cc-option, $(CC), $(REALMODE_CFLAGS), -ffreestanding) REALMODE_CFLAGS += $(call __cc-option, $(CC), $(REALMODE_CFLAGS), -fno-stack-protector)
@@ -57,8 +79,7 @@ endif # # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383 # -KBUILD_CFLAGS += -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -KBUILD_CFLAGS += $(call cc-option,-mno-avx,) +KBUILD_CFLAGS += $(CFLAGS_NO_FP) ifeq ($(CONFIG_X86_32),y) BITS := 32
arch/x86/boot/compressed/head_64.S +4 -0
@@ -517,8 +517,12 @@ relocated: leaq _bss(%rip), %rdi leaq _ebss(%rip), %rcx subq %rdi, %rcx +#ifdef CONFIG_MARCH_NATIVE_REP_STOSB + rep stosb +#else shrq $3, %rcx rep stosq +#endif /* * Do the extraction, and jump to the new kernel..
arch/x86/crypto/des3_ede-asm_64.S +28 -0
@@ -159,6 +159,15 @@ #define dummy2(a, b) /*_*/ +#ifdef CONFIG_MARCH_NATIVE_MOVBE +#define read_block(io, left, right) \ + movbe (io), left##d; \ + movbe 4(io), right##d; + +#define write_block(io, left, right) \ + movbe left##d, (io); \ + movbe right##d, 4(io); +#else #define read_block(io, left, right) \ movl (io), left##d; \ movl 4(io), right##d; \
@@ -170,6 +179,7 @@ bswapl right##d; \ movl left##d, (io); \ movl right##d, 4(io); +#endif ENTRY(des3_ede_x86_64_crypt_blk) /* input:
@@ -443,6 +453,14 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way) pushq %rsi /* dst */ /* load input */ +#ifdef CONFIG_MARCH_NATIVE_MOVBE + movbe 0 * 4(%rdx), RL0d; + movbe 1 * 4(%rdx), RR0d; + movbe 2 * 4(%rdx), RL1d; + movbe 3 * 4(%rdx), RR1d; + movbe 4 * 4(%rdx), RL2d; + movbe 5 * 4(%rdx), RR2d; +#else movl 0 * 4(%rdx), RL0d; movl 1 * 4(%rdx), RR0d; movl 2 * 4(%rdx), RL1d;
@@ -456,6 +474,7 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way) bswapl RR1d; bswapl RL2d; bswapl RR2d; +#endif initial_permutation3(RL, RR);
@@ -516,6 +535,14 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way) final_permutation3(RR, RL); +#ifdef CONFIG_MARCH_NATIVE_MOVBE + movbe RR0d, 0 * 4(%rsi); + movbe RL0d, 1 * 4(%rsi); + movbe RR1d, 2 * 4(%rsi); + movbe RL1d, 3 * 4(%rsi); + movbe RR2d, 4 * 4(%rsi); + movbe RL2d, 5 * 4(%rsi); +#else bswapl RR0d; bswapl RL0d; bswapl RR1d;
@@ -530,6 +557,7 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way) movl RL1d, 3 * 4(%rsi); movl RR2d, 4 * 4(%rsi); movl RL2d, 5 * 4(%rsi); +#endif popq %r15; popq %r14;
arch/x86/crypto/sha1_ssse3_asm.S +6 -1
@@ -94,10 +94,15 @@ SHA1_PIPELINED_MAIN_BODY # cleanup workspace - mov $8, %ecx mov %rsp, %rdi xor %eax, %eax +#ifdef CONFIG_MARCH_NATIVE_REP_STOSB + mov $64, %ecx + rep stosb +#else + mov $8, %ecx rep stosq +#endif mov %rbp, %rsp # deallocate workspace pop %rbp
arch/x86/include/asm/arch_hweight.h +26 -2
@@ -2,6 +2,30 @@#ifndef _ASM_X86_HWEIGHT_H#define _ASM_X86_HWEIGHT_H +#define __HAVE_ARCH_SW_HWEIGHT + +#ifdef CONFIG_MARCH_NATIVE_POPCNT +static inline unsigned int __arch_hweight64 (uint64_t x) +{ +return __builtin_popcountll (x); +} + +static inline unsigned int __arch_hweight32 (uint32_t x) +{ +return __builtin_popcount (x); +} + +static inline unsigned int __arch_hweight16 (uint16_t x) +{ +return __builtin_popcount (x); +} + +static inline unsigned int __arch_hweight8 (uint8_t x) +{ +return __builtin_popcount (x); +} +#else +#include <asm/cpufeatures.h> #ifdef CONFIG_64BIT
