yading@10: /* yading@10: * Copyright (c) 2011 Janne Grunau yading@10: * yading@10: * This file is part of Libav. yading@10: * yading@10: * Libav 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: * Libav 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 Libav; 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 "libavutil/arm/asm.S" yading@10: #include "neon.S" yading@10: yading@10: .macro rv34_inv_transform r0 yading@10: vld1.16 {q14-q15}, [\r0,:128] yading@10: vmov.s16 d0, #13 yading@10: vshll.s16 q12, d29, #3 yading@10: vshll.s16 q13, d29, #4 yading@10: vshll.s16 q9, d31, #3 yading@10: vshll.s16 q1, d31, #4 yading@10: vmull.s16 q10, d28, d0 yading@10: vmlal.s16 q10, d30, d0 yading@10: vmull.s16 q11, d28, d0 yading@10: vmlsl.s16 q11, d30, d0 yading@10: vsubw.s16 q12, q12, d29 @ z2 = block[i+4*1]*7 yading@10: vaddw.s16 q13, q13, d29 @ z3 = block[i+4*1]*17 yading@10: vsubw.s16 q9, q9, d31 yading@10: vaddw.s16 q1, q1, d31 yading@10: vadd.s32 q13, q13, q9 @ z3 = 17*block[i+4*1] + 7*block[i+4*3] yading@10: vsub.s32 q12, q12, q1 @ z2 = 7*block[i+4*1] - 17*block[i+4*3] yading@10: vadd.s32 q1, q10, q13 @ z0 + z3 yading@10: vadd.s32 q2, q11, q12 @ z1 + z2 yading@10: vsub.s32 q8, q10, q13 @ z0 - z3 yading@10: vsub.s32 q3, q11, q12 @ z1 - z2 yading@10: vtrn.32 q1, q2 yading@10: vtrn.32 q3, q8 yading@10: vswp d3, d6 yading@10: vswp d5, d16 yading@10: vmov.s32 d0, #13 yading@10: vadd.s32 q10, q1, q3 yading@10: vsub.s32 q11, q1, q3 yading@10: vshl.s32 q12, q2, #3 yading@10: vshl.s32 q9, q2, #4 yading@10: vmul.s32 q13, q11, d0[0] yading@10: vshl.s32 q11, q8, #4 yading@10: vadd.s32 q9, q9, q2 yading@10: vshl.s32 q15, q8, #3 yading@10: vsub.s32 q12, q12, q2 yading@10: vadd.s32 q11, q11, q8 yading@10: vmul.s32 q14, q10, d0[0] yading@10: vsub.s32 q8, q15, q8 yading@10: vsub.s32 q12, q12, q11 yading@10: vadd.s32 q9, q9, q8 yading@10: vadd.s32 q2, q13, q12 @ z1 + z2 yading@10: vadd.s32 q1, q14, q9 @ z0 + z3 yading@10: vsub.s32 q3, q13, q12 @ z1 - z2 yading@10: vsub.s32 q15, q14, q9 @ z0 - z3 yading@10: .endm yading@10: yading@10: /* void rv34_idct_add_c(uint8_t *dst, int stride, int16_t *block) */ yading@10: function ff_rv34_idct_add_neon, export=1 yading@10: mov r3, r0 yading@10: rv34_inv_transform r2 yading@10: vmov.i16 q12, #0 yading@10: vrshrn.s32 d16, q1, #10 @ (z0 + z3) >> 10 yading@10: vrshrn.s32 d17, q2, #10 @ (z1 + z2) >> 10 yading@10: vrshrn.s32 d18, q3, #10 @ (z1 - z2) >> 10 yading@10: vrshrn.s32 d19, q15, #10 @ (z0 - z3) >> 10 yading@10: vld1.32 {d28[]}, [r0,:32], r1 yading@10: vld1.32 {d29[]}, [r0,:32], r1 yading@10: vtrn.32 q8, q9 yading@10: vld1.32 {d28[1]}, [r0,:32], r1 yading@10: vld1.32 {d29[1]}, [r0,:32], r1 yading@10: vst1.16 {q12}, [r2,:128]! @ memset(block, 0, 