WebSVN – Kolibri OS – Blame – /kernel/trunk/drivers/mixer.asm

Rev	Author	Line No.	Line
431	serge	1	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
		2	;; ;;
694	serge	3	;; Copyright (C) KolibriOS team 2006-2008. All rights reserved. ;;
431	serge	4	;; Distributed under terms of the GNU General Public License ;;
		5	;; ;;
		6	;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
		7
168	serge	8	; (C) copyright Serge 2006
		9	; email: infinity_sound@mail.ru
		10
378	serge	11
168	serge	12	align 4
378	serge	13
694	serge	14	mix_list rd 32*3
378	serge	15
		16	align 4
168	serge	17	proc new_mix stdcall, output:dword
281	serge	18	locals
378	serge	19	main_count rd 1
		20	fpu_state rb 528 ;512+16
281	serge	21	endl
168	serge	22
378	serge	23	mov [main_count], 32
281	serge	24	call prepare_playlist
378	serge	25	cmp [play_count], 0
		26	je .clear
168	serge	27
378	serge	28	lea eax, [fpu_state+16]
		29	and eax, -16 ;must be 16b aligned
188	serge	30	call FpuSave
168	serge	31
378	serge	32	call update_stream
		33	.mix:
		34	lea eax, [mix_list]
		35	call do_mix_list
		36	test eax, eax
		37	je .done
168	serge	38
697	serge	39	if USE_SSE2_MIXER
		40
695	serge	41	cmp eax, 1
		42	ja @F
		43	;use fast path
		44	mov edi, [output]
		45	lea edx, [mix_list]
		46	call mix_fast
		47	jmp .next
		48	@@:
697	serge	49	cmp eax, 2
		50	ja @F
		51
		52	mov edi, [output]
		53	lea edx, [mix_list]
		54	call mix_fast_2_stream
		55	jmp .next
		56	@@:
		57
		58	end if
		59
378	serge	60	lea ebx, [mix_list]
		61	stdcall mix_all, [output], ebx, eax
695	serge	62	.next:
378	serge	63	add [output], 512
293	serge	64	dec [main_count]
378	serge	65	jnz .mix
		66	.exit:
		67	lea eax, [fpu_state+16]
		68	and eax, -16
188	serge	69	call FpuRestore
281	serge	70	ret
378	serge	71	.done:
		72	mov ecx, [main_count]
		73	shl ecx, 7 ;ecx*= 512/4
		74
281	serge	75	mov edi, [output]
		76	xor eax, eax
		77	cld
		78	rep stosd
378	serge	79	jmp .exit
		80	.clear:
		81	mov edi, [output]
		82	mov ecx, 4096
		83	xor eax, eax
		84	cld
		85	rep stosd
281	serge	86	ret
168	serge	87	endp
		88
		89	align 4
		90	proc update_stream
281	serge	91	locals
170	serge	92	stream_index dd ?
568	serge	93	event rd 6
281	serge	94	endl
168	serge	95
281	serge	96	mov [stream_index], 0
168	serge	97	.l1:
281	serge	98	mov edx, [stream_index]
		99	mov esi, [play_list+edx*4]
168	serge	100
378	serge	101	mov eax, [esi+STREAM.out_rp]
		102	cmp eax, [esi+STREAM.out_top]
281	serge	103	jb @f
328	serge	104	sub eax, 64*1024
168	serge	105	@@:
378	serge	106	mov [esi+STREAM.out_rp], eax
168	serge	107
378	serge	108	cmp [esi+STREAM.out_count], 16384
		109	ja .skip
168	serge	110
378	serge	111	test [esi+STREAM.format], PCM_RING
		112	jnz .ring
168	serge	113
