forked from Lephenixnoir/Azur
azur: replace P8 with P8_RGB565A (4.5 c/p), P8_RGB565 (3 c/p)
The code for P8 failed in some non-transparent cases and I'll admit I could not be bothered to fix it when the superiors formats were already designed and promised a significant boost.
This commit is contained in:
parent
18ee037693
commit
ddff9f6d6b
GPG Key ID:
1BBA026E13FC0495
|
|
@ -53,9 +53,10 @@ _azrp_shader_tex2d:
|
|||
.formats:
|
||||
.long _RGB565
|
||||
.long _RGB565A
|
||||
.long _P8
|
||||
.long _NOP
|
||||
.long _P4
|
||||
.long _P8_RGB565
|
||||
.long _P8_RGB565A
|
||||
|
||||
/* [Loop macros]
|
||||
|
||||
|
|
@ -105,7 +106,7 @@ _azrp_shader_tex2d:
|
|||
here (adding sub-cases); a super-heavy renderer with more hypotheses (like a
|
||||
tileset shader) should aim for that route though. Also, movua.l followed by
|
||||
mov.l is even slower (5 cycles). */
|
||||
|
||||
.align 4
|
||||
_RGB565:
|
||||
mov #8, r0 /* Maximum width for naive method */
|
||||
cmp/ge r7, r0
|
||||
|
|
@ -183,8 +184,8 @@ _RGB565.naive:
|
|||
longword-based optimization. Instead, we just go as fast as possible with
|
||||
each pixel, using DSP instructions because conditional execution is pretty
|
||||
damn good. This takes 4 cycles/pixel. I tried a number of reductions to
|
||||
3 cycles/pixel but could not get that to work. */
|
||||
|
||||
3 cycles/pixel but could not get any of them to work. */
|
||||
.align 4
|
||||
_RGB565A:
|
||||
shll16 r6
|
||||
mov #0x0004, r0 /* DC Zero mode */
|
||||
|
|
@ -202,152 +203,216 @@ _RGB565A:
|
|||
3: movx.w x0, @r5+
|
||||
TEX2D_END()
|
||||
|
||||
/* [Rendering strategy for the P8 format]
|
||||
/* [Rendering strategy for the P8_RGB565A format]
|
||||
|
||||
The work needed for each pixel gets more difficult as we go. In P8 there is
|
||||
both a palette indexing step (which induces some latency when moving values
|
||||
read from memory to the ALU, unlike RGB565), and an alpha comparison check.
|
||||
The work needed for each pixel gets more difficult as we go, with alpha
|
||||
being the major culprit due to its additional comparisons, jumps, and
|
||||
limited interweaving opportunities due to conditionally-executed code.
|
||||
|
||||
The rendering uses a 2-interwoven open loop. This reduces stall cycles and
|
||||
increases parallelism. Dealing with non-multiple widths is annoying as
|
||||
usual. Instead this routine avoids the clipping problem by overwriting then
|
||||
restoring the next pixel. (A delightfully smart workaround if you ask me.)
|
||||
Because arithmetic is unavoidable and there are 1-cycle delays between both
|
||||
loading-arithmetic, and arithmetic-indexing pairs, the loop has 2 interwoven
|
||||
iterations with an open structure. This fills the stall cycles and increases
|
||||
parallelism significantly. Pure interweaving handbook.
|
||||
|
||||
Unless I have missed something this routine achieves 5.5 cycles/pixel.
|
||||
Dealing with odd widths is a major pain as usual. Instead of adding logic to
|
||||
handle the extra pixel separately, this routine lets the loop overwrite it,
|
||||
then restores its original value afterwards - a delightfully elegant trick.
|
||||
|
||||
The format is not extremely friendly. It has alpha for all images, and uses
|
||||
a non-zero value for it, which burns a register. Palette indices are
|
||||
unsigned, which requires an extu.b even though the palette could be indexed
|
||||
with signed values by moving the pointer. Also the palette always takes up
|
||||
512 bytes even when a low amount of colors is used.
