yading@10: @ yading@10: @ ARMv4 optimized DSP utils yading@10: @ Copyright (c) 2004 AGAWA Koji yading@10: @ yading@10: @ This file is part of FFmpeg. yading@10: @ yading@10: @ FFmpeg is free software; you can redistribute it and/or yading@10: @ modify it under the terms of the GNU Lesser General Public yading@10: @ License as published by the Free Software Foundation; either yading@10: @ version 2.1 of the License, or (at your option) any later version. yading@10: @ yading@10: @ FFmpeg is distributed in the hope that it will be useful, yading@10: @ but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: @ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: @ Lesser General Public License for more details. yading@10: @ yading@10: @ You should have received a copy of the GNU Lesser General Public yading@10: @ License along with FFmpeg; if not, write to the Free Software yading@10: @ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: @ yading@10: yading@10: #include "config.h" yading@10: #include "libavutil/arm/asm.S" yading@10: yading@10: #if !HAVE_ARMV5TE_EXTERNAL yading@10: #define pld @ yading@10: #endif yading@10: yading@10: .macro ALIGN_QWORD_D shift, Rd0, Rd1, Rd2, Rd3, Rn0, Rn1, Rn2, Rn3, Rn4 yading@10: mov \Rd0, \Rn0, lsr #(\shift * 8) yading@10: mov \Rd1, \Rn1, lsr #(\shift * 8) yading@10: mov \Rd2, \Rn2, lsr #(\shift * 8) yading@10: mov \Rd3, \Rn3, lsr #(\shift * 8) yading@10: orr \Rd0, \Rd0, \Rn1, lsl #(32 - \shift * 8) yading@10: orr \Rd1, \Rd1, \Rn2, lsl #(32 - \shift * 8) yading@10: orr \Rd2, \Rd2, \Rn3, lsl #(32 - \shift * 8) yading@10: orr \Rd3, \Rd3, \Rn4, lsl #(32 - \shift * 8) yading@10: .endm yading@10: .macro ALIGN_DWORD shift, R0, R1, R2 yading@10: mov \R0, \R0, lsr #(\shift * 8) yading@10: orr \R0, \R0, \R1, lsl #(32 - \shift * 8) yading@10: mov \R1, \R1, lsr #(\shift * 8) yading@10: orr \R1, \R1, \R2, lsl #(32 - \shift * 8) yading@10: .endm yading@10: .macro ALIGN_DWORD_D shift, Rdst0, Rdst1, Rsrc0, Rsrc1, Rsrc2 yading@10: mov \Rdst0, \Rsrc0, lsr #(\shift * 8) yading@10: mov \Rdst1, \Rsrc1, lsr #(\shift * 8) yading@10: orr \Rdst0, \Rdst0, \Rsrc1, lsl #(32 - (\shift * 8)) yading@10: orr \Rdst1, \Rdst1, \Rsrc2, lsl #(32 - (\shift * 8)) yading@10: .endm yading@10: yading@10: .macro RND_AVG32 Rd0, Rd1, Rn0, Rn1, Rm0, Rm1, Rmask yading@10: @ Rd = (Rn | Rm) - (((Rn ^ Rm) & ~0x01010101) >> 1) yading@10: @ Rmask = 0xFEFEFEFE yading@10: @ Rn = destroy yading@10: eor \Rd0, \Rn0, \Rm0 yading@10: eor \Rd1, \Rn1, \Rm1 yading@10: orr \Rn0, \Rn0, \Rm0 yading@10: orr \Rn1, \Rn1, \Rm1 yading@10: and \Rd0, \Rd0, \Rmask yading@10: and \Rd1, \Rd1, \Rmask yading@10: sub \Rd0, \Rn0, \Rd0, lsr #1 yading@10: sub \Rd1, \Rn1, \Rd1, lsr #1 yading@10: .endm yading@10: yading@10: .macro NO_RND_AVG32 Rd0, Rd1, Rn0, Rn1, Rm0, Rm1, Rmask yading@10: @ Rd = (Rn & Rm) - (((Rn ^ Rm) & ~0x01010101) >> 1) yading@10: @ Rmask = 0xFEFEFEFE yading@10: @ Rn = destroy yading@10: eor \Rd0, \Rn0, \Rm0 yading@10: eor \Rd1, \Rn1, \Rm1 yading@10: and \Rn0, \Rn0, \Rm0 yading@10: and \Rn1, \Rn1, \Rm1 yading@10: and \Rd0, \Rd0, \Rmask yading@10: and \Rd1, \Rd1, \Rmask yading@10: add \Rd0, \Rn0, \Rd0, lsr #1 yading@10: add \Rd1, \Rn1, \Rd1, lsr #1 yading@10: .endm yading@10: yading@10: .macro JMP_ALIGN tmp, reg yading@10: ands \tmp, \reg, #3 yading@10: bic \reg, \reg, #3 yading@10: beq 1f yading@10: subs \tmp, \tmp, #1 yading@10: beq 2f yading@10: subs \tmp, \tmp, #1 yading@10: beq 3f yading@10: b 4f yading@10: .endm yading@10: yading@10: @ ---------------------------------------------------------------- yading@10: .align 5 yading@10: function ff_put_pixels16_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r11, lr} yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r7} yading@10: add r1, r1, r2 yading@10: stm r0, {r4-r7} yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: pop {r4-r11, pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r8} yading@10: add r1, r1, r2 yading@10: ALIGN_QWORD_D 1, r9, r10, r11, r12, r4, r5, r6, r7, r8 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r9-r12} yading@10: add r0, r0, r2 yading@10: bne 2b yading@10: pop {r4-r11, pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r8} yading@10: add r1, r1, r2 yading@10: ALIGN_QWORD_D 2, r9, r10, r11, r12, r4, r5, r6, r7, r8 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r9-r12} yading@10: add r0, r0, r2 yading@10: bne 3b yading@10: pop {r4-r11, pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r8} yading@10: add r1, r1, r2 yading@10: ALIGN_QWORD_D 3, r9, r10, r11, r12, r4, r5, r6, r7, r8 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r9-r12} yading@10: add r0, r0, r2 yading@10: bne 4b yading@10: pop {r4-r11,pc} yading@10: endfunc yading@10: yading@10: @ ---------------------------------------------------------------- yading@10: .align 5 yading@10: function ff_put_pixels8_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r5,lr} yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r5} yading@10: add r1, r1, r2 yading@10: subs r3, r3, #1 yading@10: pld [r1] yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: pop {r4-r5,pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r5, r12} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD 1, r4, r5, r12 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 2b yading@10: pop {r4-r5,pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r5, r12} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD 2, r4, r5, r12 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 3b yading@10: pop {r4-r5,pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r5, r12} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD 3, r4, r5, r12 yading@10: pld [r1] yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 4b yading@10: pop {r4-r5,pc} yading@10: endfunc yading@10: yading@10: @ ---------------------------------------------------------------- yading@10: .align 5 yading@10: function ff_put_pixels8_x2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r10,lr} yading@10: ldr r12, =0xfefefefe yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 1, r6, r7, r4, r5, r10 yading@10: pld [r1] yading@10: RND_AVG32 r8, r9, r4, r5, r6, r7, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 1, r6, r7, r4, r5, r10 yading@10: ALIGN_DWORD_D 2, r8, r9, r4, r5, r10 yading@10: pld [r1] yading@10: RND_AVG32 r4, r5, r6, r7, r8, r9, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 2b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 2, r6, r7, r4, r5, r10 yading@10: ALIGN_DWORD_D 3, r8, r9, r4, r5, r10 yading@10: pld [r1] yading@10: RND_AVG32 r4, r5, r6, r7, r8, r9, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 3b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 3, r6, r7, r4, r5, r10 yading@10: pld [r1] yading@10: RND_AVG32 r8, r9, r6, r7, r5, r10, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 4b yading@10: pop {r4-r10,pc} yading@10: endfunc yading@10: yading@10: .align 5 