format ELF64 executable 3 entry start segment readable executable ;============================================================================== macro IACA_START { ; mov ebx,111 ; db 0x64,0x67,0x90 } macro IACA_END { ; mov ebx,222 ; db 0x64,0x67,0x90 } ;============================================================================== align 16 compute_color: ; in: , normalized x, y coordinates ; out: , , r, g, b color components ;------------------------------------------------------------------------------ vaddps ymm0,ymm0,[.c_xoffset] vxorps ymm2,ymm2,ymm2 ; ymm2 = x vmovaps ymm3,ymm2 ; ymm3 = y vmovaps ymm4,ymm2 ; ymm4 = x^2 vmovaps ymm5,ymm2 ; ymm5 = y^2 vxorps ymm8,ymm8,ymm8 mov eax,256 align 32 .loop: IACA_START ; 12.00 cycles on HSW vmulps ymm3,ymm2,ymm3 ; ymm3 = x*y vsubps ymm2,ymm4,ymm5 ; ymm2 = x^2 - y^2 vaddps ymm3,ymm3,ymm3 ; ymm3 = 2*x*y vaddps ymm2,ymm2,ymm0 ; ymm2 = xn = x^2 - y^2 + a vaddps ymm3,ymm3,ymm1 ; ymm3 = yn = 2*x*y + b vmulps ymm4,ymm2,ymm2 ; ymm4 = xn^2 vmulps ymm5,ymm3,ymm3 ; ymm5 = yn^2 vaddps ymm6,ymm2,ymm3 ; ymm6 = xn^2 + yn^2 = m2 vcmpltps ymm7,ymm6,[.c_4_0] vandnps ymm7,ymm7,[.c_1_0] vaddps ymm8,ymm8,ymm7 sub eax,1 jnz .loop IACA_END vmovaps ymm0,ymm8 vdivps ymm0,ymm0,[.c_255_0] vmulps ymm0,ymm0,ymm0 vmulps ymm0,ymm0,ymm0 vmulps ymm0,ymm0,ymm0 vmovaps ymm1,ymm0 vmovaps ymm2,ymm0 ret align 32 .c_4_0: dd 8 dup 4.0 .c_1_0: dd 8 dup 1.0 .c_255_0: dd 8 dup 255.0 .c_xoffset: dd 8 dup -0.5 ;============================================================================== align 16 thread: ;------------------------------------------------------------------------------ virtual at rsp .startx dd ? .starty dd ? end virtual push rbp rbx mov rbp,rsp sub rsp,32 and rsp,-32 .tile: mov r15d,1 lock xadd [image_tile],r15d cmp r15d,TILE_COUNT jge .finish xor edx,edx mov eax,r15d mov edi,TILE_X_COUNT div edi imul eax,TILE_SIZE imul edx,TILE_SIZE mov [.startx],edx mov [.starty],eax imul eax,IMAGE_SIZE add eax,edx shl eax,2 mov rbx,[image_ptr] add rbx,rax xor r14d,r14d .row: xor r13d,r13d align 16 .pixel: mov edx,[.startx] mov eax,[.starty] add edx,r13d add eax,r14d ;------------------------------------ compute color vmovaps ymm2,[image_size] vmovaps ymm3,ymm2 vrcpps ymm3,ymm3 vcvtsi2ss xmm0,xmm0,edx vcvtsi2ss xmm1,xmm1,eax vbroadcastss ymm0,xmm0 vbroadcastss ymm1,xmm1 vaddps ymm0,ymm0,[.c_01234567f] vaddps ymm1,ymm1,ymm1 vaddps ymm0,ymm0,ymm0 vsubps ymm0,ymm0,ymm2 vsubps ymm1,ymm1,ymm2 vmulps ymm0,ymm0,ymm3 vmulps ymm1,ymm1,ymm3 call compute_color ;------------------------------------ convert color from RGB32F to BGRA8 vxorps ymm3,ymm3,ymm3 vmovaps ymm4,[.c_1_0] vminps ymm0,ymm0,ymm4 vminps ymm1,ymm1,ymm4 vminps ymm2,ymm2,ymm4 vmaxps ymm0,ymm0,ymm3 vmaxps ymm1,ymm1,ymm3 vmaxps ymm2,ymm2,ymm3 vmovaps ymm3,[.c_255_0] vmulps ymm0,ymm0,ymm3 vmulps ymm1,ymm1,ymm3 vmulps ymm2,ymm2,ymm3 vcvttps2dq ymm0,ymm0 vcvttps2dq ymm1,ymm1 vcvttps2dq ymm2,ymm2 vpslld ymm1,ymm1,8 vpslld ymm0,ymm0,16 vpor ymm0,ymm0,[.c_ff000000] vpor ymm1,ymm1,ymm2 vpor ymm0,ymm0,ymm1 vmovdqa [rbx+r13*4],ymm0 ;------------------------------------ add r13d,8 cmp r13d,TILE_SIZE jne .pixel add rbx,IMAGE_SIZE*4 add r14d,1 cmp r14d,TILE_SIZE jne .row jmp .tile .finish: mov rsp,rbp pop rbx rbp ret align 32 .c_01234567f: dd 0.0,1.0,2.0,3.0,4.0,5.0,6.0,7.0 .c_1_0: dd 8 dup 1.0 .c_255_0: dd 8 dup 