[picodrive.git] / pico / 32x / draw_arm.s

@ vim:filetype=armasm

.extern Pico32x
.extern PicoDraw2FB
.extern HighPal

.equiv P32XV_PRI,  (1<< 7)

.bss
.align 2
.global Pico32xNativePal
Pico32xNativePal:
    .word 0

.text
.align 2


.macro call_scan_prep cond
.if \cond
    ldr     r4, =PicoScan32xBegin
    ldr     r5, =PicoScan32xEnd
    ldr     r6, =DrawLineDest
    ldr     r4, [r4]
    ldr     r5, [r5]
    stmfd   sp!, {r4,r5,r6}
.endif
.endm

.macro call_scan_fin_ge cond
.if \cond
    addge   sp, sp, #4*3
.endif
.endm

.macro call_scan_begin cond
.if \cond
    stmfd   sp!, {r1-r3}
    and     r0, r2, #0xff
    add     r0, r0, r4
    mov     lr, pc
    ldr     pc, [sp, #(3+0)*4]
    ldr     r0, [sp, #(3+2)*4] @ &DrawLineDest
    ldmfd   sp!, {r1-r3}
    ldr     r0, [r0]
.endif
.endm

.macro call_scan_end cond
.if \cond
    stmfd   sp!, {r0-r3}
    and     r0, r2, #0xff
    add     r0, r0, r4
    mov     lr, pc
    ldr     pc, [sp, #(4+1)*4]
    ldmfd   sp!, {r0-r3}
.endif
.endm

@ direct color
@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_dc name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r10,=Pico32x
    ldr     r11,=PicoDraw2FB
    ldr     r10,[r10, #0x40] @ Pico32x.vdp_regs[0]
    ldr     r11,[r11]
    ldr     r9, =HighPal     @ palmd
    add     r11,r11,#(328*8) @ r11 = pmd: md data
    tst     r10,#P32XV_PRI
    moveq   r10,#0
    movne   r10,#0x8000      @ r10 = inv_bit
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    sub     r11,r11,#1       @ adjust for prev read
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]

2: @ loop_inner:
    ldrb    r7, [r11], #1    @ MD pixel
    subs    r6, r6, #1
    blt     0b @ loop_outer
    ldrh    r8, [r5], #2     @ 32x pixel
    cmp     r7, r3           @ MD has bg pixel?
    beq     3f @ draw32x
    eor     r12,r8, r10
    ands    r12,r12,#0x8000  @ !((t ^ inv) & 0x8000)
.if \do_md
    mov     r7, r7, lsl #1
    ldreqh  r12,[r9, r7]
    streqh  r12,[r0], #2     @ *dst++ = palmd[*pmd]
.endif
    beq     2b @ loop_inner

3: @ draw32x:
    and     r12,r8, #0x03e0
    mov     r8, r8, lsl #11
    orr     r8, r8, r8, lsr #(10+11)
    orr     r8, r8, r12,lsl #1
    bic     r8, r8, #0x0020  @ kill prio bit
    strh    r8, [r0], #2     @ *dst++ = bgr2rgb(*p32x++)
    b       2b @ loop_inner
.endm


@ packed pixel
.macro do_pixel_pp do_md
    ldrb    r7, [r11], #1    @ MD pixel
    eor     r12,r5, #1
    ldrb    r8, [r12]        @ palette index
    cmp     r7, r3           @ MD has bg pixel?
    mov     r12,r8,lsl #1
    ldrh    r8, [r10,r12]    @ t = 32x pixel
    mov     r7, r7, lsl #1
    add     r5, r5, #1
    eorne   r12,r8, #0x20
    tstne   r12, #0x20
.if \do_md
    ldrneh  r8, [r9, r7]     @ t = palmd[*pmd]
    subs    r6, r6, #1
    strh    r8, [r0], #2     @ *dst++ = t
.else
    streqh  r8, [r0], #2
    addne   r0, r0, #2
    subs    r6, r6, #1
.endif
.endm

