[sdl_omap.git] / src / hermes / x86p_16.asm

;
; x86 format converters for HERMES
; Copyright (c) 1998 Glenn Fielder (gaffer@gaffer.org)
; This source code is licensed under the GNU LGPL
; 
; Please refer to the file COPYING.LIB contained in the distribution for
; licensing conditions		
; 
; Routines adjusted for Hermes by Christian Nentwich (brn@eleet.mcb.at)
; Used with permission.
; 

BITS 32

%include "common.inc"

SDL_FUNC _ConvertX86p16_16BGR565
SDL_FUNC _ConvertX86p16_16RGB555
SDL_FUNC _ConvertX86p16_16BGR555
SDL_FUNC _ConvertX86p16_8RGB332

EXTERN _ConvertX86

SECTION .text

_ConvertX86p16_16BGR565:

    ; check short
    cmp ecx,BYTE 16
    ja .L3


.L1 ; short loop
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    and ebx,11111100000b
    shl edx,11
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx
    jnz .L1
.L2
    retn

.L3 ; head
    mov eax,edi
    and eax,BYTE 11b
    jz .L4
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    and ebx,11111100000b
    shl edx,11
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx

.L4 ; save count
    push ecx

    ; unroll twice
    shr ecx,1
    
    ; point arrays to end
    lea esi,[esi+ecx*4]
    lea edi,[edi+ecx*4]

    ; negative counter 
    neg ecx
    jmp SHORT .L6
                              
.L5     mov [edi+ecx*4-4],eax
.L6     mov eax,[esi+ecx*4]

        mov ebx,[esi+ecx*4]
        and eax,07E007E0h         

        mov edx,[esi+ecx*4]
        and ebx,0F800F800h

        shr ebx,11
        and edx,001F001Fh

        shl edx,11
        add eax,ebx

        add eax,edx                 
        inc ecx

        jnz .L5                 
         
    mov [edi+ecx*4-4],eax

    ; tail
    pop ecx
    and ecx,BYTE 1
    jz .L7
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    and ebx,11111100000b
    shl edx,11
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2

.L7
    retn


_ConvertX86p16_16RGB555:

    ; check short
    cmp ecx,BYTE 32
    ja .L3


.L1 ; short loop
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    shr ebx,1
    and ebx,     0111111111100000b
    and eax,BYTE 0000000000011111b
    add eax,ebx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx
    jnz .L1
.L2
    retn

.L3 ; head
    mov eax,edi
    and eax,BYTE 11b
    jz .L4
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    shr ebx,1
    and ebx,     0111111111100000b
    and eax,BYTE 0000000000011111b
    add eax,ebx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx

.L4 ; save ebp
    push ebp

    ; save count
    push ecx

    ; unroll four times
    shr ecx,2
    
    ; point arrays to end
    lea esi,[esi+ecx*8]
    lea edi,[edi+ecx*8]

    ; negative counter 
    xor ebp,ebp
    sub ebp,ecx

.L5     mov eax,[esi+ebp*8]        ; agi?
        mov ecx,[esi+ebp*8+4]
       
        mov ebx,eax
        mov edx,ecx

        and eax,0FFC0FFC0h
        and ecx,0FFC0FFC0h

        shr eax,1
        and ebx,001F001Fh

        shr ecx,1
        and edx,001F001Fh

        add eax,ebx
        add ecx,edx

        mov [edi+ebp*8],eax
        mov [edi+ebp*8+4],ecx

        inc ebp
        jnz .L5                 

    ; tail
    pop ecx
.L6 and ecx,BYTE 11b
    jz .L7
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    shr ebx,1
    and ebx,     0111111111100000b
    and eax,BYTE 0000000000011111b
    add eax,ebx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx
    jmp SHORT .L6

.L7 pop ebp
    retn


_ConvertX86p16_16BGR555:

    ; check short
    cmp ecx,BYTE 16
    ja .L3

	
.L1 ; short loop
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    shr ebx,1
    and ebx,1111100000b
    shl edx,10
    and edx,0111110000000000b
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx
    jnz .L1
.L2
    retn

