comparison core/FormatConvert.S @ 318:f7b19ea31bbb prerelease

Added NEON vectorised float<->int converters. Curiously, its performance is worse than the C version. I guess clang is better at vectorising than we thought! The new code remains disabled for now.

+@
+@ FormatConvert.S
+@
+@ NEON-based vector functions for converting data between int
+@ and float with clipping and optional level scaling. 
+@ Written in assembly for performance reasons.
+@
+@ Bela: http://bela.io
+@
+@ (c) 2016 Andrew McPherson
+@ Centre for Digital Music
+@ Queen Mary University of London
+@
+
+	.syntax unified
+	.arch armv7-a
+	.fpu neon
+
+@	void int16_to_float_audio(int numSamples, int16_t *inBuffer, float *outBuffer);
+@
+@	Convert 16-bit signed ints to floats between -1.0 and 1.0. Used for audio input.
+@
+@	numSamples should be a multiple of 4 (i.e. with stereo audio, frames should be even)
+@	inBuffer should be aligned on an 8-byte boundary.
+@ 	outBuffer should be aligned on a 16-byte boundary.
+
+@ Registers:
+@    r0: numSamples            How many frames to convert
+@    r1: inBuffer              Buffer for input samples from ADC
+@    r2: outBuffer             Buffer to store output samples for render function
+
+dIFAu_Input		.dn		D0.S16
+qIFAu_Expanded	.qn		Q1.S32
+qIFAu_Output	.qn		Q2.F32
+dIFAu_Output_0	.dn		D4.F32
+dIFAu_Output_1	.dn		D5.F32
+
+	.align	2
+	.global	int16_to_float_audio
+	.thumb
+	.thumb_func
+	.type	int16_to_float_audio, %function
+int16_to_float_audio:
+	lsr r0, r0, #2
+	lsl r0, r0, #2				@ Clip off the last two bits of numSamples (only multiples of 4 allowed)
+	cmp r0, #0					@ Check for trivial case: zero samples
+	it eq
+	bxeq lr						@ Return if that's the case
+
+	@ ---- loop: iterate over the number of samples ----
+int16_to_float_audio_loop:
+	vld1 {dIFAu_Input}, [r1]!						@ Load four 16-bit signed ints from inBuffer++
+	vmovl qIFAu_Expanded, dIFAu_Input				@ Convert four 16-bit signed ints to 32-bit signed ints
+	vcvt qIFAu_Output, qIFAu_Expanded, #15			@ Convert four Q17.15 fixed points to floats (equiv. to / 32768)
+	vst1 {dIFAu_Output_0, dIFAu_Output_1}, [r2]!	@ Store floats back into memory
+
+	subs r0, r0, #4							@ numSamples -= 4
+	it gt
+	bgt int16_to_float_audio_loop			@ Loop if numSamples > 0
+
+	bx lr
+	
+
+
+@	void int16_to_float_analog(int numSamples, uint16_t *inBuffer, float *outBuffer);
+@
+@	Convert 16-bit unsigned ints to floats between 0.0 and 1.0. Used for analog input.
+@
+@	numSamples should be a multiple of 4 (i.e. integer numFrames at 4 channels)
+@	inBuffer should be aligned on an 8-byte boundary.
+@ 	outBuffer should be aligned on a 16-byte boundary.
+
+@ Registers:
+@    r0: numSamples            How many frames to convert
+@    r1: inBuffer              Buffer for input samples from ADC
+@    r2: outBuffer             Buffer to store output samples for render function
+
+dIFAn_Input		.dn		D0.U16
+qIFAn_Expanded	.qn		Q1.U32
+qIFAn_Output	.qn		Q2.F32
+dIFAn_Output_0	.dn		D4.F32
+dIFAn_Output_1	.dn		D5.F32
+
+	.align	2
+	.global	int16_to_float_analog
+	.thumb
+	.thumb_func
+	.type	int16_to_float_analog, %function
+int16_to_float_analog:
+	lsr r0, r0, #2
+	lsl r0, r0, #2				@ Clip off the last two bits of numSamples (only multiples of 4 allowed)
+	cmp r0, #0					@ Check for trivial case: zero samples
+	it eq
+	bxeq lr						@ Return if that's the case
+
+	@ ---- loop: iterate over the number of samples ----
+int16_to_float_analog_loop:
+	vld1 {dIFAn_Input}, [r1]!						@ Load four 16-bit signed ints from inBuffer++
+	vmovl qIFAn_Expanded, dIFAn_Input				@ Convert four 16-bit signed ints to 32-bit signed ints
+	vcvt qIFAn_Output, qIFAn_Expanded, #16			@ Convert four Q16.16 fixed points to floats (equiv. to / 65536)
+	vst1 {dIFAn_Output_0, dIFAn_Output_1}, [r2]!	@ Store floats back into memory
+
+	subs r0, r0, #4							@ numSamples -= 4
+	it gt
+	bgt int16_to_float_analog_loop			@ Loop if numSamples > 0
+
+	bx lr
+	
+
+@	void float_to_int16_audio(int numSamples, float *inBuffer, int16_t *outBuffer);
+@
+@	Convert floats between -1.0 and 1.0 to 16-bit signed ints, with saturation.