yading@10: function ff_put_no_rnd_pixels8_x2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r10,lr} yading@10: ldr r12, =0xfefefefe yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 1, r6, r7, r4, r5, r10 yading@10: pld [r1] yading@10: NO_RND_AVG32 r8, r9, r4, r5, r6, r7, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 1b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 1, r6, r7, r4, r5, r10 yading@10: ALIGN_DWORD_D 2, r8, r9, r4, r5, r10 yading@10: pld [r1] yading@10: NO_RND_AVG32 r4, r5, r6, r7, r8, r9, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 2b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 2, r6, r7, r4, r5, r10 yading@10: ALIGN_DWORD_D 3, r8, r9, r4, r5, r10 yading@10: pld [r1] yading@10: NO_RND_AVG32 r4, r5, r6, r7, r8, r9, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bne 3b yading@10: pop {r4-r10,pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r5, r10} yading@10: add r1, r1, r2 yading@10: ALIGN_DWORD_D 3, r6, r7, r4, r5, r10 yading@10: pld [r1] yading@10: NO_RND_AVG32 r8, r9, r6, r7, r5, r10, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 4b yading@10: pop {r4-r10,pc} yading@10: endfunc yading@10: yading@10: yading@10: @ ---------------------------------------------------------------- yading@10: .align 5 yading@10: function ff_put_pixels8_y2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r11,lr} yading@10: mov r3, r3, lsr #1 yading@10: ldr r12, =0xfefefefe yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r5} yading@10: add r1, r1, r2 yading@10: 6: ldm r1, {r6-r7} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: RND_AVG32 r8, r9, r4, r5, r6, r7, r12 yading@10: ldm r1, {r4-r5} yading@10: add r1, r1, r2 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: pld [r1] yading@10: RND_AVG32 r8, r9, r6, r7, r4, r5, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r7, r8, r9 yading@10: RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r7, r8, r9 yading@10: RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r7, r8, r9 yading@10: RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: endfunc yading@10: yading@10: .align 5 yading@10: function ff_put_no_rnd_pixels8_y2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r11,lr} yading@10: mov r3, r3, lsr #1 yading@10: ldr r12, =0xfefefefe yading@10: JMP_ALIGN r5, r1 yading@10: 1: yading@10: ldm r1, {r4-r5} yading@10: add r1, r1, r2 yading@10: 6: ldm r1, {r6-r7} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: NO_RND_AVG32 r8, r9, r4, r5, r6, r7, r12 yading@10: ldm r1, {r4-r5} yading@10: add r1, r1, r2 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: pld [r1] yading@10: NO_RND_AVG32 r8, r9, r6, r7, r4, r5, r12 yading@10: subs r3, r3, #1 yading@10: stm r0, {r8-r9} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 2: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r7, r8, r9 yading@10: NO_RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 1, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: NO_RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 3: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r7, r8, r9 yading@10: NO_RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 2, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: NO_RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: .align 5 yading@10: 4: yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r4, r5, r6 yading@10: 6: ldm r1, {r7-r9} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r7, r8, r9 