sdl, complete overhaul of hardware/software scaling

This commit is contained in:
kub 2021-09-10 18:46:05 +02:00
parent 6651998e9f
commit d5d1778252
21 changed files with 1233 additions and 547 deletions

View file

@ -1653,9 +1653,9 @@ FinalizeLine555:
tst r12, #1
movne r2, #320/8 @ len
bne .fl_no32colRGB555
ldr r4, [r10, #OFS_EST_PicoOpt]
ldr r5, [r10, #OFS_EST_PicoOpt]
mov r2, #256/8
ldr r4, [r4]
ldr r4, [r5]
tst r4, #0x4000
bne .fl_32scale_RGB555
tst r4, #0x0100
@ -1705,15 +1705,167 @@ FinalizeLine555:
.fl_32scale_RGB555:
mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
orr r9, r9, #0x00e7
ldr r5, [r5, #OFS_PicoIn_filter-OFS_PicoIn_opt]
mov r9, #0xf700 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
orr r9, r9, #0x00de
#ifdef UNALIGNED_DRAWLINEDEST
tst r0, #2
bne .fl_32scale_RGB555u
#endif
.fl_loop32scale_RGB555:
ands r5, r5, #0x3
addne pc, pc, r5, lsl #2
b .fl_32scale_nn
b .fl_32scale_nn
b .fl_32scale_snn
b .fl_32scale_bl2
b .fl_32scale_bl4
.fl_32scale_nn:
ldr r12, [r1], #4
ldr r7, [r1], #4
and r4, lr, r12, lsl #1
ldrh r4, [r3, r4]
and r5, lr, r12, lsr #7
ldrh r5, [r3, r5]
and r6, lr, r12, lsr #15
ldrh r6, [r3, r6]
and r10,lr, r12, lsr #23
ldrh r10,[r3, r10]
orr r4, r4, r5, lsl #16
orr r5, r6, r6, lsl #16
and r6, lr, r7, lsl #1
ldrh r6, [r3, r6]
and r8, lr, r7, lsr #7
ldrh r8, [r3, r8]
and r12,lr, r7, lsr #15
ldrh r12,[r3, r12]
and r7, lr, r7, lsr #23
ldrh r7, [r3, r7]
orr r6, r10,r6, lsl #16
orr r8, r8,r12, lsl #16
subs r2, r2, #1
orr r10,r12,r7, lsl #16
stmia r0!, {r4,r5,r6,r8,r10}
bne .fl_32scale_nn
ldmfd sp!, {r4-r10,pc}
.fl_32scale_snn:
ldr r12, [r1], #4
ldr r7, [r1], #4
and r4, lr, r12, lsl #1
ldrh r4, [r3, r4]
and r5, lr, r12, lsr #7
ldrh r5, [r3, r5]
and r6, lr, r12, lsr #15
ldrh r6, [r3, r6]
and r10,lr, r12, lsr #23
ldrh r10,[r3, r10]
and r4, r4, r9
and r5, r5, r9
orr r4, r4, r5, lsl #16
and r6, r6, r9
add r5, r5, r6
mov r5, r5, lsr #1
orr r5, r5, r6, lsl #16
and r6, lr, r7, lsl #1
ldrh r6, [r3, r6]
and r8, lr, r7, lsr #7
ldrh r8, [r3, r8]
and r12,lr, r7, lsr #15
ldrh r12,[r3, r12]
and r7, lr, r7, lsr #23
ldrh r7, [r3, r7]
and r6, r6, r9
and r10,r10,r9
orr r6, r10,r6, lsl #16
and r12,r12,r9
and r7, r7, r9
orr r10,r12,r7, lsl #16
and r8, r8, r9
add r12,r12,r8
mov r12,r12,lsr #1
orr r8, r8,r12, lsl #16
subs r2, r2, #1
stmia r0!, {r4,r5,r6,r8,r10}
bne .fl_32scale_snn
ldmfd sp!, {r4-r10,pc}
.fl_32scale_bl2:
ldr r12, [r1], #4
ldr r7, [r1], #4
and r4, lr, r12, lsl #1
ldrh r4, [r3, r4]
and r5, lr, r12, lsr #7
ldrh r5, [r3, r5]
and r6, lr, r12, lsr #15
ldrh r6, [r3, r6]
and r4, r4, r9
and r5, r5, r9
add r10,r4, r5
mov r10,r10,lsr #1
orr r4, r4, r10,lsl #16 @ px0 | (px0+px1)/2
and r6, r6, r9
add r5, r5, r6
mov r5, r5, lsr #1
orr r5, r5, r6, lsl #16 @ (px1+px2)/2 | px2
and r10,lr, r12, lsr #23
ldrh r10,[r3, r10]
and r8, lr, r7, lsl #1
ldrh r8, [r3, r8]
and r10,r10,r9
and r8, r8, r9
orr r6, r10,r8, lsl #16 @ px3 | px4
and r12,lr, r7, lsr #15
ldrh r12,[r3, r12]
and r10, lr, r7, lsr #23
ldrh r10, [r3, r10]
and r7, lr, r7, lsr #7
ldrh r7, [r3, r7]
and r12,r12,r9
and r10,r10,r9
orr r10,r12,r10, lsl #16 @ px6 | px7
and r7, r7, r9
add r12,r12,r7
add r8, r8, r7
mov r8, r8, lsr #1
mov r12,r12,lsr #1