@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_pp name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r11,=PicoDraw2FB
    ldr     r10,=Pico32xNativePal
    ldr     r11,[r11]
    ldr     r10,[r10]
    ldr     r9, =HighPal     @ palmd
    add     r11,r11,#(328*8) @ r11 = pmd: md data
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
    and     r12,r2, #0x100      @ shift
    add     r5, r5, r12,lsr #8

2: @ loop_inner:
    do_pixel_pp \do_md
    do_pixel_pp \do_md
    bgt     2b @ loop_inner
    b       0b @ loop_outer
.endm


@ run length
@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_rl name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r11,=PicoDraw2FB
    ldr     r10,=Pico32xNativePal
    ldr     r11,[r11]
    ldr     r10,[r10]
    ldr     r9, =HighPal     @ palmd
    add     r11,r11,#(328*8) @ r11 = pmd: md data
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    sub     r11,r11,#1       @ adjust for prev read
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]

2: @ loop_inner:
    ldrh    r8, [r5], #2     @ control word
    and     r12,r8, #0xff
    mov     r12,r12,lsl #1
    ldrh    lr, [r10,r12]    @ t = 32x pixel
    eor     lr, lr, #0x20

3: @ loop_innermost:
    ldrb    r7, [r11], #1    @ MD pixel
    subs    r6, r6, #1
    blt     0b @ loop_outer
    cmp     r7, r3           @ MD has bg pixel?
    mov     r7, r7, lsl #1
    tstne   lr, #0x20
.if \do_md
    ldrneh  r12,[r9, r7]     @ t = palmd[*pmd]
    streqh  lr, [r0], #2
    strneh  r12,[r0], #2     @ *dst++ = t
.else
    streqh  lr, [r0]
    add     r0, r0, #2
.endif
    subs    r8, r8, #0x100
    bge     3b @ loop_innermost
    b       2b @ loop_inner
.endm


make_do_loop_dc do_loop_dc,         0, 0
make_do_loop_dc do_loop_dc_md,      0, 1
make_do_loop_dc do_loop_dc_scan,    1, 0
make_do_loop_dc do_loop_dc_scan_md, 1, 1

make_do_loop_pp do_loop_pp,         0, 0
make_do_loop_pp do_loop_pp_md,      0, 1
make_do_loop_pp do_loop_pp_scan,    1, 0
make_do_loop_pp do_loop_pp_scan_md, 1, 1

