yading@11: /* yading@11: * ARM NEON optimised Float DSP functions yading@11: * Copyright (c) 2008 Mans Rullgard yading@11: * yading@11: * This file is part of Libav. yading@11: * yading@11: * Libav is free software; you can redistribute it and/or yading@11: * modify it under the terms of the GNU Lesser General Public yading@11: * License as published by the Free Software Foundation; either yading@11: * version 2.1 of the License, or (at your option) any later version. yading@11: * yading@11: * Libav is distributed in the hope that it will be useful, yading@11: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@11: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@11: * Lesser General Public License for more details. yading@11: * yading@11: * You should have received a copy of the GNU Lesser General Public yading@11: * License along with Libav; if not, write to the Free Software yading@11: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@11: */ yading@11: yading@11: #include "config.h" yading@11: #include "asm.S" yading@11: yading@11: function ff_vector_fmul_neon, export=1 yading@11: subs r3, r3, #8 yading@11: vld1.32 {d0-d3}, [r1,:128]! yading@11: vld1.32 {d4-d7}, [r2,:128]! yading@11: vmul.f32 q8, q0, q2 yading@11: vmul.f32 q9, q1, q3 yading@11: beq 3f yading@11: bics ip, r3, #15 yading@11: beq 2f yading@11: 1: subs ip, ip, #16 yading@11: vld1.32 {d0-d1}, [r1,:128]! yading@11: vld1.32 {d4-d5}, [r2,:128]! yading@11: vmul.f32 q10, q0, q2 yading@11: vld1.32 {d2-d3}, [r1,:128]! yading@11: vld1.32 {d6-d7}, [r2,:128]! yading@11: vmul.f32 q11, q1, q3 yading@11: vst1.32 {d16-d19},[r0,:128]! yading@11: vld1.32 {d0-d1}, [r1,:128]! yading@11: vld1.32 {d4-d5}, [r2,:128]! yading@11: vmul.f32 q8, q0, q2 yading@11: vld1.32 {d2-d3}, [r1,:128]! yading@11: vld1.32 {d6-d7}, [r2,:128]! yading@11: vmul.f32 q9, q1, q3 yading@11: vst1.32 {d20-d23},[r0,:128]! yading@11: bne 1b yading@11: ands r3, r3, #15 yading@11: beq 3f yading@11: 2: vld1.32 {d0-d1}, [r1,:128]! yading@11: vld1.32 {d4-d5}, [r2,:128]! yading@11: vst1.32 {d16-d17},[r0,:128]! yading@11: vmul.f32 q8, q0, q2 yading@11: vld1.32 {d2-d3}, [r1,:128]! yading@11: vld1.32 {d6-d7}, [r2,:128]! yading@11: vst1.32 {d18-d19},[r0,:128]! yading@11: vmul.f32 q9, q1, q3 yading@11: 3: vst1.32 {d16-d19},[r0,:128]! yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_vector_fmac_scalar_neon, export=1 yading@11: VFP len .req r2 yading@11: VFP acc .req r3 yading@11: NOVFP len .req r3 yading@11: NOVFP acc .req r2 yading@11: VFP vdup.32 q15, d0[0] yading@11: NOVFP vdup.32 q15, r2 yading@11: bics r12, len, #15 yading@11: mov acc, r0 yading@11: beq 3f yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vld1.32 {q8}, [acc,:128]! yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vld1.32 {q9}, [acc,:128]! yading@11: 1: vmla.f32 q8, q0, q15 yading@11: vld1.32 {q2}, [r1,:128]! yading@11: vld1.32 {q10}, [acc,:128]! yading@11: vmla.f32 