.L3 ; head
    mov eax,edi
    and eax,BYTE 11b
    jz .L4
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    shr ebx,1
    and ebx,1111100000b
    shl edx,10
    and edx,0111110000000000b
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2
    dec ecx

.L4 ; save count
    push ecx

    ; unroll twice
    shr ecx,1
    
    ; point arrays to end
    lea esi,[esi+ecx*4]
    lea edi,[edi+ecx*4]

    ; negative counter 
    neg ecx
    jmp SHORT .L6
                              
.L5     mov [edi+ecx*4-4],eax
.L6     mov eax,[esi+ecx*4]

        shr eax,1
        mov ebx,[esi+ecx*4]
        
        and eax,03E003E0h         
        mov edx,[esi+ecx*4]

        and ebx,0F800F800h

        shr ebx,11
        and edx,001F001Fh

        shl edx,10
        add eax,ebx

        add eax,edx                 
        inc ecx

        jnz .L5                 
         
    mov [edi+ecx*4-4],eax

    ; tail
    pop ecx
    and ecx,BYTE 1
    jz .L7
    mov al,[esi]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    shr eax,11
    and eax,BYTE 11111b
    shr ebx,1
    and ebx,1111100000b
    shl edx,10
    and edx,0111110000000000b
    add eax,ebx
    add eax,edx
    mov [edi],al
    mov [edi+1],ah
    add esi,BYTE 2
    add edi,BYTE 2

.L7
    retn


_ConvertX86p16_8RGB332:

    ; check short
    cmp ecx,BYTE 16
    ja .L3


.L1 ; short loop
    mov al,[esi+0]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    and eax,BYTE 11000b         ; blue
    shr eax,3
    and ebx,11100000000b        ; green
    shr ebx,6
    and edx,1110000000000000b   ; red
    shr edx,8
    add eax,ebx
    add eax,edx
    mov [edi],al
    add esi,BYTE 2
    inc edi
    dec ecx
    jnz .L1
.L2
    retn

.L3 mov eax,edi
    and eax,BYTE 11b
    jz .L4
    mov al,[esi+0]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    and eax,BYTE 11000b         ; blue
    shr eax,3
    and ebx,11100000000b        ; green
    shr ebx,6
    and edx,1110000000000000b   ; red
    shr edx,8
    add eax,ebx
    add eax,edx
    mov [edi],al
    add esi,BYTE 2
    inc edi
    dec ecx
    jmp SHORT .L3

.L4 ; save ebp
    push ebp

    ; save count
    push ecx

    ; unroll 4 times
    shr ecx,2

    ; prestep
    mov dl,[esi+0]
    mov bl,[esi+1]
    mov dh,[esi+2]
        
.L5     shl edx,16
        mov bh,[esi+3]
        
        shl ebx,16
        mov dl,[esi+4]

        mov dh,[esi+6]
        mov bl,[esi+5]

        and edx,00011000000110000001100000011000b
        mov bh,[esi+7]

        ror edx,16+3
        mov eax,ebx                                     ; setup eax for reds

        and ebx,00000111000001110000011100000111b
        and eax,11100000111000001110000011100000b       ; reds

        ror ebx,16-2
        add esi,BYTE 8

        ror eax,16
        add edi,BYTE 4

        add eax,ebx
        mov bl,[esi+1]                                  ; greens

        add eax,edx
        mov dl,[esi+0]                                  ; blues

        mov [edi-4],eax
        mov dh,[esi+2]

        dec ecx
        jnz .L5                 
    
    ; check tail
    pop ecx
    and ecx,BYTE 11b
    jz .L7

.L6 ; tail
    mov al,[esi+0]
    mov ah,[esi+1]
    mov ebx,eax
    mov edx,eax
    and eax,BYTE 11000b         ; blue
    shr eax,3
    and ebx,11100000000b        ; green
    shr ebx,6
    and edx,1110000000000000b   ; red
    shr edx,8
    add eax,ebx
    add eax,edx
    mov [edi],al
    add esi,BYTE 2
    inc edi
    dec ecx
    jnz .L6