orr r8, r8,r12, lsl #16 @ (px4+px5)/2 | (px5+px6)/2
subs r2, r2, #1
stmia r0!, {r4,r5,r6,r8,r10}
bne .fl_32scale_bl2
ldmfd sp!, {r4-r10,pc}
.fl_32scale_bl4:
ldr r12, [r1], #4
ldr r7, [r1], #4
@ -1721,16 +1873,21 @@ FinalizeLine555:
ldrh r4, [r3, r4]
and r5, lr, r12,lsr #7
ldrh r5, [r3, r5]
and r4, r4, r9, lsl #2
@ r4 = 1/4px0+3/4px1 : px0
and r4, r4, r9
orr r4, r4, r4, lsl #14 @ r4[31:16] = 1/4 pix_s 0
and r5, r5, r9, lsl #2
and r5, r5, r9
sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1
add r4, r4, r6, lsl #16 @ pix_d 0, 1
and r6, lr, r12,lsr #15
ldrh r6, [r3, r6]
and r12,lr, r12,lsr #23
ldrh r12,[r3, r12]
and r6, r6, r9, lsl #2
@ r5 = 3/4px2+1/4px3 : (px1+px2)/2
and r6, r6, r9
add r5, r5, r6
mov r5, r5, lsr #1
sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
@ -1738,32 +1895,38 @@ FinalizeLine555:
and r6, lr, r7, lsl #1
ldrh r6, [r3, r6]
and r12,r12,r9, lsl #2
and r12,r12,r9
add r5, r5, r12,lsl #14 @ pix_d 2, 3
and r6, r6, r9, lsl #2
@ r6 = px4 : px3
and r6, r6, r9
orr r6, r12,r6, lsl #16 @ pix_d 4, 5
@ r8 = (px5+px6)/2 : 1/4px4+3/4px5
and r12,lr, r7, lsr #7
ldrh r12,[r3, r12]
and r10,lr, r7, lsr #15
ldrh r10,[r3, r10]
and r12,r12,r9, lsl #2
and r12,r12,r9
sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1
add r8, r8, r6, lsr #18
and r7, lr, r7, lsr #23
ldrh r7, [r3, r7]
and r10,r10,r9, lsl #2
and r10,r10,r9
orr r8, r8, r10,lsl #15
add r8, r8, r12,lsl #15 @ pix_d 6, 7
@ r10 = px7 : 3/4px6+1/4px7
sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2
and r7, r7, r9, lsl #2
and r7, r7, r9
add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3
orr r10,r10,r7, lsl #16 @ pix_d 8, 9
subs r2, r2, #1
stmia r0!, {r4,r5,r6,r8,r10}
bne .fl_loop32scale_RGB555
bne .fl_32scale_bl4
ldmfd sp!, {r4-r10,pc}
@ -1826,10 +1989,10 @@ FinalizeLine555:
ldrh r6, [r3, r6]
and r5, lr, r12,lsr #7
ldrh r5, [r3, r5]
and r6, r6, r9, lsl #2
and r6, r6, r9
orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0
and r5, r5, r9, lsl #2
and r5, r5, r9
sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1
add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)
orr r5, r6, r5, lsl #15
@ -1838,20 +2001,20 @@ FinalizeLine555:
ldrh r6, [r3, r6]
and r12,lr, r12,lsr #23
ldrh r12,[r3, r12]
and r6, r6, r9, lsl #2
and r6, r6, r9
add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2
and r8, lr, r7, lsl #1
ldrh r8, [r3, r8]
and r10,lr, r7, lsr #7
ldrh r10,[r3, r10]
and r12,r12,r9, lsl #2
and r12,r12,r9
sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2
add r6, r6, r12,lsr #2
orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4
and r8, r8, r9, lsl #2
and r10,r10,r9, lsl #2
and r8, r8, r9
and r10,r10,r9
sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5
orr r8, r8, r8, lsl #14
add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6
@ -1859,12 +2022,12 @@ FinalizeLine555:
ldrh r12,[r3, r12]
and r7, lr, r7, lsr #23
ldrh r7, [r3, r7]
and r12,r12,r9, lsl #2
and r12,r12,r9
add r10,r10,r12
mov r10,r10, lsr #1
sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6
orr r10,r10,r12,lsl #16
and r7, r7, r9, lsl #2
and r7, r7, r9
add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8
subs r2, r2, #1