q9, q1, q15 yading@11: vld1.32 {q3}, [r1,:128]! yading@11: vld1.32 {q11}, [acc,:128]! yading@11: vmla.f32 q10, q2, q15 yading@11: vst1.32 {q8}, [r0,:128]! yading@11: vmla.f32 q11, q3, q15 yading@11: vst1.32 {q9}, [r0,:128]! yading@11: subs r12, r12, #16 yading@11: beq 2f yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vld1.32 {q8}, [acc,:128]! yading@11: vst1.32 {q10}, [r0,:128]! yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vld1.32 {q9}, [acc,:128]! yading@11: vst1.32 {q11}, [r0,:128]! yading@11: b 1b yading@11: 2: vst1.32 {q10}, [r0,:128]! yading@11: vst1.32 {q11}, [r0,:128]! yading@11: ands len, len, #15 yading@11: it eq yading@11: bxeq lr yading@11: 3: vld1.32 {q0}, [r1,:128]! yading@11: vld1.32 {q8}, [acc,:128]! yading@11: vmla.f32 q8, q0, q15 yading@11: vst1.32 {q8}, [r0,:128]! yading@11: subs len, len, #4 yading@11: bgt 3b yading@11: bx lr yading@11: .unreq len yading@11: endfunc yading@11: yading@11: function ff_vector_fmul_scalar_neon, export=1 yading@11: VFP len .req r2 yading@11: NOVFP len .req r3 yading@11: VFP vdup.32 q8, d0[0] yading@11: NOVFP vdup.32 q8, r2 yading@11: bics r12, len, #15 yading@11: beq 3f yading@11: vld1.32 {q0},[r1,:128]! yading@11: vld1.32 {q1},[r1,:128]! yading@11: 1: vmul.f32 q0, q0, q8 yading@11: vld1.32 {q2},[r1,:128]! yading@11: vmul.f32 q1, q1, q8 yading@11: vld1.32 {q3},[r1,:128]! yading@11: vmul.f32 q2, q2, q8 yading@11: vst1.32 {q0},[r0,:128]! yading@11: vmul.f32 q3, q3, q8 yading@11: vst1.32 {q1},[r0,:128]! yading@11: subs r12, r12, #16 yading@11: beq 2f yading@11: vld1.32 {q0},[r1,:128]! yading@11: vst1.32 {q2},[r0,:128]! yading@11: vld1.32 {q1},[r1,:128]! yading@11: vst1.32 {q3},[r0,:128]! yading@11: b 1b yading@11: 2: vst1.32 {q2},[r0,:128]! yading@11: vst1.32 {q3},[r0,:128]! yading@11: ands len, len, #15 yading@11: it eq yading@11: bxeq lr yading@11: 3: vld1.32 {q0},[r1,:128]! yading@11: vmul.f32 q0, q0, q8 yading@11: vst1.32 {q0},[r0,:128]! yading@11: subs len, len, #4 yading@11: bgt 3b yading@11: bx lr yading@11: .unreq len yading@11: endfunc yading@11: yading@11: function ff_vector_fmul_window_neon, export=1 yading@11: push {r4,r5,lr} yading@11: ldr lr, [sp, #12] yading@11: sub r2, r2, #8 yading@11: sub r5, lr, #2 yading@11: add r2, r2, r5, lsl #2 yading@11: add r4, r3, r5, lsl #3 yading@11: add ip, r0, r5, lsl #3 yading@11: mov r5, #-16 yading@11: vld1.32 {d0,d1}, [r1,:128]! yading@11: vld1.32 {d2,d3}, [r2,:128], r5 yading@11: vld1.32 {d4,d5}, [r3,:128]! yading@11: vld1.32 {d6,d7}, [r4,:128], r5 yading@11: 1: subs lr, lr, #4 yading@11: vmul.f32 d22, d0, d4 yading@11: vrev64.32 q3, q3 yading@11: vmul.f32 d23, d1, d5 yading@11: vrev64.32 q1, q1 yading@11: vmul.f32 d20, d0, d7 yading@11: vmul.f32 d21, d1, d6 yading@11: beq 2f yading@11: vmla.f32 d22, d3, d7 yading@11: vld1.32 {d0,d1}, [r1,:128]! yading@11: vmla.f32 d23, d2, d6 yading@11: vld1.32 {d18,d19},[r2,:128], r5 yading@11: vmls.f32 d20, d3, d4 