Use hand written fast fill to save a few more cycles

This commit is contained in:
Gwilym Kuiper 2022-06-12 16:53:46 +01:00
parent e051a17836
commit 4969672c73
3 changed files with 26 additions and 28 deletions

View file

@ -136,12 +136,34 @@ agb_arm_func agb_rs__mixer_add_stereo
agb_arm_end agb_rs__mixer_add_stereo
.section .iwram
.balign 4
constant_zero:
.rept 4
.word 0
.endr
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)
push {r4-r11}
@ zero registers r4-r7 (4 of them)
mov r8, #constant_zero
ldm r8, {r4-r7}
@ get the size of the buffer
ldr r9, agb_rs__buffer_size
@ make a copy of the output buffer pointer
mov r10, r0
1:
@ zero 4 words worth of the output buffer
stmia r10, {r4-r7}
subs r9, r9, #16
@ loop if we haven't zeroed everything
bne 1b
CONST_0 .req r7
CONST_FF .req r8
CONST_127 .req r9

View file

@ -5,7 +5,7 @@ use bare_metal::{CriticalSection, Mutex};
use super::hw;
use super::hw::LeftOrRight;
use super::{SoundChannel, SoundPriority};
use crate::syscall::cpu_fast_fill_i8;
use crate::{
fixnum::Num,
interrupt::free,
@ -245,9 +245,6 @@ impl MixerBuffer {
channel.playback_speed
};
let right_amount = ((channel.panning + 1) / 2) * channel.volume;
let left_amount = ((-channel.panning + 1) / 2) * channel.volume;
if (channel.pos + playback_speed * constants::SOUND_BUFFER_SIZE).floor()
>= channel.data.len()
{
@ -268,6 +265,9 @@ impl MixerBuffer {
);
}
} else {
let right_amount = ((channel.panning + 1) / 2) * channel.volume;
let left_amount = ((-channel.panning + 1) / 2) * channel.volume;
unsafe {
agb_rs__mixer_add(
channel.data.as_ptr().add(channel.pos.floor()),
@ -285,7 +285,6 @@ impl MixerBuffer {
let write_buffer_index = free(|cs| self.state.borrow(cs).borrow_mut().active_advanced());
let write_buffer = &mut self.buffers[write_buffer_index].0;
cpu_fast_fill_i8(write_buffer, 0);
unsafe {
agb_rs__mixer_collapse(write_buffer.as_mut_ptr(), buffer.as_ptr());

View file

@ -120,29 +120,6 @@ pub fn arc_tan2(x: i16, y: i32) -> i16 {
result
}
pub(crate) fn cpu_fast_fill_i8(input: &mut [i8], new_content: i32) {
assert_eq!(
input.len() % (4 * 8),
0,
"Input length must be divisible by 32"
);
let input_ptr = [new_content].as_ptr();
let output_ptr = input.as_mut_ptr();
let length_mode = (1 << 24) | // copy
(input.len() / 4);
unsafe {
asm!(
"swi 0x0c",
in("r0") input_ptr,
in("r1") output_ptr,
in("r2") length_mode,
lateout("r3") _,
);
}
}
// pub fn affine_matrix(
// x_scale: Num<i16, 8>,
// y_scale: Num<i16, 8>,