make_do_loop_rl do_loop_rl,         0, 0
make_do_loop_rl do_loop_rl_md,      0, 1
make_do_loop_rl do_loop_rl_scan,    1, 0
make_do_loop_rl do_loop_rl_scan_md, 1, 1
Commit	Line	Data
5a681086	1	@ vim:filetype=armasm
	2
	3	.extern Pico32x
	4	.extern PicoDraw2FB
	5	.extern HighPal
	6
	7	.equiv P32XV_PRI, (1<< 7)
	8
	9	.bss
	10	.align 2
	11	.global Pico32xNativePal
	12	Pico32xNativePal:
	13	.word 0
	14
	15	.text
	16	.align 2
	17
	18
	19	.macro call_scan_prep cond
	20	.if \cond
	21	ldr r4, =PicoScan32xBegin
	22	ldr r5, =PicoScan32xEnd
	23	ldr r6, =DrawLineDest
	24	ldr r4, [r4]
	25	ldr r5, [r5]
	26	stmfd sp!, {r4,r5,r6}
	27	.endif
	28	.endm
	29
	30	.macro call_scan_fin_ge cond
	31	.if \cond
	32	addge sp, sp, #4*3
	33	.endif
	34	.endm
	35
	36	.macro call_scan_begin cond
	37	.if \cond
	38	stmfd sp!, {r1-r3}
	39	and r0, r2, #0xff
	40	add r0, r0, r4
	41	mov lr, pc
	42	ldr pc, [sp, #(3+0)*4]
	43	ldr r0, [sp, #(3+2)*4] @ &DrawLineDest
	44	ldmfd sp!, {r1-r3}
	45	ldr r0, [r0]
	46	.endif
	47	.endm
	48
	49	.macro call_scan_end cond
	50	.if \cond
	51	stmfd sp!, {r0-r3}
	52	and r0, r2, #0xff
	53	add r0, r0, r4
	54	mov lr, pc
	55	ldr pc, [sp, #(4+1)*4]
	56	ldmfd sp!, {r0-r3}
	57	.endif
	58	.endm
	59
	60	@ direct color
e51e5983	61	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
5a681086	62	.macro make_do_loop_dc name call_scan do_md
	63	.global \name
	64	\name:
	65	stmfd sp!, {r4-r11,lr}
	66
	67	ldr r10,=Pico32x
	68	ldr r11,=PicoDraw2FB
	69	ldr r10,[r10, #0x40] @ Pico32x.vdp_regs[0]
	70	ldr r11,[r11]
	71	ldr r9, =HighPal @ palmd
	72	add r11,r11,#(328*8) @ r11 = pmd: md data
	73	tst r10,#P32XV_PRI
	74	moveq r10,#0
	75	movne r10,#0x8000 @ r10 = inv_bit
	76	call_scan_prep \call_scan
	77
	78	mov r4, #0 @ line
	79	b 1f @ loop_outer_entry
	80
	81	0: @ loop_outer:
	82	call_scan_end \call_scan
	83	add r4, r4, #1
	84	sub r11,r11,#1 @ adjust for prev read
	85	cmp r4, r2, lsr #16
	86	call_scan_fin_ge \call_scan
	87	ldmgefd sp!, {r4-r11,pc}
	88
	89	1: @ loop_outer_entry:
	90	call_scan_begin \call_scan
	91	mov r12,r4, lsl #1
	92	ldrh r12,[r1, r12]
	93	add r11,r11,#8
	94	mov r6, #320
	95	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
	96
	97	2: @ loop_inner:
	98	ldrb r7, [r11], #1 @ MD pixel
	99	subs r6, r6, #1
	100	blt 0b @ loop_outer
	101	ldrh r8, [r5], #2 @ 32x pixel
	102	cmp r7, r3 @ MD has bg pixel?
	103	beq 3f @ draw32x
	104	eor r12,r8, r10
	105	ands r12,r12,#0x8000 @ !((t ^ inv) & 0x8000)
	106	.if \do_md
	107	mov r7, r7, lsl #1
	108	ldreqh r12,[r9, r7]
	109	streqh r12,[r0], #2 @ dst++ = palmd[pmd]
	110	.endif
	111	beq 2b @ loop_inner
	112
	113	3: @ draw32x:
	114	and r12,r8, #0x03e0
	115	mov r8, r8, lsl #11
	116	orr r8, r8, r8, lsr #(10+11)
	117	orr r8, r8, r12,lsl #1
	118	bic r8, r8, #0x0020 @ kill prio bit
	119	strh r8, [r0], #2 @ dst++ = bgr2rgb(p32x++)
	120	b 2b @ loop_inner
	121	.endm
	122
	123
	124	@ packed pixel
	125	.macro do_pixel_pp do_md
126	ldrb r7, [r11], #1 @ MD pixel
127	eor r12,r5, #1
128	ldrb r8, [r12] @ palette index
129	cmp r7, r3 @ MD has bg pixel?
130	mov r12,r8,lsl #1
131	ldrh r8, [r10,r12] @ t = 32x pixel
132	mov r7, r7, lsl #1
133	add r5, r5, #1
134	eorne r12,r8, #0x20
135	tstne r12, #0x20
136	.if \do_md
137	ldrneh r8, [r9, r7] @ t = palmd[*pmd]
138	subs r6, r6, #1