yading@11: vld1.32 {d24,d25},[r3,:128]! yading@11: vmls.f32 d21, d2, d5 yading@11: vld1.32 {d6,d7}, [r4,:128], r5 yading@11: vmov q1, q9 yading@11: vrev64.32 q11, q11 yading@11: vmov q2, q12 yading@11: vswp d22, d23 yading@11: vst1.32 {d20,d21},[r0,:128]! yading@11: vst1.32 {d22,d23},[ip,:128], r5 yading@11: b 1b yading@11: 2: vmla.f32 d22, d3, d7 yading@11: vmla.f32 d23, d2, d6 yading@11: vmls.f32 d20, d3, d4 yading@11: vmls.f32 d21, d2, d5 yading@11: vrev64.32 q11, q11 yading@11: vswp d22, d23 yading@11: vst1.32 {d20,d21},[r0,:128]! yading@11: vst1.32 {d22,d23},[ip,:128], r5 yading@11: pop {r4,r5,pc} yading@11: endfunc yading@11: yading@11: function ff_vector_fmul_add_neon, export=1 yading@11: ldr r12, [sp] yading@11: vld1.32 {q0-q1}, [r1,:128]! yading@11: vld1.32 {q8-q9}, [r2,:128]! yading@11: vld1.32 {q2-q3}, [r3,:128]! yading@11: vmul.f32 q10, q0, q8 yading@11: vmul.f32 q11, q1, q9 yading@11: 1: vadd.f32 q12, q2, q10 yading@11: vadd.f32 q13, q3, q11 yading@11: pld [r1, #16] yading@11: pld [r2, #16] yading@11: pld [r3, #16] yading@11: subs r12, r12, #8 yading@11: beq 2f yading@11: vld1.32 {q0}, [r1,:128]! yading@11: vld1.32 {q8}, [r2,:128]! yading@11: vmul.f32 q10, q0, q8 yading@11: vld1.32 {q1}, [r1,:128]! yading@11: vld1.32 {q9}, [r2,:128]! yading@11: vmul.f32 q11, q1, q9 yading@11: vld1.32 {q2-q3}, [r3,:128]! yading@11: vst1.32 {q12-q13},[r0,:128]! yading@11: b 1b yading@11: 2: vst1.32 {q12-q13},[r0,:128]! yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_vector_fmul_reverse_neon, export=1 yading@11: add r2, r2, r3, lsl #2 yading@11: sub r2, r2, #32 yading@11: mov r12, #-32 yading@11: vld1.32 {q0-q1}, [r1,:128]! yading@11: vld1.32 {q2-q3}, [r2,:128], r12 yading@11: 1: pld [r1, #32] yading@11: vrev64.32 q3, q3 yading@11: vmul.f32 d16, d0, d7 yading@11: vmul.f32 d17, d1, d6 yading@11: pld [r2, #-32] yading@11: vrev64.32 q2, q2 yading@11: vmul.f32 d18, d2, d5 yading@11: vmul.f32 d19, d3, d4 yading@11: subs r3, r3, #8 yading@11: beq 2f yading@11: vld1.32 {q0-q1}, [r1,:128]! yading@11: vld1.32 {q2-q3}, [r2,:128], r12 yading@11: vst1.32 {q8-q9}, [r0,:128]! yading@11: b 1b yading@11: 2: vst1.32 {q8-q9}, [r0,:128]! yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_butterflies_float_neon, export=1 yading@11: 1: vld1.32 {q0},[r0,:128] yading@11: vld1.32 {q1},[r1,:128] yading@11: vsub.f32 q2, q0, q1 yading@11: vadd.f32 q1, q0, q1 yading@11: vst1.32 {q2},[r1,:128]! yading@11: vst1.32 {q1},[r0,:128]! yading@11: subs r2, r2, #4 yading@11: bgt 1b yading@11: bx lr yading@11: endfunc yading@11: yading@11: function ff_scalarproduct_float_neon, export=1 yading@11: vmov.f32 q2, #0.0 yading@11: 1: vld1.32 {q0},[r0,:128]! yading@11: vld1.32 {q1},[r1,:128]! yading@11: vmla.f32 q2, q0, q1 yading@11: subs r2, r2, #4 yading@11: bgt 1b yading@11: vadd.f32 d0, d4, d5 yading@11: vpadd.f32 d0, d0, d0 yading@11: NOVFP vmov.32 r0, d0[0] yading@11: bx lr yading@11: endfunc