281	serge	114	stdcall refill, esi
378	serge	115	.skip:
281	serge	116	inc [stream_index]
		117	dec [play_count]
		118	jnz .l1
		119	ret
378	serge	120	.ring:
		121	stdcall refill_ring, esi
		122	jmp .skip
		123	endp
		124
		125	align 4
		126	proc refill stdcall, str:dword
		127	locals
		128	r_size rd 1
		129	endl
		130
		131	mov ebx, [str]
		132	mov edi, [ebx+STREAM.out_wp]
		133	cmp edi, [ebx+STREAM.out_top]
		134	jb @F
		135	sub edi, 0x10000
		136	mov [ebx+STREAM.out_wp], edi
168	serge	137	@@:
378	serge	138	mov eax, [ebx+STREAM.in_count]
		139	test eax, eax
		140	jz .done
227	serge	141
378	serge	142	mov ecx, [ebx+STREAM.r_size]
		143	cmp eax, ecx
		144	jle @F
		145
		146	mov eax, ecx
		147	@@:
		148	mov ecx, eax
		149	cmp word [ebx+STREAM.format], PCM_1_16_8
		150	ja @F
		151
		152	shr eax, 1 ;two channles
		153	@@:
		154	test [ebx+STREAM.format], 1 ;even formats mono
		155	jz @F
		156
		157	shr eax, 1 ;eax= samples
		158	@@:
		159	shl eax, 15 ;eax*=32768 =r_end
		160
		161	mov [r_size], ecx
		162
		163	mov esi, [ebx+STREAM.in_rp]
		164	mov edi, [ebx+STREAM.out_wp]
		165
		166	stdcall [ebx+STREAM.resample], edi, esi, \
		167	[ebx+STREAM.r_dt], ecx, eax
		168
		169	mov ebx, [str]
		170
		171	add [ebx+STREAM.out_count], eax;
		172	add [ebx+STREAM.out_wp], eax;
		173
		174	mov eax, [ebx+STREAM.in_rp]
		175	mov ecx, [r_size]
		176	add eax, ecx
		177	add [ebx+STREAM.in_free], ecx
		178	sub [ebx+STREAM.in_count], ecx
		179
		180	cmp eax, [ebx+STREAM.in_top]
		181	jb @f
		182
		183	sub eax, [ebx+STREAM.in_size]
		184	@@:
		185	mov [ebx+STREAM.in_rp], eax
		186
		187	.done:
		188	mov eax, [ebx+STREAM.notify_event]
168	serge	189	test eax, eax
378	serge	190	jz .exit
227	serge	191
378	serge	192	mov ebx, [ebx+STREAM.notify_id]
1058	Galkov	193	mov edx, EVENT_WATCHED
		194	xor esi, esi
		195	call RaiseEvent ;eax, ebx, edx, esi
378	serge	196	.exit:
281	serge	197	ret
168	serge	198	endp
		199
		200	align 4
378	serge	201	proc refill_ring stdcall, str:dword
281	serge	202	locals
378	serge	203	event rd 6
281	serge	204	endl
168	serge	205
281	serge	206	mov ebx, [str]
378	serge	207	mov edi, [ebx+STREAM.out_wp]
		208	cmp edi, [ebx+STREAM.out_top]
328	serge	209	jb @F
378	serge	210	sub edi, 0x10000
		211	mov [ebx+STREAM.out_wp], edi
328	serge	212	@@:
378	serge	213	mov ecx, [ebx+STREAM.r_size]
		214	mov eax, ecx
		215	cmp word [ebx+STREAM.format], PCM_1_16_8
		216	ja @F
168	serge	217
378	serge	218	shr eax, 1 ;two channles
		219	@@:
		220	test [ebx+STREAM.format], 1 ;even formats mono
		221	jz @F
		222
		223	shr eax, 1 ;eax= samples
		224	@@:
		225	shl eax, 15 ;eax*=32768 =r_end
		226
		227	mov esi, [ebx+STREAM.in_rp]
		228	mov edi, [ebx+STREAM.out_wp]
		229
285	serge	230	stdcall [ebx+STREAM.resample], edi, esi, \
