yading@10: /* yading@10: * ARM NEON optimised FFT yading@10: * yading@10: * Copyright (c) 2009 Mans Rullgard yading@10: * Copyright (c) 2009 Naotoshi Nojiri yading@10: * yading@10: * This algorithm (though not any of the implementation details) is yading@10: * based on libdjbfft by D. J. Bernstein. 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 "libavutil/arm/asm.S" yading@10: yading@10: #define M_SQRT1_2 0.70710678118654752440 yading@10: yading@10: yading@10: function fft4_neon yading@10: vld1.32 {d0-d3}, [r0,:128] yading@10: yading@10: vext.32 q8, q1, q1, #1 @ i2,r3 d3=i3,r2 yading@10: vsub.f32 d6, d0, d1 @ r0-r1,i0-i1 yading@10: vsub.f32 d7, d16, d17 @ r3-r2,i2-i3 yading@10: vadd.f32 d4, d0, d1 @ r0+r1,i0+i1 yading@10: vadd.f32 d5, d2, d3 @ i2+i3,r2+r3 yading@10: vadd.f32 d1, d6, d7 yading@10: vsub.f32 d3, d6, d7 yading@10: vadd.f32 d0, d4, d5 yading@10: vsub.f32 d2, d4, d5 yading@10: yading@10: vst1.32 {d0-d3}, [r0,:128] yading@10: yading@10: bx lr yading@10: endfunc yading@10: yading@10: function fft8_neon yading@10: mov r1, r0 yading@10: vld1.32 {d0-d3}, [r1,:128]! yading@10: vld1.32 {d16-d19}, [r1,:128] yading@10: yading@10: movw r2, #0x04f3 @ sqrt(1/2) yading@10: movt r2, #0x3f35 yading@10: eor r3, r2, #1<<31 yading@10: vdup.32 d31, r2 yading@10: yading@10: vext.32 q11, q1, q1, #1 @ i2,r3,i3,r2 yading@10: vadd.f32 d4, d16, d17 @ r4+r5,i4+i5 yading@10: vmov d28, r3, r2 yading@10: vadd.f32 d5, d18, d19 @ r6+r7,i6+i7 yading@10: vsub.f32 d17, d16, d17 @ r4-r5,i4-i5 yading@10: vsub.f32 d19, d18, d19 @ r6-r7,i6-i7 yading@10: vrev64.32 d29, d28 yading@10: vadd.f32 d20, d0, d1 @ r0+r1,i0+i1 yading@10: vadd.f32 d21, d2, d3 @ r2+r3,i2+i3 yading@10: vmul.f32 d26, d17, d28 @ -a2r*w,a2i*w yading@10: vext.32 q3, q2, q2, #1 yading@10: vmul.f32 d27, d19, d29 @ a3r*w,-a3i*w yading@10: vsub.f32 d23, d22, d23 @ i2-i3,r3-r2 yading@10: vsub.f32 d22, d0, d1 @ r0-r1,i0-i1 yading@10: vmul.f32 d24, d17, d31 @ a2r*w,a2i*w yading@10: vmul.f32 d25, d19, d31 @ a3r*w,a3i*w yading@10: vadd.f32 d0, d20, d21 yading@10: vsub.f32 d2, d20, d21 yading@10: vadd.f32 d1, d22, d23 yading@10: vrev64.32 q13, q13 yading@10: vsub.f32 d3, d22, d23 yading@10: vsub.f32 d6, d6, d7 yading@10: vadd.f32 d24, d24, d26 @ a2r+a2i,a2i-a2r t1,t2 yading@10: vadd.f32 d25, d25, d27 @ a3r-a3i,a3i+a3r t5,t6 yading@10: vadd.f32 d7, d4, d5 yading@10: vsub.f32 d18, d2, d6 yading@10: vext.32 q13, q12, q12, #1 yading@10: vadd.f32 d2, d2, d6 yading@10: vsub.f32 d16, d0, d7 yading@10: vadd.f32 d5, d25, d24 yading@10: vsub.f32 d4, d26, d27 