\r
.if UNALIGNED_DRAWLINEDEST\r
@ unaligned versions of loops\r
+ @ warning: starts drawing 2bytes before dst\r
\r
.fl_RGB555u:\r
- ldr r12, [r1], #4\r
- ldr r7, [r1], #4\r
-\r
- and r4, lr, r12, lsl #1\r
- ldrh r4, [r3, r4]\r
- and r5, lr, r12, lsr #7\r
- ldrh r5, [r3, r5]\r
- strh r4, [r0], #2\r
- b .fl_loopRGB555u_enter\r
+ sub r0, r0, #2 @ initial adjustment\r
+ mov r8, #0\r
\r
.fl_loopRGB555u:\r
ldr r12, [r1], #4\r
ldr r7, [r1], #4\r
\r
- and r4, lr, r12, lsl #1\r
- ldrh r4, [r3, r4]\r
- and r5, lr, r12, lsr #7\r
+ and r6, lr, r12,lsl #1\r
+ ldrh r6, [r3, r6]\r
+ and r5, lr, r12,lsr #7\r
ldrh r5, [r3, r5]\r
+ orr r4, r8, r6, lsl #16\r
\r
- orr r4, r8, r4, lsl #16\r
- str r4, [r0], #4\r
-\r
-.fl_loopRGB555u_enter:\r
- and r6, lr, r12, lsr #15\r
+ and r6, lr, r12,lsr #15\r
ldrh r6, [r3, r6]\r
- and r8, lr, r12, lsr #23\r
+ and r8, lr, r12,lsr #23\r
ldrh r8, [r3, r8]\r
- orr r4, r5, r6, lsl #16\r
+ orr r5, r5, r6, lsl #16\r
\r
- and r5, lr, r7, lsl #1\r
- ldrh r5, [r3, r5]\r
- and r6, lr, r7, lsr #7\r
+ and r6, lr, r7, lsl #1\r
ldrh r6, [r3, r6]\r
- orr r5, r8, r5, lsl #16\r
-\r
- and r12,lr, r7, lsr #15\r
+ and r12,lr, r7, lsr #7\r
ldrh r12,[r3, r12]\r
- and r8, lr, r7, lsr #23\r
+ orr r6, r8, r6, lsl #16\r
+\r
+ and r8, lr, r7, lsr #15\r
ldrh r8, [r3, r8]\r
- orr r6, r6, r12,lsl #16\r
+ and r7, lr, r7, lsr #23\r
\r
subs r2, r2, #1\r
- stmia r0!, {r4,r5,r6}\r
+ orr r12,r12,r8, lsl #16\r
+ ldrh r8, [r3, r7]\r
+\r
+ stmia r0!, {r4,r5,r6,r12}\r
bne .fl_loopRGB555u\r
\r
strh r8, [r0], #2\r
\r
\r
.fl_32scale_RGB555u:\r
+ sub r0, r0, #2 @ initial adjustment\r
+ mov r4, #0\r
\r
@ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007\r
.fl_loop32scale_RGB555u:\r
ldr r12, [r1], #4\r
ldr r7, [r1], #4\r
\r
- and r4, lr, r12,lsl #1\r
- ldrh r4, [r3, r4]\r
+ and r6, lr, r12,lsl #1\r
+ ldrh r6, [r3, r6]\r
and r5, lr, r12,lsr #7\r
ldrh r5, [r3, r5]\r
- and r4, r4, r9, lsl #2\r
-@ orr r4, rx, r4, lsl #16\r
- str r4, [r0], #4 @ pix_d -1, 0\r
-\r
-.fl_loop32scale_RGB555u_enter:\r
- and r4, r9, r4, lsr #2 @ r4=1/4 pix_s 0\r
+ and r6, r6, r9, lsl #2\r
+ orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0\r
\r
and r5, r5, r9, lsl #2\r
- sub r6, r5, r5, lsr #2 @ r6 = 3/4 pix_s 1\r
- add r4, r6, r4, lsr #2 @ r4=(1/4 pix_s 0) + (3/4 pix_s 1)\r
- orr r4, r4, r5, lsl #15\r
+ sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1\r
+ add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1)\r
+ orr r5, r6, r5, lsl #15\r
\r
and r6, lr, r12,lsr #15\r
ldrh r6, [r3, r6]\r
and r12,lr, r12,lsr #23\r
ldrh r12,[r3, r12]\r
and r6, r6, r9, lsl #2\r
- add r4, r4, r6, lsl #15 @ pix_d 1, 2\r
-\r
-@@ TODO...\r
-\r
- mov r5, r5, lsr #1\r
- sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2\r
- orr r5, r5, r6, lsl #16\r
+ add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2\r
\r
- and r6, lr, r7, lsl #1\r
- ldrh r6, [r3, r6]\r
+ and r8, lr, r7, lsl #1\r
+ ldrh r8, [r3, r8]\r
+ and r10,lr, r7, lsr #7\r
+ ldrh r10,[r3, r10]\r
and r12,r12,r9, lsl #2\r
- add r5, r5, r12,lsl #14 @ pix_d 2, 3\r
- and r6, r6, r9, lsl #2\r
- orr r6, r12,r6, lsl #16 @ pix_d 4, 5\r
+ sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2\r
+ add r6, r6, r12,lsr #2\r
+ orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4\r
\r
- and r12,lr, r7, lsr #7\r
+ and r8, r8, r9, lsl #2\r
+ and r10,r10,r9, lsl #2\r
+ sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5\r
+ orr r8, r8, r8, lsl #14\r
+ add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6\r
+ and r12,lr, r7, lsr #15\r
ldrh r12,[r3, r12]\r
- and r10,lr, r7, lsr #15\r
- ldrh r10,[r3, r10]\r
- and r12,r12,r9, lsl #2\r
- sub r8, r12,r12,lsr #2 @ r8 = 3/4 pix_s 1\r
- add r8, r8, r6, lsr #18\r
and r7, lr, r7, lsr #23\r
ldrh r7, [r3, r7]\r
- and r10,r10,r9, lsl #2\r
- orr r8, r8, r10,lsl #15\r
- add r8, r8, r12,lsl #15 @ pix_d 6, 7\r
- sub r10,r10,r10,lsr #2 @ r10= 3/4 pix_s 2\r
+ and r12,r12,r9, lsl #2\r
+ add r10,r10,r12\r
+ mov r10,r10, lsr #1\r
+ sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6\r
+ orr r10,r10,r12,lsl #16\r
and r7, r7, r9, lsl #2\r
- add r10,r10,r7, lsr #2 @ += 1/4 pix_s 3\r
- orr r10,r10,r7, lsl #16 @ pix_d 8, 9\r
+ add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8\r
\r
subs r2, r2, #1\r
\r
stmia r0!, {r4,r5,r6,r8,r10}\r
+ mov r4, r7\r
bne .fl_loop32scale_RGB555u\r
\r
+ strh r4, [r0], #2\r
+\r
ldmfd sp!, {r9,r10}\r
ldmfd sp!, {r4-r8,lr}\r
bx lr\r
notaz.gp2x.de / picodrive.git / blobdiff