.L7 pop ebp
    retn

%ifidn __OUTPUT_FORMAT__,elf
section .note.GNU-stack noalloc noexec nowrite progbits
%endif
Commit	Line	Data
e14743d1	1	;
	2	; x86 format converters for HERMES
	3	; Copyright (c) 1998 Glenn Fielder (gaffer@gaffer.org)
	4	; This source code is licensed under the GNU LGPL
	5	;
	6	; Please refer to the file COPYING.LIB contained in the distribution for
	7	; licensing conditions
	8	;
	9	; Routines adjusted for Hermes by Christian Nentwich (brn@eleet.mcb.at)
	10	; Used with permission.
	11	;
	12
	13	BITS 32
	14
	15	%include "common.inc"
	16
	17	SDL_FUNC _ConvertX86p16_16BGR565
	18	SDL_FUNC _ConvertX86p16_16RGB555
	19	SDL_FUNC _ConvertX86p16_16BGR555
	20	SDL_FUNC _ConvertX86p16_8RGB332
	21
	22	EXTERN _ConvertX86
	23
	24	SECTION .text
	25
	26	_ConvertX86p16_16BGR565:
	27
	28	; check short
	29	cmp ecx,BYTE 16
	30	ja .L3
	31
	32
	33	.L1 ; short loop
	34	mov al,[esi]
	35	mov ah,[esi+1]
	36	mov ebx,eax
	37	mov edx,eax
	38	shr eax,11
	39	and eax,BYTE 11111b
	40	and ebx,11111100000b
	41	shl edx,11
	42	add eax,ebx
	43	add eax,edx
	44	mov [edi],al
	45	mov [edi+1],ah
	46	add esi,BYTE 2
	47	add edi,BYTE 2
	48	dec ecx
	49	jnz .L1
	50	.L2
	51	retn
	52
	53	.L3 ; head
	54	mov eax,edi
	55	and eax,BYTE 11b
	56	jz .L4
	57	mov al,[esi]
	58	mov ah,[esi+1]
	59	mov ebx,eax
	60	mov edx,eax
	61	shr eax,11
	62	and eax,BYTE 11111b
	63	and ebx,11111100000b
	64	shl edx,11
65	add eax,ebx
66	add eax,edx
67	mov [edi],al
68	mov [edi+1],ah
69	add esi,BYTE 2
70	add edi,BYTE 2
71	dec ecx
72
73	.L4 ; save count
74	push ecx
75
76	; unroll twice
77	shr ecx,1
78
79	; point arrays to end
80	lea esi,[esi+ecx*4]
81	lea edi,[edi+ecx*4]
82
83	; negative counter
84	neg ecx
85	jmp SHORT .L6
86
87	.L5 mov [edi+ecx*4-4],eax
88	.L6 mov eax,[esi+ecx*4]
89
90	mov ebx,[esi+ecx*4]
91	and eax,07E007E0h
92
93	mov edx,[esi+ecx*4]
94	and ebx,0F800F800h
95
96	shr ebx,11
97	and edx,001F001Fh
98
99	shl edx,11
100	add eax,ebx
101
102	add eax,edx
103	inc ecx
104
105	jnz .L5
106
107	mov [edi+ecx*4-4],eax
108
109	; tail
110	pop ecx
111	and ecx,BYTE 1
112	jz .L7
113	mov al,[esi]
114	mov ah,[esi+1]
115	mov ebx,eax
116	mov edx,eax
117	shr eax,11
118	and eax,BYTE 11111b
119	and ebx,11111100000b
120	shl edx,11
121	add eax,ebx
122	add eax,edx
123	mov [edi],al
124	mov [edi+1],ah
125	add esi,BYTE 2
126	add edi,BYTE 2
127
128	.L7
129	retn
130
131
132
133
134
135
136	_ConvertX86p16_16RGB555:
137
138	; check short
139	cmp ecx,BYTE 32
140	ja .L3
141
142
143	.L1 ; short loop
144	mov al,[esi]
145	mov ah,[esi+1]
146	mov ebx,eax
147	shr ebx,1
148	and ebx, 0111111111100000b
149	and eax,BYTE 0000000000011111b
150	add eax,ebx
151	mov [edi],al