378	serge	231	[ebx+STREAM.r_dt], ecx, eax
168	serge	232
281	serge	233	mov ebx, [str]
168	serge	234
378	serge	235	add [ebx+STREAM.out_count], eax;
		236	add [ebx+STREAM.out_wp], eax;
168	serge	237
378	serge	238	mov eax, [ebx+STREAM.in_rp]
		239	mov ecx, [ebx+STREAM.r_size]
		240	add eax, ecx
		241	add [ebx+STREAM.in_free], ecx
		242	sub [ebx+STREAM.in_count], ecx
		243
		244	cmp eax, [ebx+STREAM.in_top]
281	serge	245	jb @f
		246
378	serge	247	sub eax, [ebx+STREAM.in_size]
168	serge	248	@@:
378	serge	249	mov [ebx+STREAM.in_rp], eax
168	serge	250
378	serge	251	sub eax, [ebx+STREAM.in_base]
		252	sub eax, 128
1058	Galkov	253	lea esi, [event]
378	serge	254
1058	Galkov	255	mov dword [esi], RT_INP_EMPTY
		256	mov dword [esi+4], 0
		257	mov dword [esi+8], ebx
		258	mov dword [esi+12], eax
378	serge	259
		260	mov eax, [ebx+STREAM.notify_event]
		261	test eax, eax
		262	jz .exit
		263
		264	mov ebx, [ebx+STREAM.notify_id]
1058	Galkov	265	xor edx, edx
		266	call RaiseEvent ;eax, ebx, edx, esi
378	serge	267	.exit:
		268	ret
		269	endp
168	serge	270
694	serge	271	if USE_SSE2_MIXER
		272
378	serge	273	align 4
		274	proc mix_all stdcall, dest:dword, list:dword, count:dword
		275
		276	mov edi, [dest]
694	serge	277	mov ebx, 32
		278	.mix:
		279	mov edx, [list]
		280	mov ecx, [count]
		281
		282	mov eax, [edx]
		283
		284	movdqa xmm1, [eax]
		285	movss xmm2, [edx+4]
		286	movss xmm3, [edx+8]
		287
		288	punpcklwd xmm0, xmm1
		289	punpckhwd xmm1, xmm1
		290
		291	shufps xmm2, xmm3, 0
		292	shufps xmm2, xmm2, 0x88
		293
		294	psrad xmm0, 16
		295	psrad xmm1, 16
695	serge	296	cvtdq2ps xmm0, xmm0
		297	cvtdq2ps xmm1, xmm1
694	serge	298	mulps xmm0, xmm2
		299	mulps xmm1, xmm2
		300
		301	.mix_loop:
		302	add dword [edx], 16
		303	add edx, 12
		304	dec ecx
		305	jz @F
		306
		307	mov eax, [edx]
		308
		309	movdqa xmm3, [eax]
		310	movss xmm4, [edx+4]
		311	movss xmm5, [edx+8]
		312
		313	punpcklwd xmm2, xmm3
		314	punpckhwd xmm3, xmm3
		315
		316	shufps xmm4, xmm5, 0
		317	shufps xmm4, xmm4, 0x88
		318
		319	psrad xmm2, 16
		320	psrad xmm3, 16
		321
		322	cvtdq2ps xmm2, xmm2
		323	cvtdq2ps xmm3, xmm3
		324
		325	mulps xmm2, xmm4
		326	mulps xmm3, xmm4
		327	addps xmm0, xmm2
		328	addps xmm1, xmm3
		329
		330	jmp .mix_loop
		331	@@:
		332	cvtps2dq xmm0, xmm0
		333	cvtps2dq xmm1, xmm1
		334	packssdw xmm0, xmm0
		335	packssdw xmm1, xmm1
		336	punpcklqdq xmm0, xmm1
		337	movntdq [edi], xmm0
		338
		339	add edi, 16
		340	dec ebx
		341	jnz .mix
		342
		343	ret
		344	endp
		345
695	serge	346	; param
		347	; edi = dest
		348	; edx = mix_list
		349
		350	align 4
		351	mix_fast:
		352
		353	mov ebx, 32
		354	mov eax, [edx]
		355
		356	movss xmm2, [edx+4] ; vol Lf
		357	movss xmm3, [edx+8] ; vol Rf