@@ -12,8 +36,6 @@#define REG_OUT "a"#endif -#define __HAVE_ARCH_SW_HWEIGHT - static __always_inline unsigned int __arch_hweight32(unsigned int w) {unsigned int res;
@@ -55,3 +77,5 @@ static __always_inline unsigned long __arch_hweight64(__u64 w)#endif /* CONFIG_X86_32 */ #endif + +#endif
arch/x86/include/asm/page_64.h +26 -0
@@ -40,6 +40,18 @@ extern unsigned long __phys_addr_symbol(unsigned long);#define pfn_valid (pfn) ((pfn) < max_pfn)#endif +#ifdef CONFIG_MARCH_NATIVE_REP_STOSB +static __always_inline void clear_page(void *page) +{ +uint32_t len =PAGE_SIZE ; + asmvolatile ( +"rep stosb" + :"+D" (page),"+c" (len) + :"a" (0 ) + :"memory" + ); +} +#else void clear_page_orig (void *page);void clear_page_rep (void *page);void clear_page_erms (void *page);
@@ -53,8 +65,22 @@ static inline void clear_page(void *page) "0" (page) : "cc", "memory", "rax", "rcx"); } +#endif +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB +static __always_inline void copy_page(void *to, void *from) +{ +uint32_t len =PAGE_SIZE ; + asmvolatile ( +"rep movsb" + :"+D" (to),"+S" (from),"+c" (len) + : + :"memory" + ); +} +#else void copy_page (void *to,void *from); +#endif #endif /* !__ASSEMBLY__ */
arch/x86/include/asm/segment.h +1 -0
@@ -4,6 +4,7 @@#include <linux/const.h> #include <asm/alternative.h> +#include <asm/cpufeatures.h> /* * Constructor for a conventional segment GDT (or LDT) entry.
arch/x86/kernel/relocate_kernel_64.S +15 -0
@@ -268,18 +268,33 @@ swap_pages: movq %rsi, %rax movq %r10, %rdi +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB + mov $4096, %ecx + rep movsb +#else movl $512, %ecx rep ; movsq +#endif movq %rax, %rdi movq %rdx, %rsi +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB + mov $4096, %ecx + rep movsb +#else movl $512, %ecx rep ; movsq +#endif movq %rdx, %rdi movq %r10, %rsi +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB + mov $4096, %ecx + rep movsb +#else movl $512, %ecx rep ; movsq +#endif lea PAGE_SIZE(%rax), %rsi jmp 0b
arch/x86/kernel/verify_cpu.S +27 -0
@@ -136,6 +136,33 @@ ENTRY(verify_cpu) movl $1,%eax ret .Lverify_cpu_sse_ok: + +#ifdef CONFIG_MARCH_NATIVE_POPCNT + mov $1, %eax + cpuid + bt $23, %ecx + jnc .Lverify_cpu_no_longmode +#endif + +#ifdef CONFIG_MARCH_NATIVE_MOVBE + mov $1, %eax + cpuid + bt $22, %ecx + jnc .Lverify_cpu_no_longmode +#endif + +#if defined(CONFIG_MARCH_NATIVE_REP_MOVSB) || defined(CONFIG_MARCH_NATIVE_REP_STOSB) + xor %eax, %eax + cpuid + cmp $7, %eax + jb .Lverify_cpu_no_longmode + mov $7, %eax + xor %ecx, %ecx + cpuid + bt $9, %ebx + jnc .Lverify_cpu_no_longmode +#endif + popf # Restore caller passed flags xorl %eax, %eax ret
arch/x86/lib/Makefile +10 -2