+@	Used for audio output.
+@
+@	numSamples should be a multiple of 4 (i.e. with stereo audio, frames should be even)
+@	inBuffer should be aligned on a 16-byte boundary.
+@ 	outBuffer should be aligned on an 8-byte boundary.
+
+@ Registers:
+@    r0: numSamples            How many frames to convert
+@    r1: inBuffer              Buffer for input samples from render function
+@    r2: outBuffer             Buffer to store output samples for DAC
+
+qFIAu_Input		.qn		Q0.F32
+dFIAu_Input_0	.dn		D0.F32
+dFIAu_Input_1,	.dn		D1.F32
+qFIAu_Converted	.qn		Q1.S32
+dFIAu_Narrowed	.dn		D4.S16	
+
+	.align	2
+	.global	float_to_int16_audio
+	.thumb
+	.thumb_func
+	.type	float_to_int16_audio, %function
+	
+float_to_int16_audio:
+	lsr r0, r0, #2
+	lsl r0, r0, #2				@ Clip off the last two bits of numSamples (only multiples of 4 allowed)
+	cmp r0, #0					@ Check for trivial case: zero samples
+	it eq
+	bxeq lr						@ Return if that's the case
+
+	@ ---- loop: iterate over the number of samples ----
+float_to_int16_audio_loop:
+	vld1 {dFIAu_Input_0, dFIAu_Input_1}, [r1]!		@ Load four floats from inBuffer++
+	vcvt qFIAu_Converted, qFIAu_Input, #15			@ Convert four floats into four Q17.15 fixed points (equiv. to * 32768)
+													@ This will truncate the result to a 32-bit representable value
+	vqmovn dFIAu_Narrowed, qFIAu_Converted			@ Convert four 32-bit signed ints to 16-bit signed ints, with saturation
+	vst1 {dFIAu_Narrowed}, [r2]!					@ Store ints back into memory
+
+	subs r0, r0, #4							@ numSamples -= 4
+	it gt
+	bgt float_to_int16_audio_loop			@ Loop if numSamples > 0
+
+	bx lr
+	
+
+@	void float_to_int16_analog(int numSamples, float *inBuffer, uint16_t *outBuffer);
+@
+@	Convert floats between 0.0 and 1.0 to 16-bit unsigned ints, with saturation.
+@	Used for analog output.
+@
+@	numSamples should be a multiple of 4 (i.e. with stereo audio, frames should be even)
+@	inBuffer should be aligned on a 16-byte boundary.
+@ 	outBuffer should be aligned on an 8-byte boundary.
+
+@ Registers:
+@    r0: numSamples            How many frames to convert
+@    r1: inBuffer              Buffer for input samples from render function
+@    r2: outBuffer             Buffer to store output samples for DAC
+
+qFIAn_Input		.qn		Q0.F32
+dFIAn_Input_0	.dn		D0.F32
+dFIAn_Input_1,	.dn		D1.F32
+qFIAn_Converted	.qn		Q1.U32
+dFIAn_Narrowed	.dn		D4.U16	
+
+	.align	2
+	.global	float_to_int16_analog
+	.thumb
+	.thumb_func
+	.type	float_to_int16_analog, %function
+	
+float_to_int16_analog:
+	lsr r0, r0, #2
+	lsl r0, r0, #2				@ Clip off the last two bits of numSamples (only multiples of 4 allowed)
+	cmp r0, #0					@ Check for trivial case: zero samples
+	it eq
+	bxeq lr						@ Return if that's the case
+
+	@ ---- loop: iterate over the number of samples ----
+float_to_int16_analog_loop:
+	vld1 {dFIAn_Input_0, dFIAn_Input_1}, [r1]!		@ Load four floats from inBuffer++
+	vcvt qFIAn_Converted, qFIAn_Input, #16			@ Convert four floats into four Q16.16 fixed points (equiv. to * 65536)
+													@ This will truncate the result to a 32-bit representable value
+	vqmovn dFIAn_Narrowed, qFIAn_Converted			@ Convert four 32-bit unsigned ints to 16-bit unsigned ints, with saturation
+	vst1 {dFIAn_Narrowed}, [r2]!					@ Store ints back into memory
+
+	subs r0, r0, #4							@ numSamples -= 4
+	it gt
+	bgt float_to_int16_analog_loop			@ Loop if numSamples > 0
+
+	bx lr
+