		358	shufps xmm2, xmm3, 0 ; Rf Rf Lf Lf
		359	shufps xmm2, xmm2, 0x88 ; volume level Rf Lf Rf Lf
		360	.mix:
		361	movdqa xmm1, [eax] ; R3w L3w R2w L2w R1w L1w R0w L0w
697	serge	362	add eax, 16
695	serge	363	punpcklwd xmm0, xmm1 ; R1w R1w L1w L1W R0w R0w L0w L0w
		364	punpckhwd xmm1, xmm1 ; R3w R3w L3w L3w R2w R2w L2w L2w
		365
		366	psrad xmm0, 16 ; R1d L1d R0d L0d
		367	psrad xmm1, 16 ; R3d L3d R2d L2d
		368
		369	cvtdq2ps xmm0, xmm0 ; time to use all power
		370	cvtdq2ps xmm1, xmm1 ; of the dark side
		371
		372	mulps xmm0, xmm2 ; R1f' L1f' R0f' L0f'
		373	mulps xmm1, xmm2 ; R3f' L3f' R2f' L2f'
		374
		375	cvtps2dq xmm0, xmm0 ; R1d' L1d' R0d' L0d'
		376	cvtps2dq xmm1, xmm1 ; R3d' L3d' R2d' L2d'
		377	packssdw xmm0, xmm0 ; R1w' L1w' R0w' L0w' R1w' L1w' R0w' L0w'
		378	packssdw xmm1, xmm1 ; R3w' L3w' R2w' L2w' R3w' L3w' R2w' L2w'
		379	punpcklqdq xmm0, xmm1 ; R3w' L3w' R2w' L2w' R1w' L1w' R0w' L0w'
		380	movntdq [edi], xmm0
		381
		382	add edi, 16
		383	dec ebx
		384	jnz .mix
		385
		386	ret
		387
697	serge	388	align 4
		389	mix_fast_2_stream:
		390
		391	mov ebx, 32
		392	mov eax, [edx]
		393
		394	movss xmm4, [edx+4] ; vol Lf
		395	movss xmm5, [edx+8] ; vol Rf
		396	mov ecx, [edx+12]
		397
		398	movss xmm6, [edx+16] ; vol Lf
		399	movss xmm7, [edx+20] ; vol Rf
		400
		401	shufps xmm4, xmm5, 0 ; Rf Rf Lf Lf
		402	shufps xmm4, xmm4, 0x88 ; volume level Rf Lf Rf Lf
		403
		404	shufps xmm6, xmm7, 0 ; Rf Rf Lf Lf
		405	shufps xmm6, xmm6, 0x88 ; volume level Rf Lf Rf Lf
		406
		407	.mix:
		408	movdqa xmm1, [eax] ; R3w L3w R2w L2w R1w L1w R0w L0w
		409	movdqa xmm3, [ecx] ; R3w L3w R2w L2w R1w L1w R0w L0w
		410
		411	add eax, 16
		412	add ecx, 16
		413
		414	punpcklwd xmm0, xmm1 ; R1w R1w L1w L1W R0w R0w L0w L0w
		415	punpckhwd xmm1, xmm1 ; R3w R3w L3w L3w R2w R2w L2w L2w
		416
		417	psrad xmm0, 16 ; R1d L1d R0d L0d
		418	psrad xmm1, 16 ; R3d L3d R2d L2d
		419
		420	cvtdq2ps xmm0, xmm0 ; time to use all power
		421	cvtdq2ps xmm1, xmm1 ; of the dark side
		422
		423	mulps xmm0, xmm4 ; R1f' L1f' R0f' L0f'
		424	mulps xmm1, xmm4 ; R3f' L3f' R2f' L2f'
		425
		426	punpcklwd xmm2, xmm3 ; R1w R1w L1w L1W R0w R0w L0w L0w
		427	punpckhwd xmm3, xmm3 ; R3w R3w L3w L3w R2w R2w L2w L2w
		428
		429	psrad xmm2, 16 ; R1d L1d R0d L0d
		430	psrad xmm3, 16 ; R3d L3d R2d L2d
		431
		432	cvtdq2ps xmm2, xmm2 ; time to use all power
		433	cvtdq2ps xmm3, xmm3 ; of the dark side
		434
		435	mulps xmm2, xmm6 ; R1f' L1f' R0f' L0f'
		436	mulps xmm3, xmm6 ; R3f' L3f' R2f' L2f'
		437
		438	addps xmm0, xmm2
		439	addps xmm1, xmm3
		440
		441	cvtps2dq xmm0, xmm0 ; R1d' L1d' R0d' L0d'
		442	cvtps2dq xmm1, xmm1 ; R3d' L3d' R2d' L2d'
		443	packssdw xmm0, xmm0 ; R1w' L1w' R0w' L0w' R1w' L1w' R0w' L0w'
