alex/agb
1
0
Fork 0
mirror of https://github.com/italicsjenga/agb.git synced 2024-12-23 08:11:33 +11:00
Code Issues Packages Projects Releases Wiki Activity

Even faster mixer (#447)

Browse source 
If you do an ldmia for loading lots of samples at once, the mixer uses
significantly less CPU (10% compared to previous, 19800 cycles per frame
-> 17701 cycles per frame for 32768Hz).

I've also added a really simple unit test for the `collapse` function to
at least gain some confidence in it.

- [x] Changelog updated / no changelog update needed
This commit is contained in:
Gwilym Inzani 2023-06-20 22:04:16 +01:00 committed by GitHub
parent 20a7c262e9 dbf7715e67
commit c3268e3de2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 119 additions and 35 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
CHANGELOG.md
View file

@ -13,6 +13,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
- Changed the default template game.
- `DynamicSprite` has a new API which changes the constructor and adds a `set_pixel` method.
- You no longer need to install arm-none-eabi-binutils. In order to write games using `agb`, you now only need to install rust nightly.
- 10% performance improvement with the software mixer.
## [0.15.0] - 2023/04/25

67
agb/src/sound/mixer/mixer.s
View file

@ -98,18 +98,15 @@ agb_arm_func agb_rs__mixer_add_stereo
@ r2 - volume to play the sound at
@
@ The sound buffer must be SOUND_BUFFER_SIZE * 2 in size = 176 * 2
push {{r4-r9}}
push {{r4-r11}}
mov r9, r2
ldr r5, =0x00000FFF
ldr r8, =agb_rs__buffer_size
ldr r8, [r8]
1:
.rept 4
ldrsh r6, [r0], #2 @ load the current sound sample to r6
ldr r4, [r1] @ read the current value
.macro add_stereo_sample sample_reg:req
ldrsh r6, [r0], #2 @ load the current sound sample to r6
@ This is slightly convoluted, but is mainly done for performance reasons. It is better
@ to hit ROM just once and then do 3 really simple instructions then do 2 ldrsbs however annoying
@ -130,15 +127,23 @@ agb_arm_func agb_rs__mixer_add_stereo
lsl r6, r6, #24 @ r6 = | R | 0 | 0 | 0 | drop everything except the right sample
orr r6, r7, r6, asr #8 @ r6 = | 1 | R | 1 | L | now we have it perfectly set up
mla r4, r6, r9, r4 @ r4 += r6 * r9 (calculating both the left and right samples together)
mla \sample_reg, r6, r2, \sample_reg @ r4 += r6 * r2 (calculating both the left and right samples together)
.endm
str r4, [r1], #4 @ store the new value, and increment the pointer
.endr
1:
ldmia r1, {{r9-r12}} @ read the current values
add_stereo_sample r9
add_stereo_sample r10
add_stereo_sample r11
add_stereo_sample r12
stmia r1!, {{r9-r12}} @ store the new value, and increment the pointer
subs r8, r8, #4 @ loop counter
bne 1b @ jump back if we're done with the loop
pop {{r4-r9}}
pop {{r4-r11}}
bx lr
agb_arm_end agb_rs__mixer_add_stereo
@ -147,22 +152,19 @@ agb_arm_func agb_rs__mixer_collapse
@ Arguments:
@ r0 = target buffer (i8)
@ r1 = input buffer (i16) of fixnums with 4 bits of precision (read in sets of i16 in an i32)
@ r2 = loop counter
push {{r4-r11}}
push {{r4-r11,lr}}
CONST_0 .req r7
CONST_FF .req r8
CONST_127 .req r9
CONST_128 .req r8
TEMP .req r10
SWAP_SIGN .req r11
ldr CONST_0, =0
ldr CONST_FF, =0xff
ldr CONST_127, =127
ldr CONST_128, =128
ldr SWAP_SIGN, =0x80808080
ldr r2, =agb_rs__buffer_size @ loop counter
ldr r2, [r2]
mov r4, r2
@ The idea for this solution came from pimpmobile:
@ -171,7 +173,7 @@ SWAP_SIGN .req r11
@ The register should be 127 bigger then what you actually want, and we'll correct for that later. Hence the
@ add instructions in `load_sample`.
@
@ The idea behind this is in the bit patters of -128 and 127 which are 10000000 and 01111111 respectively,
@ The idea behind this is in the bit patters of -128 and 127 which are 10000000 and 01111111 respectively, -x = !x + 1 => !x = -x-1
@ and we want to clamp the value between them.
@
@ The first instruction calculates `-((sample + 128) >> 8)`. If sample is between -128 and 127, then
@ -185,16 +187,13 @@ SWAP_SIGN .req r11
@ So (-1 logical >> 24) gives 11111111 and (1 logical >> 24) gives 00000000 so register is clamped between these two values.
.macro clamp_s8 reg:req
subs TEMP, CONST_0, \reg, asr #8
andne \reg, CONST_FF, TEMP, lsr #24
movne \reg, TEMP, lsr #24
.endm
.macro load_sample left_reg:req right_reg:req
@ left_reg = *r1; r1++
ldr \left_reg, [r1], #4
mov \right_reg, \left_reg, lsl #16 @ push the sample 16 bits first
add \right_reg, CONST_127, \right_reg, asr #20 @ move right sample back to being the correct value
add \left_reg, CONST_127, \left_reg, asr #20 @ now we only have the left sample
add \right_reg, CONST_128, \right_reg, asr #20 @ move right sample back to being the correct value
add \left_reg, CONST_128, \left_reg, asr #20 @ now we only have the left sample
clamp_s8 \left_reg @ clamp the audio to 8 bit values
clamp_s8 \right_reg
@ -202,19 +201,21 @@ SWAP_SIGN .req r11
1:
.rept 4
ldmia r1!, {{r3,r5,r6,r9}}
load_sample r3, r12
load_sample r5, r6
load_sample r5, lr
orr r3, r3, r5, lsl #8
orr r12, r12, r6, lsl #8
orr r12, r12, lr, lsl #8
load_sample r5, r6
orr r3, r3, r5, lsl #16
orr r12, r12, r6, lsl #16
load_sample r6, lr
orr r3, r3, r6, lsl #16
orr r12, r12, lr, lsl #16
load_sample r5, r6
orr r3, r3, r5, lsl #24
orr r12, r12, r6, lsl #24
load_sample r9, lr
orr r3, r3, r9, lsl #24
orr r12, r12, lr, lsl #24
eor r3, r3, SWAP_SIGN
eor r12, r12, SWAP_SIGN
@ -226,6 +227,6 @@ SWAP_SIGN .req r11
subs r2, r2, #16 @ r2 -= 16
bne 1b @ loop if not 0
pop {{r4-r11}}
pop {{r4-r11,lr}}
bx lr
agb_arm_end agb_rs__mixer_collapse