152	mov [edi+1],ah
153	add esi,BYTE 2
154	add edi,BYTE 2
155	dec ecx
156	jnz .L1
157	.L2
158	retn
159
160	.L3 ; head
161	mov eax,edi
162	and eax,BYTE 11b
163	jz .L4
164	mov al,[esi]
165	mov ah,[esi+1]
166	mov ebx,eax
167	shr ebx,1
168	and ebx, 0111111111100000b
169	and eax,BYTE 0000000000011111b
170	add eax,ebx
171	mov [edi],al
172	mov [edi+1],ah
173	add esi,BYTE 2
174	add edi,BYTE 2
175	dec ecx
176
177	.L4 ; save ebp
178	push ebp
179
180	; save count
181	push ecx
182
183	; unroll four times
184	shr ecx,2
185
186	; point arrays to end
187	lea esi,[esi+ecx*8]
188	lea edi,[edi+ecx*8]
189
190	; negative counter
191	xor ebp,ebp
192	sub ebp,ecx
193
194	.L5 mov eax,[esi+ebp*8] ; agi?
195	mov ecx,[esi+ebp*8+4]
196
197	mov ebx,eax
198	mov edx,ecx
199
200	and eax,0FFC0FFC0h
201	and ecx,0FFC0FFC0h
202
203	shr eax,1
204	and ebx,001F001Fh
205
206	shr ecx,1
207	and edx,001F001Fh
208
209	add eax,ebx
210	add ecx,edx
211
212	mov [edi+ebp*8],eax
213	mov [edi+ebp*8+4],ecx
214
215	inc ebp
216	jnz .L5
217
218	; tail
219	pop ecx
220	.L6 and ecx,BYTE 11b
221	jz .L7
222	mov al,[esi]
223	mov ah,[esi+1]
224	mov ebx,eax
225	shr ebx,1
226	and ebx, 0111111111100000b
227	and eax,BYTE 0000000000011111b
228	add eax,ebx
229	mov [edi],al
230	mov [edi+1],ah
231	add esi,BYTE 2
232	add edi,BYTE 2
233	dec ecx
234	jmp SHORT .L6
235
236	.L7 pop ebp
237	retn
238
239
240
241
242
243
244	_ConvertX86p16_16BGR555:
245
246	; check short
247	cmp ecx,BYTE 16
248	ja .L3
249
250
251	.L1 ; short loop
252	mov al,[esi]
253	mov ah,[esi+1]
254	mov ebx,eax
255	mov edx,eax
256	shr eax,11
257	and eax,BYTE 11111b
258	shr ebx,1
259	and ebx,1111100000b
260	shl edx,10
261	and edx,0111110000000000b
262	add eax,ebx
263	add eax,edx
264	mov [edi],al
265	mov [edi+1],ah
266	add esi,BYTE 2
267	add edi,BYTE 2
268	dec ecx
269	jnz .L1
270	.L2
271	retn
272
273	.L3 ; head
274	mov eax,edi
275	and eax,BYTE 11b
276	jz .L4
277	mov al,[esi]
278	mov ah,[esi+1]
279	mov ebx,eax
280	mov edx,eax
281	shr eax,11
282	and eax,BYTE 11111b
283	shr ebx,1
284	and ebx,1111100000b
285	shl edx,10
286	and edx,0111110000000000b
287	add eax,ebx
288	add eax,edx
289	mov [edi],al
290	mov [edi+1],ah
291	add esi,BYTE 2
292	add edi,BYTE 2
293	dec ecx
294
295	.L4 ; save count
296	push ecx
297
298	; unroll twice
299	shr ecx,1
300
301	; point arrays to end
302	lea esi,[esi+ecx*4]
303	lea edi,[edi+ecx*4]
304
305	; negative counter
306	neg ecx
307	jmp SHORT .L6
308
309	.L5 mov [edi+ecx*4-4],eax
310	.L6 mov eax,[esi+ecx*4]
311
312	shr eax,1
313	mov ebx,[esi+ecx*4]
314
315	and eax,03E003E0h
316	mov edx,[esi+ecx*4]
317
318	and ebx,0F800F800h
319
320	shr ebx,11
321	and edx,001F001Fh
322
323	shl edx,10