255.0 .c_ff000000: dd 8 dup 0xff000000 ;============================================================================== align 16 mem_alloc: ; in: size in bytes ; out: pointer to the allocated memory ;------------------------------------------------------------------------------ mov eax,9 ; sys_mmap mov rsi,rdi ; length xor edi,edi ; addr mov edx,0x1+0x2 ; PROT_READ | PROT_WRITE mov r10d,0x02+0x20 ; MAP_PRIVATE | MAP_ANONYMOUS mov r8,-1 ; fd xor r9d,r9d ; offset syscall ret ;============================================================================== align 16 thread_create: ; in: mutex address ;------------------------------------------------------------------------------ mov edi,4096 call mem_alloc mov rsi,rax add rsi,4096 mov eax,56 ; sys_clone mov edi,0x100+0x200+0x400+0x800+0x10000 xor edx,edx syscall test eax,eax jnz .ret call thread mov dword [rbx],1 mov eax,202 ; sys_futex mov rdi,rbx ; mutex address mov esi,1 ; FUTEX_WAKE mov edx,1 ; wake 1 thread syscall mov eax,60 ; sys_exit xor edi,edi ; exit code syscall .ret: ret ;============================================================================== align 16 thread_wait: ; in: mutex address ;------------------------------------------------------------------------------ mov eax,202 ; sys_futex mov rdi,rbx ; mutex address mov esi,0 ; FUTEX_WAIT mov edx,0 ; mutex 'running' value xor r10d,r10d ; unused but must be zero syscall ret ;============================================================================== align 16 image_save: ;------------------------------------------------------------------------------ push rbx mov eax,85 mov rdi,.tga_name mov esi,110000000b syscall mov rbx,rax mov eax,1 mov rdi,rbx mov rsi,.tga_head mov edx,18 syscall mov eax,1 mov rdi,rbx mov rsi,[image_ptr] mov edx,IMAGE_SIZE*IMAGE_SIZE*4 syscall pop rbx ret .tga_name db 'mandelbrot.tga',0 .tga_head db 0,0,2,9 dup 0 db (IMAGE_SIZE and 0x00ff),(IMAGE_SIZE and 0xff00) shr 8 db (IMAGE_SIZE and 0x00ff),(IMAGE_SIZE and 0xff00) shr 8,32,0 ;============================================================================== align 16 main: ;------------------------------------------------------------------------------ mov rdi,IMAGE_SIZE*IMAGE_SIZE*4 call mem_alloc mov [image_ptr],rax mov rbx,thread0_mutex call thread_create mov rbx,thread1_mutex call thread_create mov rbx,thread2_mutex call thread_create mov rbx,thread3_mutex call thread_create mov rbx,thread0_mutex call thread_wait mov rbx,thread1_mutex call thread_wait mov rbx,thread2_mutex call thread_wait mov rbx,thread3_mutex call thread_wait call image_save ret ;============================================================================== align 16 start: ;------------------------------------------------------------------------------ call main mov eax,60 ; sys_exit xor edi,edi ; exit code syscall ;============================================================================== segment readable writeable align 4 thread0_mutex dd 0 thread1_mutex dd 0 thread2_mutex dd 0 thread3_mutex dd 0 align 8 image_ptr dq 0 align 4 image_tile dd 0 TILE_SIZE = 64 TILE_X_COUNT = IMAGE_SIZE / TILE_SIZE TILE_Y_COUNT = IMAGE_SIZE / TILE_SIZE TILE_COUNT = TILE_X_COUNT * TILE_Y_COUNT IMAGE_SIZE = 1024 align 32 image_size: dd 8 dup 1024.0 ;==============================================================================