@@ -29,7 +29,10 @@ lib-$(CONFIG_RANDOMIZE_BASE) += kaslr.o lib-$(CONFIG_FUNCTION_ERROR_INJECTION) += error-inject.o lib-$(CONFIG_RETPOLINE) += retpoline.o -obj-y += msr.o msr-reg.o msr-reg-export.o hweight.o +obj-y += msr.o msr-reg.o msr-reg-export.o +ifneq ($(CONFIG_MARCH_NATIVE_POPCNT),y) + obj-y += hweight.o +endif obj-y += iomem.o ifeq ($(CONFIG_X86_32),y)
@@ -45,7 +48,12 @@ endif else obj-y += iomap_copy_64.o lib-y += csum-partial_64.o csum-copy_64.o csum-wrappers_64.o - lib-y += clear_page_64.o copy_page_64.o +ifneq ($(CONFIG_MARCH_NATIVE_REP_STOSB),y) + lib-y += clear_page_64.o +endif +ifneq ($(CONFIG_MARCH_NATIVE_REP_MOVSB),y) + lib-y += copy_page_64.o +endif lib-y += memmove_64.o memset_64.o lib-y += copy_user_64.o lib-y += cmpxchg16b_emu.o
arch/x86/lib/memcpy_64.S +13 -0
@@ -16,6 +16,18 @@ .weak memcpy +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB +ENTRY(__memcpy) +ENTRY(memcpy) + mov %rdi, %rax + mov %rdx, %rcx + rep movsb + ret +ENDPROC(memcpy) +ENDPROC(__memcpy) +EXPORT_SYMBOL(memcpy) +EXPORT_SYMBOL(__memcpy) +#else /* * memcpy - Copy a memory block. *
@@ -182,6 +194,7 @@ ENTRY(memcpy_orig) .Lend: retq ENDPROC(memcpy_orig) +#endif #ifndef CONFIG_UML
arch/x86/lib/memset_64.S +15 -0
@@ -8,6 +8,20 @@ .weak memset +#ifdef CONFIG_MARCH_NATIVE_REP_STOSB +ENTRY(memset) +ENTRY(__memset) + mov %esi, %eax + mov %rdi, %rsi + mov %rdx, %rcx + rep stosb + mov %rsi, %rax + ret +ENDPROC(memset) +ENDPROC(__memset) +EXPORT_SYMBOL(memset) +EXPORT_SYMBOL(__memset) +#else /* * ISO C memset - set a memory block to a byte value. This function uses fast * string to get better performance than the original function. The code is
@@ -140,3 +154,4 @@ ENTRY(memset_orig) jmp .Lafter_bad_alignment .Lfinal: ENDPROC(memset_orig) +#endif
arch/x86/lib/usercopy_64.c +15 -1
@@ -15,11 +15,23 @@unsigned long __clear_user (void __user *addr,unsigned long size) { -long __d0;might_fault ();/* no memory constraint because it doesn't change any memory gcc knows about */ stac (); + +#ifdef CONFIG_MARCH_NATIVE_REP_STOSB + asmvolatile ( +"0: rep stosb\n" +"1:\n" + _ASM_EXTABLE(0b,1b) + : "+D" (addr), "+c" (size) + : "a" (0) + : "memory" + ); +#else + { +long __d0; asmvolatile (" testq %[size8],%[size8]\n" " jz 4f\n"
@@ -41,6 +53,8 @@ unsigned long __clear_user(void __user *addr, unsigned long size)_ASM_EXTABLE_UA (1b ,2b ) : [size8] "=&c"(size), [dst] "=&D" (__d0) : [size1] "r"(size & 7), "[size8]" (size / 8), "[dst]"(addr)); + } +#endif clac ();return size; }
arch/x86/platform/pvh/head.S +4 -0
@@ -64,9 +64,13 @@ ENTRY(pvh_start_xen) mov $_pa(pvh_start_info), %edi mov %ebx, %esi mov _pa(pvh_start_info_sz), %ecx +#ifdef CONFIG_MARCH_NATIVE_REP_MOVSB + rep movsb +#else shr $2,%ecx rep movsl +#endif mov $_pa(early_stack_end), %esp
drivers/net/wireless/mediatek/mt76/mac80211.c +1 -1
@@ -124,7 +124,7 @@ static void mt76_init_stream_cap(struct mt76_dev *dev, bool vht) {struct ieee80211_sta_ht_cap *ht_cap = &sband->ht_cap; -int i, nstream =__sw_hweight8 (dev->antenna_mask); +int i, nstream =hweight8 (dev->antenna_mask);struct ieee80211_sta_vht_cap *vht_cap;u16 mcs_map =0 ;