324	add eax,ebx
325
326	add eax,edx
327	inc ecx
328
329	jnz .L5
330
331	mov [edi+ecx*4-4],eax
332
333	; tail
334	pop ecx
335	and ecx,BYTE 1
336	jz .L7
337	mov al,[esi]
338	mov ah,[esi+1]
339	mov ebx,eax
340	mov edx,eax
341	shr eax,11
342	and eax,BYTE 11111b
343	shr ebx,1
344	and ebx,1111100000b
345	shl edx,10
346	and edx,0111110000000000b
347	add eax,ebx
348	add eax,edx
349	mov [edi],al
350	mov [edi+1],ah
351	add esi,BYTE 2
352	add edi,BYTE 2
353
354	.L7
355	retn
356
357
358
359
360
361
362	_ConvertX86p16_8RGB332:
363
364	; check short
365	cmp ecx,BYTE 16
366	ja .L3
367
368
369	.L1 ; short loop
370	mov al,[esi+0]
371	mov ah,[esi+1]
372	mov ebx,eax
373	mov edx,eax
374	and eax,BYTE 11000b ; blue
375	shr eax,3
376	and ebx,11100000000b ; green
377	shr ebx,6
378	and edx,1110000000000000b ; red
379	shr edx,8
380	add eax,ebx
381	add eax,edx
382	mov [edi],al
383	add esi,BYTE 2
384	inc edi
385	dec ecx
386	jnz .L1
387	.L2
388	retn
389
390	.L3 mov eax,edi
391	and eax,BYTE 11b
392	jz .L4
393	mov al,[esi+0]
394	mov ah,[esi+1]
395	mov ebx,eax
396	mov edx,eax
397	and eax,BYTE 11000b ; blue
398	shr eax,3
399	and ebx,11100000000b ; green
400	shr ebx,6
401	and edx,1110000000000000b ; red
402	shr edx,8
403	add eax,ebx
404	add eax,edx
405	mov [edi],al
406	add esi,BYTE 2
407	inc edi
408	dec ecx
409	jmp SHORT .L3
410
411	.L4 ; save ebp
412	push ebp
413
414	; save count
415	push ecx
416
417	; unroll 4 times
418	shr ecx,2
419
420	; prestep
421	mov dl,[esi+0]
422	mov bl,[esi+1]
423	mov dh,[esi+2]
424
425	.L5 shl edx,16
426	mov bh,[esi+3]
427
428	shl ebx,16
429	mov dl,[esi+4]
430
431	mov dh,[esi+6]
432	mov bl,[esi+5]
433
434	and edx,00011000000110000001100000011000b
435	mov bh,[esi+7]
436
437	ror edx,16+3
438	mov eax,ebx ; setup eax for reds
439
440	and ebx,00000111000001110000011100000111b
441	and eax,11100000111000001110000011100000b ; reds
442
443	ror ebx,16-2
444	add esi,BYTE 8
445
446	ror eax,16
447	add edi,BYTE 4
448
449	add eax,ebx
450	mov bl,[esi+1] ; greens
451
452	add eax,edx
453	mov dl,[esi+0] ; blues
454
455	mov [edi-4],eax
456	mov dh,[esi+2]
457
458	dec ecx
459	jnz .L5
460
461	; check tail
462	pop ecx
463	and ecx,BYTE 11b
464	jz .L7
465
466	.L6 ; tail
467	mov al,[esi+0]
468	mov ah,[esi+1]
469	mov ebx,eax
470	mov edx,eax
471	and eax,BYTE 11000b ; blue
472	shr eax,3
473	and ebx,11100000000b ; green
474	shr ebx,6
475	and edx,1110000000000000b ; red
476	shr edx,8
477	add eax,ebx
478	add eax,edx
479	mov [edi],al
480	add esi,BYTE 2
481	inc edi
482	dec ecx
483	jnz .L6
484
485	.L7 pop ebp
486	retn
487
488	%ifidn __OUTPUT_FORMAT__,elf
489	section .note.GNU-stack noalloc noexec nowrite progbits
490	%endif