|
||||
The P8 format is actually so bad that spending precious time grinding cycles
|
||||
felt completely inappropriate without first refining it. This led to two new
|
||||
variations, P8_RGB565 and P8_RGB565A, which fix the following problems.
|
||||
|
||||
The P8_RGB565 and P8_RGB565A address these issues and supplant P8. In the
|
||||
interim this version of P8 is reasonably elegant despite ample extra
|
||||
registers and initial computations. */
|
||||
_P8:
|
||||
-> First there is alpha for all images, which is the most costly feature,
|
||||
single-handedly accounting for half of the work per pixel. P8_RGB565
|
||||
does no support alpha, which basically doubles performance.
|
||||
|
||||
-> Then, there is the alpha value itself. In P8 it is a variable (and fxconv
|
||||
sets it to 0xff), which burns a register for the comparison and enforces
|
||||
a fixed order between comparison and left-shift. P8_RGB565A always sets
|
||||
an alpha value of 0x00 which lifts both constraints.
|
||||
|
||||
-> Then, there are palette indices. In P8 they are unsigned, which requires
|
||||
an extu.b. In P8_RGB565 and P8_RGB565A they are signed, so the sign-
|
||||
extended value of the mov.b can be used directly (once doubled). The
|
||||
palette base is simply offset by 128 entries, with colors numbered
|
||||
-128..-1 first and only then 0..127.
|
||||
|
||||
-> Finally, there's the palette itself. In P8 it always has 256 entries,
|
||||
even when only a few are used. For small images this is a huge waste, so
|
||||
P8_RGB565 and P8_RGB565A only store colors that are actually used.
|
||||
|
||||
P8_RGB565A achieves 4.5 cycles/pixel asymptotically, which is really good
|
||||
compared to 4 cycles/pixel for RGB565A. */
|
||||
.align 4
|
||||
_P8_RGB565A:
|
||||
mov.l r13, @-r15
|
||||
add #2, r8 /* Palette */
|
||||
add #-2, r9 /* Input stride compensation for openness */
|
||||
|
||||
mov r7, r13
|
||||
shlr r7
|
||||
|
||||
mov.l r12, @-r15
|
||||
movt r12
|
||||
|
||||
mov.l r11, @-r15
|
||||
add r12, r7
|
||||
movt r6
|
||||
|
||||
mov.l r10, @-r15
|
||||
extu.b r6, r6
|
||||
|
||||
shll r13
|
||||
|
||||
add #-1, r9
|
||||
mov.w _P8_RGB565A.palette_distance, r0
|
||||
add r6, r7
|
||||
|
||||
sub r12, r9
|
||||
sub r6, r9
|
||||
|
||||
sub r12, r4
|
||||
sub r6, r4
|
||||
|
||||
sub r12, r4
|
||||
sub r6, r4
|
||||
|
||||
add r0, r8
|
||||
|
||||
add r5, r13
|
||||
mov r7, r2
|
||||
|
||||
add #-4, r5 /* Output offset compensation in the loop */
|
||||
|
||||
shll2 r2
|
||||
|
||||
add r4, r2
|
||||
nop /* 4-alignment */
|
||||
|
||||
TEX2D_START()
|
||||
|
||||
/* Save the first pixel after the line. It will be restored at the end
|
||||