86
agb/src/sound/mixer/sw_mixer.rs
View file

@ -35,7 +35,11 @@ extern "C" {
volume: Num<i16, 4>,
);
fn agb_rs__mixer_collapse(sound_buffer: *mut i8, input_buffer: *const Num<i16, 4>);
fn agb_rs__mixer_collapse(
sound_buffer: *mut i8,
input_buffer: *const Num<i16, 4>,
num_samples: usize,
);
}
/// The main software mixer struct.
@ -458,7 +462,85 @@ impl MixerBuffer {
let write_buffer = free(|cs| self.state.borrow(cs).borrow_mut().active_advanced());
unsafe {
agb_rs__mixer_collapse(write_buffer, working_buffer.as_ptr());
agb_rs__mixer_collapse(
write_buffer,
working_buffer.as_ptr(),
self.frequency.buffer_size(),
);
}
}
}
#[cfg(test)]
mod test {
use crate::fixnum::num;
use alloc::vec;
use super::*;
#[test_case]
fn collapse_should_correctly_reduce_size_of_input(_: &mut crate::Gba) {
#[repr(align(4))]
struct AlignedNumbers<const N: usize>([Num<i16, 4>; N]);
let input = &AlignedNumbers([
num!(10.0),
num!(10.0),
num!(5.0),
num!(5.0),
num!(-10.0),
num!(-10.5),
num!(-5.9),
num!(-5.2),
num!(0.0),
num!(1.1),
num!(2.2),
num!(3.3),
num!(155.4),
num!(-230.5),
num!(400.6),
num!(-700.7),
num!(10.0),
num!(10.0),
num!(5.0),
num!(5.0),
num!(-10.0),
num!(-10.5),
num!(-5.9),
num!(-5.2),
num!(0.0),
num!(1.1),
num!(2.2),
num!(3.3),
num!(155.4),
num!(-230.5),
num!(400.6),
num!(-700.7),
]);
let input = &input.0;
let mut output_buffer = vec![0i32; input.len() / 4];
unsafe {
agb_rs__mixer_collapse(
output_buffer.as_mut_ptr().cast(),
input.as_ptr(),
input.len() / 2,
);
}
// output will be unzipped, so input is LRLRLRLRLRLRLR... and output is LLLLLLRRRRRR
assert_eq!(
output_buffer
.iter()
.flat_map(|x| x.to_le_bytes())
.map(|x| x as i8)
.collect::<alloc::vec::Vec<_>>(),
&[
10, 5, -10, -6, 0, 2, 127, 127, 10, 5, -10, -6, 0, 2, 127, 127, 10, 5, -11, -6, 1,
3, -128, -128, 10, 5, -11, -6, 1, 3, -128, -128
]
);
}
}