include/linux/bitops.h +2 -0
@@ -7,10 +7,12 @@#define BITS_PER_TYPE (type) (sizeof(type) * BITS_PER_BYTE)#define BITS_TO_LONGS (nr) DIV_ROUND_UP(nr, BITS_PER_TYPE(long)) +#ifndef CONFIG_MARCH_NATIVE_POPCNT extern unsigned int __sw_hweight8 (unsigned int w);extern unsigned int __sw_hweight16 (unsigned int w);extern unsigned int __sw_hweight32 (unsigned int w);extern unsigned long __sw_hweight64 (__u64 w); +#endif /* * Include this here because some architectures need generic_ffs/fls in
lib/Makefile +2 -0
@@ -93,7 +93,9 @@ obj-$(CONFIG_DEBUG_LOCKING_API_SELFTESTS) += locking-selftest.o obj-y += logic_pio.o +ifneq ($(CONFIG_MARCH_NATIVE_POPCNT),y) obj-$(CONFIG_GENERIC_HWEIGHT) += hweight.o +endif obj-$(CONFIG_BTREE) += btree.o obj-$(CONFIG_INTERVAL_TREE) += interval_tree.o
scripts/kconfig/.gitignore +1 -0
@@ -8,6 +8,7 @@ # configuration programs # conf +cpuid mconf nconf qconf
scripts/kconfig/Makefile +6 -1
@@ -65,8 +65,9 @@ simple-targets := oldconfig allnoconfig allyesconfig allmodconfig \ alldefconfig randconfig listnewconfig olddefconfig syncconfig PHONY += $(simple-targets) -$(simple-targets): $(obj)/conf +$(simple-targets): $(obj)/conf $(obj)/cpuid $< $(silent) --$@ $(Kconfig) + $(Q)$(srctree)/scripts/march-native.sh $(CC) $(obj)/cpuid PHONY += savedefconfig defconfig
@@ -149,6 +150,10 @@ $(obj)/zconf.lex.o: $(obj)/zconf.tab.h HOSTCFLAGS_zconf.lex.o := -I$(src) HOSTCFLAGS_zconf.tab.o := -I$(src) +# cpuid: -march=native, CONFIG_MARCH_NATIVE_* detection +hostprogs-y += cpuid +cpuid-objs := cpuid.o + # conf: Used for defconfig, oldconfig and related targets hostprogs-y += conf conf-objs := conf.o $(common-objs)
new file mode 100644
scripts/kconfig/cpuid.c new file +108 -0
108 lines not on this page · raw message
new file mode 100755
scripts/march-native.sh new file +74 -0
74 lines not on this page · raw message
Quoted 27 lines · Alexey Dobriyan
-ad1 patchset adds support for compiling kernel with "-march=native" compiler option optimizing kernel for the specific CPU. "-march=native" has been available in userspace for a long time and is trivial to enable in Gentoo specifically.
"-march=native" can be used for folks like me who compile kernels on their home machines and never share binaries.
See the link for more information:
https://www.shlomifish.org/humour/by-others/funroll-loops/Gentoo-is-Rice.html
Requirements:
Intel CPU x86_64 archUsage:
# apply -ad1 patchset# copy regular kernel .config# enable "-march=native" support in # "Processor type and features" # "Processor family"
Quoted 5 lines
# ensure CONFIG_MARCH_NATIVE is enabled $ grep -e CONFIG_MARCH_NATIVE .config# workaround kbuild race condition if "-j" is used $ make syncconfig
Linus is unlikely to take this directly; you may want to Cc X86 MAINTAINERS.
Pavel
Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html