of the line to correct the odd-width case where the 2-interwoven
|
||||
main loop writes an additional pixel. */
|
||||
mov.b @r3+, r6
|
||||
|
||||
/* Save next pixel for the odd-width case */
|
||||
mov.w @r13, r12
|
||||
|
||||
mov.b @r3+, r0
|
||||
|
||||
2: /* 2-interwoven open main loop */
|
||||
mov.b @r3+, r10
|
||||
extu.b r0, r0
|
||||
tst r6, r6
|
||||
|
||||
cmp/eq r0, r6
|
||||
mov.w @r5+, r2
|
||||
/* 2-interwoven open main loop */
|
||||
2: add r6, r6
|
||||
mov r6, r0
|
||||
|
||||
add r0, r0 /* Don't use shll to keep T */
|
||||
mov.w @r5, r11
|
||||
|
||||
add #-2, r5
|
||||
add r10, r10
|
||||
bt.s 5f
|
||||
|
||||
extu.b r10, r10
|
||||
mov.w @(r0,r8), r2
|
||||
tst r10, r10
|
||||
mov.w @(r0,r8), r0
|
||||
|
||||
5: cmp/eq r10, r6
|
||||
mov.w r0, @(4,r5)
|
||||
|
||||
add r10, r10 /* Don't use shll to keep T */
|
||||
5: mov.b @r3+, r6
|
||||
mov r10, r0
|
||||
|
||||
mov.w r2, @r5
|
||||
add #2, r5
|
||||
bt.s 6f
|
||||
add #4, r5
|
||||
|
||||
bt 6f
|
||||
mov.w @(r0,r8), r11
|
||||
mov.w @(r0,r8), r0
|
||||
|
||||
6: mov.b @r3+, r0
|
||||
mov.w r0, @(2,r5)
|
||||
|
||||
mov.w r11, @r5
|
||||
3: add #2, r5
|
||||
6: mov.b @r3+, r10
|
||||
3: tst r6, r6
|
||||
|
||||
/* Restore last pixel */
|
||||
mov.w r12, @r13
|
||||
add r4, r13
|
||||
|
||||
mov r7, r6
|
||||
shll2 r6
|
||||
|
||||
add r6, r13
|
||||
add r2, r13
|
||||
|
||||
TEX2D_END_NORET()
|
||||
mov.l @r15+, r10
|
||||
mov.l @r15+, r11
|
||||
mov.l @r15+, r12
|
||||
mov.l @r15+, r13
|
||||
mov.l @r15+, r9
|
||||
rts
|
||||
mov.l @r15+, r8
|
||||
|
||||
/* [Rendering strategy for the P8 RGB565 format]
|
||||
_P8_RGB565A.palette_distance:
|
||||
/* Distance between image pointer and palette array base */
|
||||
.word 260
|
||||
|
||||
This format does not support alpha, lifting the requirement for comparisons,
|
||||
branches and some register logic. The palette is also designed to support
|
||||
signed indices (from -128 to 127). The interwoven setup becomes much more
|
||||
practical as a result. */
|
||||
/* [Rendering strategy for the P8_RGB565 format]
|
||||
|
||||
See P8_RGB565A for format details. Removing the checks for transparency and
|
||||
the jumps simplifies the instruction sequence and allows superior
|
||||
parallelism because all paths are unconditional. This routines achieves
|
||||
3 cycles/pixel asymptotically. */
|
||||
.align 4
|
||||
_P8_RGB565:
|
||||
mov.l r13, @-r15
|
||||
add #-2, r9 /* Input stride compensation for openness */
|
||||
|
||||
mov r7, r13
|
||||
shlr r7
|
||||
/* TODO: Odd case */
|
||||
|
||||
mov.b @r3+, r6
|
||||
add #-4, r5
|
||||
mov.l r12, @-r15
|
||||
movt r6
|
||||
|
||||
shll r6
|
||||
mov.l r10, @-r15
|
||||
shll r13
|
||||
|
||||
mov.w _P8_RGB565.palette_distance, r0