yading@10: vadd.f32 d0, d0, d7 yading@10: vsub.f32 d17, d1, d5 yading@10: vsub.f32 d19, d3, d4 yading@10: vadd.f32 d3, d3, d4 yading@10: vadd.f32 d1, d1, d5 yading@10: yading@10: vst1.32 {d16-d19}, [r1,:128] yading@10: vst1.32 {d0-d3}, [r0,:128] yading@10: yading@10: bx lr yading@10: endfunc yading@10: yading@10: function fft16_neon yading@10: movrel r1, mppm yading@10: vld1.32 {d16-d19}, [r0,:128]! @ q8{r0,i0,r1,i1} q9{r2,i2,r3,i3} yading@10: pld [r0, #32] yading@10: vld1.32 {d2-d3}, [r1,:128] yading@10: vext.32 q13, q9, q9, #1 yading@10: vld1.32 {d22-d25}, [r0,:128]! @ q11{r4,i4,r5,i5} q12{r6,i5,r7,i7} yading@10: vadd.f32 d4, d16, d17 yading@10: vsub.f32 d5, d16, d17 yading@10: vadd.f32 d18, d18, d19 yading@10: vsub.f32 d19, d26, d27 yading@10: yading@10: vadd.f32 d20, d22, d23 yading@10: vsub.f32 d22, d22, d23 yading@10: vsub.f32 d23, d24, d25 yading@10: vadd.f32 q8, q2, q9 @ {r0,i0,r1,i1} yading@10: vadd.f32 d21, d24, d25 yading@10: vmul.f32 d24, d22, d2 yading@10: vsub.f32 q9, q2, q9 @ {r2,i2,r3,i3} yading@10: vmul.f32 d25, d23, d3 yading@10: vuzp.32 d16, d17 @ {r0,r1,i0,i1} yading@10: vmul.f32 q1, q11, d2[1] yading@10: vuzp.32 d18, d19 @ {r2,r3,i2,i3} yading@10: vrev64.32 q12, q12 yading@10: vadd.f32 q11, q12, q1 @ {t1a,t2a,t5,t6} yading@10: vld1.32 {d24-d27}, [r0,:128]! @ q12{r8,i8,r9,i9} q13{r10,i10,r11,i11} yading@10: vzip.32 q10, q11 yading@10: vld1.32 {d28-d31}, [r0,:128] @ q14{r12,i12,r13,i13} q15{r14,i14,r15,i15} yading@10: vadd.f32 d0, d22, d20 yading@10: vadd.f32 d1, d21, d23 yading@10: vsub.f32 d2, d21, d23 yading@10: vsub.f32 d3, d22, d20 yading@10: sub r0, r0, #96 yading@10: vext.32 q13, q13, q13, #1 yading@10: vsub.f32 q10, q8, q0 @ {r4,r5,i4,i5} yading@10: vadd.f32 q8, q8, q0 @ {r0,r1,i0,i1} yading@10: vext.32 q15, q15, q15, #1 yading@10: vsub.f32 q11, q9, q1 @ {r6,r7,i6,i7} yading@10: vswp d25, d26 @ q12{r8,i8,i10,r11} q13{r9,i9,i11,r10} yading@10: vadd.f32 q9, q9, q1 @ {r2,r3,i2,i3} yading@10: vswp d29, d30 @ q14{r12,i12,i14,r15} q15{r13,i13,i15,r14} yading@10: vadd.f32 q0, q12, q13 @ {t1,t2,t5,t6} yading@10: vadd.f32 q1, q14, q15 @ {t1a,t2a,t5a,t6a} yading@10: movrelx r2, X(ff_cos_16) yading@10: vsub.f32 q13, q12, q13 @ {t3,t4,t7,t8} yading@10: vrev64.32 d1, d1 yading@10: vsub.f32 q15, q14, q15 @ {t3a,t4a,t7a,t8a} yading@10: vrev64.32 d3, d3 yading@10: movrel r3, pmmp yading@10: vswp d1, d26 @ q0{t1,t2,t3,t4} q13{t6,t5,t7,t8} yading@10: vswp d3, d30 @ q1{t1a,t2a,t3a,t4a} q15{t6a,t5a,t7a,t8a} yading@10: vadd.f32 q12, q0, q13 @ {r8,i8,r9,i9} yading@10: vadd.f32 q14, q1, q15 @ {r12,i12,r13,i13} yading@10: vld1.32 {d4-d5}, [r2,:64] yading@10: vsub.f32 q13, q0, q13 @ {r10,i10,r11,i11} yading@10: vsub.f32 q15, q1, q15 @ {r14,i14,r15,i15} yading@10: vswp d25, d28 @ q12{r8,i8,r12,i12} q14{r9,i9,r13,i13} yading@10: vld1.32 {d6-d7}, [r3,:128] yading@10: vrev64.32 q1, q14 yading@10: vmul.f32 q14, q14, d4[1] yading@10: vmul.f32 q1, q1, q3 yading@10: vmla.f32 q14, q1, d5[1] @ {t1a,t2a,t5a,t6a} yading@10: vswp d27, d30 @ q13{r10,i10,r14,i14} q15{r11,i11,r15,i15} yading@10: vzip.32 q12, q14 yading@10: vadd.f32 d0, d28, d24 yading@10: vadd.f32 d1, d25, d29 yading@10: vsub.f32 d2, d25, d29 yading@10: vsub.f32 d3, d28, d24 yading@10: vsub.f32 q12, q8, q0 @ {r8,r9,i8,i9} yading@10: vadd.f32 q8, q8, q0 @ {r0,r1,i0,i1} yading@10: vsub.f32 q14, q10, q1 @ {r12,r13,i12,i13} yading@10: mov r1, #32 yading@10: vadd.f32 q10, q10, q1 @ {r4,r5,i4,i5} yading@10: vrev64.32 q0, q13 yading@10: vmul.f32 q13, q13, d5[0] yading@10: vrev64.32 q1, q15 yading@10: vmul.f32 q15, q15, d5[1] yading@10: vst2.32 {d16-d17},[r0,:128], r1 yading@10: vmul.f32 q0, q0, q3 yading@10: vst2.32 {d20-d21},[r0,:128], r1 yading@10: vmul.f32 q1, q1, q3 yading@10: vmla.f32 q13, q0, d5[0] @ {t1,t2,t5,t6} yading@10: vmla.f32 q15, q1, d4[1] @ {t1a,t2a,t5a,t6a} yading@10: vst2.32 {d24-d25},[r0,:128], r1 yading@10: vst2.32 {d28-d29},[r0,:128] yading@10: vzip.32 q13, q15 yading@10: sub r0, r0, #80 yading@10: vadd.f32 d0, d30, d26 yading@10: vadd.f32 d1, d27, d31 yading@10: vsub.f32 d2, d27, d31 yading@10: vsub.f32 d3, d30, d26 yading@10: vsub.f32 q13, q9, q0 @ {r10,r11,i10,i11} yading@10: vadd.f32 q9, q9, q0 @ {r2,r3,i2,i3} yading@10: vsub.f32 q15, q11, q1 @ {r14,r15,i14,i15} yading@10: vadd.f32 q11, q11, q1 @ {r6,r7,i6,i7} yading@10: vst2.32 {d18-d19},[r0,:128], r1 yading@10: vst2.32 {d22-d23},[r0,:128], r1 yading@10: vst2.32 {d26-d27},[r0,:128], r1 yading@10: vst2.32 {d30-d31},[r0,:128] yading@10: bx lr yading@10: endfunc yading@10: yading@10: function fft_pass_neon yading@10: push {r4-r6,lr} yading@10: mov r6, r2 @ n yading@10: lsl r5, r2, #3 @ 2 * n * sizeof FFTSample yading@10: lsl r4, r2, #4 @ 2 * n * sizeof FFTComplex yading@10: lsl r2, r2, #5 @ 4 * n * sizeof FFTComplex yading@10: add r3, r2, r4 yading@10: add r4, r4, r0 @ &z[o1] yading@10: add r2, r2, r0 @ &z[o2] yading@10: add r3, r3, r0 @ &z[o3] yading@10: vld1.32 {d20-d21},[r2,:128] @ {z[o2],z[o2+1]} yading@10: movrel r12, pmmp yading@10: vld1.32 {d22-d23},[r3,:128] @ {z[o3],z[o3+1]} yading@10: add r5, r5, r1 @ wim yading@10: vld1.32 {d6-d7}, [r12,:128] @ pmmp yading@10: vswp d21, d22 yading@10: vld1.32 {d4}, [r1,:64]! @ {wre[0],wre[1]} yading@10: sub r5, r5, #4 @ wim-- yading@10: vrev64.32 q1, q11 yading@10: vmul.f32 q11, q11, d4[1] yading@10: vmul.f32 q1, q1, q3 yading@10: vld1.32 {d5[0]}, [r5,:32] @ d5[0] = wim[-1] yading@10: vmla.f32 q11, q1, d5[0] @ {t1a,t2a,t5a,t6a} yading@10: vld2.32 {d16-d17},[r0,:128] @ {z[0],z[1]} yading@10: sub r6, r6, #1 @ n-- yading@10: vld2.32 {d18-d19},[r4,:128] @ {z[o1],z[o1+1]} yading@10: vzip.32 q10, q11 yading@10: vadd.f32 d0, d22, d20 yading@10: vadd.f32 d1, d21, d23 yading@10: vsub.f32 d2, d21, d23 yading@10: vsub.f32 d3, d22, d20 yading@10: vsub.f32 q10, q8, q0 yading@10: vadd.f32 q8, q8, q0 yading@10: vsub.f32 q11, q9, q1 yading@10: vadd.f32 q9, q9, q1 yading@10: vst2.32 {d20-d21},[r2,:128]! @ {z[o2],z[o2+1]} yading@10: vst2.32 {d16-d17},[r0,:128]! @ {z[0],z[1]} yading@10: vst2.32 {d22-d23},[r3,:128]! @ {z[o3],z[o3+1]} yading@10: vst2.32 {d18-d19},[r4,:128]! @ {z[o1],z[o1+1]} yading@10: sub r5, r5, #8 @ wim -= 2 yading@10: 1: yading@10: vld1.32 {d20-d21},[r2,:128] @ {z[o2],z[o2+1]} yading@10: vld1.32 {d22-d23},[r3,:128] @ {z[o3],z[o3+1]} yading@10: vswp d21, d22 yading@10: vld1.32 {d4}, [r1]! @ {wre[0],wre[1]} yading@10: vrev64.32 q0, q10 yading@10: vmul.f32 q10, q10, d4[0] yading@10: vrev64.32 q1, q11 yading@10: vmul.f32 q11, q11, d4[1] yading@10: vld1.32 {d5}, [r5] @ {wim[-1],wim[0]} yading@10: vmul.f32 q0, q0, q3 yading@10: sub r5, r5, #8 @ wim -= 2 yading@10: vmul.f32 q1, q1, q3 yading@10: vmla.f32 q10, q0, d5[1] @ {t1,t2,t5,t6} yading@10: vmla.f32 q11, q1, d5[0] @ {t1a,t2a,t5a,t6a} yading@10: vld2.32 {d16-d17},[r0,:128] @ {z[0],z[1]} yading@10: subs r6, r6, #1 @ n-- yading@10: vld2.32 {d18-d19},[r4,:128] @ {z[o1],z[o1+1]} yading@10: vzip.32 q10, q11 yading@10: vadd.f32 d0, d22, d20 yading@10: vadd.f32 d1, d21, d23 yading@10: vsub.f32 d2, d21, d23 yading@10: vsub.f32 d3, d22, d20 yading@10: vsub.f32 q10, q8, q0 yading@10: vadd.f32 q8, q8, q0 yading@10: vsub.f32 q11, q9, q1 yading@10: vadd.f32 q9, q9, q1 yading@10: vst2.32 {d20-d21}, [r2,:128]! @ {z[o2],z[o2+1]} yading@10: vst2.32 {d16-d17}, [r0,:128]! @ {z[0],z[1]} yading@10: vst2.32 {d22-d23}, [r3,:128]! @ {z[o3],z[o3+1]} yading@10: vst2.32 {d18-d19}, [r4,:128]! @ {z[o1],z[o1+1]} yading@10: bne 1b yading@10: yading@10: pop {r4-r6,pc} yading@10: endfunc yading@10: yading@10: .macro def_fft n, n2, n4 yading@10: .align 6 yading@10: function fft\n\()_neon yading@10: push {r4, lr} yading@10: mov r4, r0 yading@10: bl