139	strh r8, [r0], #2 @ *dst++ = t
140	.else
141	streqh r8, [r0], #2
142	addne r0, r0, #2
143	subs r6, r6, #1
144	.endif
145	.endm
146
e51e5983	147	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
5a681086	148	.macro make_do_loop_pp name call_scan do_md
	149	.global \name
	150	\name:
	151	stmfd sp!, {r4-r11,lr}
	152
	153	ldr r11,=PicoDraw2FB
	154	ldr r10,=Pico32xNativePal
	155	ldr r11,[r11]
	156	ldr r10,[r10]
	157	ldr r9, =HighPal @ palmd
	158	add r11,r11,#(328*8) @ r11 = pmd: md data
	159	call_scan_prep \call_scan
	160
	161	mov r4, #0 @ line
	162	b 1f @ loop_outer_entry
	163
	164	0: @ loop_outer:
	165	call_scan_end \call_scan
	166	add r4, r4, #1
	167	cmp r4, r2, lsr #16
	168	call_scan_fin_ge \call_scan
	169	ldmgefd sp!, {r4-r11,pc}
	170
	171	1: @ loop_outer_entry:
	172	call_scan_begin \call_scan
	173	mov r12,r4, lsl #1
	174	ldrh r12,[r1, r12]
	175	add r11,r11,#8
	176	mov r6, #320
	177	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
e51e5983	178	and r12,r2, #0x100 @ shift
e51e5983	179	add r5, r5, r12,lsr #8
5a681086	180
	181	2: @ loop_inner:
	182	do_pixel_pp \do_md
	183	do_pixel_pp \do_md
	184	bgt 2b @ loop_inner
	185	b 0b @ loop_outer
	186	.endm
	187
	188
	189	@ run length
e51e5983	190	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
5a681086	191	.macro make_do_loop_rl name call_scan do_md
	192	.global \name
	193	\name:
	194	stmfd sp!, {r4-r11,lr}
	195
	196	ldr r11,=PicoDraw2FB
	197	ldr r10,=Pico32xNativePal
	198	ldr r11,[r11]
	199	ldr r10,[r10]
	200	ldr r9, =HighPal @ palmd
	201	add r11,r11,#(328*8) @ r11 = pmd: md data
	202	call_scan_prep \call_scan
	203
	204	mov r4, #0 @ line
	205	b 1f @ loop_outer_entry
	206
	207	0: @ loop_outer:
	208	call_scan_end \call_scan
	209	add r4, r4, #1
	210	sub r11,r11,#1 @ adjust for prev read
	211	cmp r4, r2, lsr #16
	212	call_scan_fin_ge \call_scan
	213	ldmgefd sp!, {r4-r11,pc}
	214
	215	1: @ loop_outer_entry:
	216	call_scan_begin \call_scan
	217	mov r12,r4, lsl #1
	218	ldrh r12,[r1, r12]
	219	add r11,r11,#8
	220	mov r6, #320
	221	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
	222
	223	2: @ loop_inner:
	224	ldrh r8, [r5], #2 @ control word
	225	and r12,r8, #0xff
	226	mov r12,r12,lsl #1
	227	ldrh lr, [r10,r12] @ t = 32x pixel
	228	eor lr, lr, #0x20
	229
	230	3: @ loop_innermost:
	231	ldrb r7, [r11], #1 @ MD pixel
	232	subs r6, r6, #1
	233	blt 0b @ loop_outer
	234	cmp r7, r3 @ MD has bg pixel?
	235	mov r7, r7, lsl #1
	236	tstne lr, #0x20
	237	.if \do_md
	238	ldrneh r12,[r9, r7] @ t = palmd[*pmd]
	239	streqh lr, [r0], #2
	240	strneh r12,[r0], #2 @ *dst++ = t
	241	.else
	242	streqh lr, [r0]
	243	add r0, r0, #2
	244	.endif
	245	subs r8, r8, #0x100
	246	bge 3b @ loop_innermost
	247	b 2b @ loop_inner
	248	.endm
	249
	250
	251	make_do_loop_dc do_loop_dc, 0, 0
	252	make_do_loop_dc do_loop_dc_md, 0, 1
	253	make_do_loop_dc do_loop_dc_scan, 1, 0
	254	make_do_loop_dc do_loop_dc_scan_md, 1, 1
255
256	make_do_loop_pp do_loop_pp, 0, 0
257	make_do_loop_pp do_loop_pp_md, 0, 1
258	make_do_loop_pp do_loop_pp_scan, 1, 0
259	make_do_loop_pp do_loop_pp_scan_md, 1, 1
260
261	make_do_loop_rl do_loop_rl, 0, 0
262	make_do_loop_rl do_loop_rl_md, 0, 1
263	make_do_loop_rl do_loop_rl_scan, 1, 0
264	make_do_loop_rl do_loop_rl_scan_md, 1, 1
265