|
||||
add r6, r7
|
||||
|
||||
sub r6, r9
|
||||
|
||||
sub r6, r4
|
||||
|
||||
sub r6, r4
|
||||
|
||||
add r0, r8
|
||||
|
||||
add r5, r13
|
||||
|
||||
add #-4, r5 /* Output offset compensation in the loop */
|
||||
mov r7, r2
|
||||
|
||||
shll2 r2
|
||||
|
||||
add r4, r2
|
||||
nop /* 4-alignment */
|
||||
|
||||
TEX2D_START()
|
||||
2: mov.b @r3+, r2
|
||||
|
||||
mov.b @r3+, r0
|
||||
|
||||
/* Save next pixel for the odd-width case */
|
||||
mov.w @r13, r12
|
||||
|
||||
mov.b @r3+, r10
|
||||
shll r0
|
||||
|
||||
/* 2-interwoven open main loop */
|
||||
2: mov.b @r3+, r6
|
||||
shll r10
|
||||
|
||||
mov.w @(r0,r8), r0
|
||||
|
||||
mov.w r0, @(4,r5)
|
||||
mov r10, r0
|
||||
|
||||
mov.b @r3+, r10
|
||||
add #4, r5
|
||||
|
||||
shll r2
|
||||
mov r6, r0
|
||||
|
||||
/* Stall for r0 */
|
||||
|
||||
mov.w @(r0,r8), r0
|
||||
shll r6
|
||||
|
||||
mov.w r0, @r5
|
||||
mov r2, r0
|
||||
mov.w r0, @(2,r5)
|
||||
3: mov r6, r0
|
||||
|
||||
mov.b @r3+, r6
|
||||
/* Restore last pixel */
|
||||
mov.w r12, @r13
|
||||
add r2, r13
|
||||
|
||||
mov.w @(r0,r8), r0
|
||||
TEX2D_END_NORET()
|
||||
mov.l @r15+, r10
|
||||
mov.l @r15+, r12
|
||||
mov.l @r15+, r13
|
||||
mov.l @r15+, r9
|
||||
rts
|
||||
mov.l @r15+, r8
|
||||
|
||||
mov.w r0, @(2, r5)
|
||||
3: shll r6
|
||||
TEX2D_END()
|
||||
_P8_RGB565.palette_distance:
|
||||
/* Distance between image pointer and palette array base */
|
||||
.word 260
|
||||
|
||||
/* [Rendering strategy for the P4 format] */
|
||||
_P4:
|
||||
|
|
@ -355,3 +420,11 @@ _P4:
|
|||
2:
|
||||
3: nop
|
||||
TEX2D_END()
|
||||
|
||||
/* [Unsupported formats]
|
||||
|
||||
P8 is unsupported, use P8_RGB565 and P8_RGB565A. */
|
||||
_NOP:
|
||||
mov.l @r15+, r9
|
||||
rts
|
||||
mov.l @r15+, r8
|
||||
|
|
|
|||
|
|
@ -17,10 +17,9 @@ void azrp_shader_tex2d_configure(void)
|
|||
|
||||
//---
|
||||
|
||||
/* Profile values from bopti */
|
||||
/* Profile IDs */
|
||||
#define PX_RGB565 0
|
||||
#define PX_RGB565A 1
|
||||
#define PX_P8 2
|
||||
#define PX_P4 3
|
||||
#define PX_P8_RGB565 4
|
||||
#define PX_P8_RGB565A 5
|
||||
|
|
@ -47,11 +46,11 @@ void azrp_subimage(int x, int y, bopti_image_t const *image,
|
|||
int input_multiplier = 1;
|
||||
void const *data = image->data;
|
||||
|
||||
if(image->profile == PX_P8 || image->profile == PX_P8_RGB565) {
|
||||
if(image->profile == PX_P8_RGB565 || image->profile == PX_P8_RGB565A) {
|
||||
input_multiplier = 0;
|
||||
data += 512;
|
||||
data += (image->data[0] * 2) + 2;
|
||||
}
|
||||
if(image->profile == PX_P4) {
|
||||
else if(image->profile == PX_P4) {
|
||||
input_multiplier = -1;
|
||||
data += 32;
|
||||
}
|
||||
|
|
|
|||
Loading…
Reference in New Issue