yading@10: NO_RND_AVG32 r10, r11, r4, r5, r7, r8, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: ldm r1, {r4-r6} yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: ALIGN_DWORD 3, r4, r5, r6 yading@10: subs r3, r3, #1 yading@10: NO_RND_AVG32 r10, r11, r7, r8, r4, r5, r12 yading@10: stm r0, {r10-r11} yading@10: add r0, r0, r2 yading@10: bne 6b yading@10: pop {r4-r11,pc} yading@10: endfunc yading@10: yading@10: .ltorg yading@10: yading@10: @ ---------------------------------------------------------------- yading@10: .macro RND_XY2_IT align, rnd yading@10: @ l1= (a & 0x03030303) + (b & 0x03030303) ?(+ 0x02020202) yading@10: @ h1= ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2) yading@10: .if \align == 0 yading@10: ldm r1, {r6-r8} yading@10: .elseif \align == 3 yading@10: ldm r1, {r5-r7} yading@10: .else yading@10: ldm r1, {r8-r10} yading@10: .endif yading@10: add r1, r1, r2 yading@10: pld [r1] yading@10: .if \align == 0 yading@10: ALIGN_DWORD_D 1, r4, r5, r6, r7, r8 yading@10: .elseif \align == 1 yading@10: ALIGN_DWORD_D 1, r4, r5, r8, r9, r10 yading@10: ALIGN_DWORD_D 2, r6, r7, r8, r9, r10 yading@10: .elseif \align == 2 yading@10: ALIGN_DWORD_D 2, r4, r5, r8, r9, r10 yading@10: ALIGN_DWORD_D 3, r6, r7, r8, r9, r10 yading@10: .elseif \align == 3 yading@10: ALIGN_DWORD_D 3, r4, r5, r5, r6, r7 yading@10: .endif yading@10: ldr r14, =0x03030303 yading@10: tst r3, #1 yading@10: and r8, r4, r14 yading@10: and r9, r5, r14 yading@10: and r10, r6, r14 yading@10: and r11, r7, r14 yading@10: it eq yading@10: andeq r14, r14, r14, \rnd #1 yading@10: add r8, r8, r10 yading@10: add r9, r9, r11 yading@10: ldr r12, =0xfcfcfcfc >> 2 yading@10: itt eq yading@10: addeq r8, r8, r14 yading@10: addeq r9, r9, r14 yading@10: and r4, r12, r4, lsr #2 yading@10: and r5, r12, r5, lsr #2 yading@10: and r6, r12, r6, lsr #2 yading@10: and r7, r12, r7, lsr #2 yading@10: add r10, r4, r6 yading@10: add r11, r5, r7 yading@10: subs r3, r3, #1 yading@10: .endm yading@10: yading@10: .macro RND_XY2_EXPAND align, rnd yading@10: RND_XY2_IT \align, \rnd yading@10: 6: push {r8-r11} yading@10: RND_XY2_IT \align, \rnd yading@10: pop {r4-r7} yading@10: add r4, r4, r8 yading@10: add r5, r5, r9 yading@10: ldr r14, =0x0f0f0f0f yading@10: add r6, r6, r10 yading@10: add r7, r7, r11 yading@10: and r4, r14, r4, lsr #2 yading@10: and r5, r14, r5, lsr #2 yading@10: add r4, r4, r6 yading@10: add r5, r5, r7 yading@10: stm r0, {r4-r5} yading@10: add r0, r0, r2 yading@10: bge 6b yading@10: pop {r4-r11,pc} yading@10: .endm yading@10: yading@10: .align 5 yading@10: function ff_put_pixels8_xy2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r11,lr} @ R14 is also called LR yading@10: JMP_ALIGN r5, r1 yading@10: 1: RND_XY2_EXPAND 0, lsl yading@10: .align 5 yading@10: 2: RND_XY2_EXPAND 1, lsl yading@10: .align 5 yading@10: 3: RND_XY2_EXPAND 2, lsl yading@10: .align 5 yading@10: 4: RND_XY2_EXPAND 3, lsl yading@10: endfunc yading@10: yading@10: .align 5 yading@10: function ff_put_no_rnd_pixels8_xy2_arm, export=1 yading@10: @ void func(uint8_t *block, const uint8_t *pixels, int line_size, int h) yading@10: @ block = word aligned, pixles = unaligned yading@10: pld [r1] yading@10: push {r4-r11,lr} yading@10: JMP_ALIGN r5, r1 yading@10: 1: RND_XY2_EXPAND 0, lsr yading@10: .align 5 yading@10: 2: RND_XY2_EXPAND 1, lsr yading@10: .align 5 yading@10: 3: RND_XY2_EXPAND 2, lsr yading@10: .align 5 yading@10: 4: RND_XY2_EXPAND 3, lsr yading@10: endfunc