fft\n2\()_neon yading@10: add r0, r4, #\n4*2*8 yading@10: bl fft\n4\()_neon yading@10: add r0, r4, #\n4*3*8 yading@10: bl fft\n4\()_neon yading@10: mov r0, r4 yading@10: pop {r4, lr} yading@10: movrelx r1, X(ff_cos_\n) yading@10: mov r2, #\n4/2 yading@10: b fft_pass_neon yading@10: endfunc yading@10: .endm yading@10: yading@10: def_fft 32, 16, 8 yading@10: def_fft 64, 32, 16 yading@10: def_fft 128, 64, 32 yading@10: def_fft 256, 128, 64 yading@10: def_fft 512, 256, 128 yading@10: def_fft 1024, 512, 256 yading@10: def_fft 2048, 1024, 512 yading@10: def_fft 4096, 2048, 1024 yading@10: def_fft 8192, 4096, 2048 yading@10: def_fft 16384, 8192, 4096 yading@10: def_fft 32768, 16384, 8192 yading@10: def_fft 65536, 32768, 16384 yading@10: yading@10: function ff_fft_calc_neon, export=1 yading@10: ldr r2, [r0] yading@10: sub r2, r2, #2 yading@10: movrel r3, fft_tab_neon yading@10: ldr r3, [r3, r2, lsl #2] yading@10: mov r0, r1 yading@10: bx r3 yading@10: endfunc yading@10: yading@10: function ff_fft_permute_neon, export=1 yading@10: push {r4,lr} yading@10: mov r12, #1 yading@10: ldr r2, [r0] @ nbits yading@10: ldr r3, [r0, #12] @ tmp_buf yading@10: ldr r0, [r0, #8] @ revtab yading@10: lsl r12, r12, r2 yading@10: mov r2, r12 yading@10: 1: yading@10: vld1.32 {d0-d1}, [r1,:128]! yading@10: ldr r4, [r0], #4 yading@10: uxth lr, r4 yading@10: uxth r4, r4, ror #16 yading@10: add lr, r3, lr, lsl #3 yading@10: add r4, r3, r4, lsl #3 yading@10: vst1.32 {d0}, [lr,:64] yading@10: vst1.32 {d1}, [r4,:64] yading@10: subs r12, r12, #2 yading@10: bgt 1b yading@10: yading@10: sub r1, r1, r2, lsl #3 yading@10: 1: yading@10: vld1.32 {d0-d3}, [r3,:128]! yading@10: vst1.32 {d0-d3}, [r1,:128]! yading@10: subs r2, r2, #4 yading@10: bgt 1b yading@10: yading@10: pop {r4,pc} yading@10: endfunc yading@10: yading@10: const fft_tab_neon yading@10: .word fft4_neon yading@10: .word fft8_neon yading@10: .word fft16_neon yading@10: .word fft32_neon yading@10: .word fft64_neon yading@10: .word fft128_neon yading@10: .word fft256_neon yading@10: .word fft512_neon yading@10: .word fft1024_neon yading@10: .word fft2048_neon yading@10: .word fft4096_neon yading@10: .word fft8192_neon yading@10: .word fft16384_neon yading@10: .word fft32768_neon yading@10: .word fft65536_neon yading@10: endconst yading@10: yading@10: const pmmp, align=4 yading@10: .float +1.0, -1.0, -1.0, +1.0 yading@10: endconst yading@10: yading@10: const mppm, align=4 yading@10: .float -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2 yading@10: endconst