andrewm@0: @ andrewm@0: @ audio_routines.S andrewm@0: @ andrewm@0: @ NEON-based functions for time-critical audio processing andrewm@0: @ andrewm@0: @ Andrew McPherson 2014 andrewm@0: @ Queen Mary University of London andrewm@0: @ andrewm@0: andrewm@0: .syntax unified andrewm@0: .arch armv7-a andrewm@0: .fpu neon andrewm@0: andrewm@0: @ void oscillator_bank_neon(int numAudioFrames, float *audioOut, andrewm@0: @ int activePartialNum, int lookupTableSize, andrewm@0: @ float *phases, float *frequencies, float *amplitudes, andrewm@0: @ float *freqDerivatives, float *ampDerivatives, andrewm@0: @ float *lookupTable); andrewm@0: andrewm@0: @ Registers: andrewm@0: @ r0: numAudioFrames How many frames to render andrewm@0: @ r1: audioOut Buffer for audio output samples [stereo] andrewm@0: @ r2: activePartialNum How many active partials to render andrewm@0: @ r3: lookupTableSize Size of lookup table andrewm@0: @ ---- other arguments start on the stack and are moved: ----- andrewm@0: @ r4: phases Phase of each oscillator (pointer) andrewm@0: @ r5: frequencies Normalised frequency of each oscillator (pointer) andrewm@0: @ r6: amplitudes Normalised amplitude of each oscillator (pointer) andrewm@0: @ r7: freqDerivatives Derivative of frequency for each oscillator (pointer) andrewm@0: @ r8: ampDerivatives Derivative of amplitude for each oscillator (pointer) andrewm@0: @ r9: lookupTable Lookup table containing one oscillation andrewm@0: @ andrewm@0: @ Alignment requirements: andrewm@0: @ audioOut: 8-byte boundary andrewm@0: @ phases: 16-byte boundary andrewm@0: @ frequencies: 16-byte boundary andrewm@0: @ amplitudes: 16-byte boundary andrewm@0: @ freqDerivatives: 16-byte bounary andrewm@0: @ ampDerivatives: 16-byte boundary andrewm@0: @ lookupTable: 4-byte boundary (TODO: check this) andrewm@0: andrewm@0: .align 2 andrewm@0: .global oscillator_bank_neon andrewm@0: .thumb andrewm@0: .thumb_func andrewm@0: .type oscillator_bank_neon, %function andrewm@0: oscillator_bank_neon: andrewm@0: andrewm@0: andrewm@0: dSample .dn D6.F32 andrewm@0: qPhases .qn Q8.F32 andrewm@0: dPhases_0 .dn D16.F32 andrewm@0: dPhases_1 .dn D17.F32 andrewm@0: qFreqs .qn Q9.F32 andrewm@0: dFreqs_0 .dn D18.F32 andrewm@0: dFreqs_1 .dn D19.F32 andrewm@0: qAmps .qn Q10.F32 andrewm@0: dAmps_0 .dn D20.F32 andrewm@0: dAmps_1 .dn D21.F32 andrewm@0: qFreqDs .qn Q11.F32 andrewm@0: dFreqDs_0 .dn D22.F32 andrewm@0: dFreqDs_1 .dn D23.F32 andrewm@0: qAmpDs .qn Q12.F32 andrewm@0: dAmpDs_0 .dn D24.F32 andrewm@0: dAmpDs_1 .dn D25.F32 andrewm@0: andrewm@0: qBaseInts .qn Q13.U32 @ Base indexes: unsigned ints x4 andrewm@0: dBaseInts_0 .dn D26.U32 andrewm@0: dBaseInts_1 .dn D27.U32 andrewm@0: qFractions .qn Q14.F32 @ Fraction indexes: floats x4 andrewm@0: qTableBase .qn Q15.U32 @ Base of lookup table andrewm@0: andrewm@0: cmp r0, #0 @ Check for trivial case 1: zero frames andrewm@0: it eq andrewm@0: bxeq lr @ Return if that's the case (otherwise might have odd behaviour) andrewm@0: cmp r2, #4 @ Check for trivial case 2: zero oscillators andrewm@0: it lt andrewm@0: bxlt lr @ Return if that's the case andrewm@0: andrewm@0: push {r4-r11} @ Now arguments start 32 bytes above SP andrewm@0: add r11, sp, #32 @ Pointer to 32 bytes into the stack andrewm@0: ldm r11, {r4-r9} @ Load 6 arguments into registers andrewm@0: andrewm@0: vdup qTableBase, r9 @ Move lookup table base index into 4 ints andrewm@0: andrewm@0: @ Outer loop: iterate over the number of oscillators, choosing 4 at a andrewm@0: @ time to work with. andrewm@0: oscbank_oscillator_loop: andrewm@0: vld1 {dPhases_0, dPhases_1}, [r4] @ no increment; will store at end of sample loop andrewm@0: vld1 {dFreqs_0, dFreqs_1}, [r5] andrewm@0: vld1 {dAmps_0, dAmps_1}, [r6] andrewm@0: vld1 {dFreqDs_0, dFreqDs_1}, [r7]! @ increment; won't update at end of sample loop andrewm@0: vld1 {dAmpDs_0, dAmpDs_1}, [r8]! andrewm@0: andrewm@0: push {r0-r1,r4-r8} andrewm@0: @ --- inner loop: iterate over the number of samples --- andrewm@0: oscbank_sample_loop: andrewm@0: vcvt qBaseInts, qPhases @ Take floor(phases) andrewm@0: vmov q2.f32, #1.0 @ Load 1.0 into every slot of q2 andrewm@0: vshl q0.U32, qBaseInts, #2 @ Shift the indexes left 2 (*4 for float addressing) andrewm@0: vcvt qFractions, qBaseInts @ int back to float andrewm@0: vadd q0.U32, q0.U32, qTableBase @ Find memory addresses andrewm@0: andrewm@0: vmov r4, r5, d0 @ Move two indexes to ARM registers andrewm@0: vmov r6, r7, d1 @ Move two more indexes to ARM registers andrewm@0: vsub qFractions, qPhases, qFractions @ fraction = phase - floor(phase) andrewm@0: andrewm@0: vldr.64 d0, [r4] @ Load two consecutive floats at each location andrewm@0: vldr.64 d1, [r5] @ These hold the previous and following samples in the table andrewm@0: vldr.64 d2, [r6] @ TODO: check whether these work at 4-byte alignment andrewm@0: vldr.64 d3, [r7] andrewm@0: andrewm@0: @ Format at this point: andrewm@0: @ Osc0(before) Osc0(after) Osc1(before) Osc1(after) Osc2(before) Osc2(after) Osc3(before) Osc3(after) andrewm@0: @ We want: andrewm@0: @ Osc0(before) Osc1(before) Osc2(before) Osc3(before) Osc0(after) Osc1(after) Osc2(after) Osc3(after) andrewm@0: andrewm@0: vuzp.32 q0, q1 @ Now q0 contains before, q1 contains after andrewm@0: vsub q2.f32, q2.f32, qFractions @ q2 = 1.0 - fraction andrewm@0: vmul q1.f32, q1.f32, qFractions @ q1 = fraction * after andrewm@0: vmul q0.f32, q0.f32, q2.f32 @ q0 = (1.0 - fraction) * before andrewm@0: andrewm@0: vadd qPhases, qPhases, qFreqs @ Update phases andrewm@0: vadd qFreqs, qFreqs, qFreqDs @ Update frequencies andrewm@0: andrewm@0: vadd q0.f32, q0.f32, q1.f32 @ Add two interpolated components to get the final sample andrewm@0: vdup q2.u32, r3 @ Put lookup table size into each element of q2 andrewm@0: vcvt qBaseInts, qPhases @ Take floor of new phases andrewm@0: vmul q0.f32, q0.f32, qAmps @ Multiply samples by current amplitude andrewm@0: andrewm@0: vld1 dSample, [r1] @ Load the current stereo samples andrewm@0: vpadd d2.f32, d0.f32, d1.f32 @ Pairwise accumulate q0 (output sample) into d2 andrewm@0: andrewm@0: vand q2, q2, qBaseInts @ Logical AND of new phase int leaves 1 bit set only if phase >= table size andrewm@0: vpadd d3.f32, d2.f32, d2.f32 @ Pairwise accumulate d2 into d0 --> d0[0] and d0[1] both hold total of 4 oscillators andrewm@0: vadd qAmps, qAmps, qAmpDs @ Update amplitudes andrewm@0: vcvt q0.f32, q2.u32 @ Convert int back to float after AND operation andrewm@0: andrewm@0: vadd dSample, dSample, d3.f32 @ Add oscillator outputs to each channel andrewm@0: andrewm@0: subs r0, r0, #1 @ numFrames-- andrewm@0: vsub qPhases, qPhases, q0.f32 @ Keep phases in table range andrewm@0: vst1 dSample, [r1]! @ Store back in buffer and increment by 8 andrewm@0: andrewm@0: it gt andrewm@0: bgt oscbank_sample_loop @ Loop if numFrames > 0 andrewm@0: andrewm@0: @ --- end inner loop --- andrewm@0: pop {r0-r1,r4-r8} @ Restore registers: restores audioOut and numFrames, among others andrewm@0: andrewm@0: vst1 {dPhases_0, dPhases_1}, [r4]! @ Store phases back to array andrewm@0: vst1 {dFreqs_0, dFreqs_1}, [r5]! @ Store frequencies back to array andrewm@0: vst1 {dAmps_0, dAmps_1}, [r6]! @ Store amplitudes back to array andrewm@0: @ No need to update r7, r8 andrewm@0: andrewm@0: subs r2, r2, #4 @ numPartials -= 4 andrewm@0: it gt andrewm@0: bgt oscbank_oscillator_loop @ Loop if numPartials > 0 andrewm@0: andrewm@0: pop {r4-r11} andrewm@0: bx lr