16) yading@10: vst1.16 {q12}, [r2,:128] @ memset(block+16, 0, 16) yading@10: vtrn.16 d16, d17 yading@10: vtrn.32 d28, d29 yading@10: vtrn.16 d18, d19 yading@10: vaddw.u8 q0, q8, d28 yading@10: vaddw.u8 q1, q9, d29 yading@10: vqmovun.s16 d28, q0 yading@10: vqmovun.s16 d29, q1 yading@10: vst1.32 {d28[0]}, [r3,:32], r1 yading@10: vst1.32 {d28[1]}, [r3,:32], r1 yading@10: vst1.32 {d29[0]}, [r3,:32], r1 yading@10: vst1.32 {d29[1]}, [r3,:32], r1 yading@10: bx lr yading@10: endfunc yading@10: yading@10: /* void rv34_inv_transform_noround_neon(int16_t *block); */ yading@10: function ff_rv34_inv_transform_noround_neon, export=1 yading@10: rv34_inv_transform r0 yading@10: vshl.s32 q11, q2, #1 yading@10: vshl.s32 q10, q1, #1 yading@10: vshl.s32 q12, q3, #1 yading@10: vshl.s32 q13, q15, #1 yading@10: vadd.s32 q11, q11, q2 yading@10: vadd.s32 q10, q10, q1 yading@10: vadd.s32 q12, q12, q3 yading@10: vadd.s32 q13, q13, q15 yading@10: vshrn.s32 d0, q10, #11 @ (z0 + z3)*3 >> 11 yading@10: vshrn.s32 d1, q11, #11 @ (z1 + z2)*3 >> 11 yading@10: vshrn.s32 d2, q12, #11 @ (z1 - z2)*3 >> 11 yading@10: vshrn.s32 d3, q13, #11 @ (z0 - z3)*3 >> 11 yading@10: vst4.16 {d0[0], d1[0], d2[0], d3[0]}, [r0,:64]! yading@10: vst4.16 {d0[1], d1[1], d2[1], d3[1]}, [r0,:64]! yading@10: vst4.16 {d0[2], d1[2], d2[2], d3[2]}, [r0,:64]! yading@10: vst4.16 {d0[3], d1[3], d2[3], d3[3]}, [r0,:64]! yading@10: bx lr yading@10: endfunc yading@10: yading@10: /* void ff_rv34_idct_dc_add_neon(uint8_t *dst, int stride, int dc) */ yading@10: function ff_rv34_idct_dc_add_neon, export=1 yading@10: mov r3, r0 yading@10: vld1.32 {d28[]}, [r0,:32], r1 yading@10: vld1.32 {d29[]}, [r0,:32], r1 yading@10: vdup.16 d0, r2 yading@10: vmov.s16 d1, #169 yading@10: vld1.32 {d28[1]}, [r0,:32], r1 yading@10: vmull.s16 q1, d0, d1 @ dc * 13 * 13 yading@10: vld1.32 {d29[1]}, [r0,:32], r1 yading@10: vrshrn.s32 d0, q1, #10 @ (dc * 13 * 13 + 0x200) >> 10 yading@10: vmov d1, d0 yading@10: vaddw.u8 q2, q0, d28 yading@10: vaddw.u8 q3, q0, d29 yading@10: vqmovun.s16 d28, q2 yading@10: vqmovun.s16 d29, q3 yading@10: vst1.32 {d28[0]}, [r3,:32], r1 yading@10: vst1.32 {d29[0]}, [r3,:32], r1 yading@10: vst1.32 {d28[1]}, [r3,:32], r1 yading@10: vst1.32 {d29[1]}, [r3,:32], r1 yading@10: bx lr yading@10: endfunc yading@10: yading@10: /* void rv34_inv_transform_dc_noround_c(int16_t *block) */ yading@10: function ff_rv34_inv_transform_noround_dc_neon, export=1 yading@10: vld1.16 {d28[]}, [r0,:16] @ block[0] yading@10: vmov.i16 d4, #251 yading@10: vorr.s16 d4, #256 @ 13^2 * 3 yading@10: vmull.s16 q3, d28, d4 yading@10: vshrn.s32 d0, q3, #11 yading@10: vmov.i16 d1, d0 yading@10: vst1.64 {q0}, [r0,:128]! yading@10: vst1.64 {q0}, [r0,:128]! yading@10: bx lr yading@10: endfunc