		444	packssdw xmm1, xmm1 ; R3w' L3w' R2w' L2w' R3w' L3w' R2w' L2w'
		445	punpcklqdq xmm0, xmm1 ; R3w' L3w' R2w' L2w' R1w' L1w' R0w' L0w'
		446	movntdq [edi], xmm0
		447
		448	add edi, 16
		449	dec ebx
		450	jnz .mix
		451
		452	ret
		453
694	serge	454	else ; fixed point mmx version
		455
		456	align 4
		457	proc mix_all stdcall, dest:dword, list:dword, count:dword
		458
		459	mov edi, [dest]
378	serge	460	mov ebx, 64
		461	.mix:
		462	mov edx, [list]
		463	mov ecx, [count]
		464
		465	mov eax, [edx]
694	serge	466
378	serge	467	movq mm0, [eax]
694	serge	468
378	serge	469	movd mm1, [edx+4]
		470	punpckldq mm1,mm1
		471	pmulhw mm0, mm1
		472	psllw mm0, 1
		473
		474	.mix_loop:
		475	add dword [edx], 8
694	serge	476	add edx, 12
378	serge	477	dec ecx
		478	jz @F
		479
		480	mov eax, [edx]
		481	movq mm1, [eax]
		482	movd mm2, [edx+4]
		483	punpckldq mm2,mm2
		484	pmulhw mm1, mm2
		485	psllw mm1, 1
		486	paddsw mm0, mm1
		487	jmp .mix_loop
168	serge	488	@@:
378	serge	489	movq [edi], mm0
		490	add edi, 8
		491	dec ebx
		492	jnz .mix
227	serge	493
291	serge	494	ret
168	serge	495	endp
		496
694	serge	497	end if
		498
		499
168	serge	500	align 4
285	serge	501	proc resample_1 stdcall, dest:dword,src:dword,\
291	serge	502	r_dt:dword, r_size:dword,r_end:dword
168	serge	503
285	serge	504	; dest equ esp+8
		505	; src equ esp+12
		506	; r_dt equ esp+16
		507	; r_size equ esp+20
378	serge	508	; r_end equ esp+24
168	serge	509
285	serge	510	mov edi, [dest]
		511	mov edx, [src]
		512	sub edx, 32*2
		513	mov eax, 16
168	serge	514
378	serge	515	align 4
168	serge	516	.l1:
291	serge	517	mov ecx, eax
		518	mov esi, eax
		519	and ecx, 0x7FFF
		520	shr esi, 15
		521	lea esi, [edx+esi*2]
168	serge	522
291	serge	523	movsx ebp, word [esi]
		524	movsx esi, word [esi+2]
		525	mov ebx, 32768
		526	imul esi, ecx
		527	sub ebx, ecx
		528	imul ebx, ebp
		529	lea ecx, [ebx+esi+16384]
		530	sar ecx, 15
		531	cmp ecx, 32767 ; 00007fffH
		532	jle @f
		533	mov ecx, 32767 ; 00007fffH
		534	jmp .write
168	serge	535	@@:
291	serge	536	cmp ecx, -32768 ; ffff8000H
		537	jge .write
		538	mov ecx, -32768 ; ffff8000H
168	serge	539	.write:
291	serge	540	mov ebx, ecx
		541	shl ebx, 16
		542	mov bx, cx
		543	mov [edi], ebx
		544	add edi, 4
168	serge	545
293	serge	546	add eax, [esp+16]
		547	cmp eax, [esp+24]
291	serge	548	jb .l1
168	serge	549
291	serge	550	mov ebp, esp
168	serge	551
291	serge	552	sub edi, [dest]
		553	mov eax, edi
		554	ret
168	serge	555	endp
		556
		557	align 4
285	serge	558	proc resample_18 stdcall, dest:dword,src:dword,\
291	serge	559	r_dt:dword, r_size:dword,r_end:dword
168	serge	560
		561
291	serge	562	mov edi, [dest]
285	serge	563	mov edx, [src]
		564	sub edx, 32
		565
291	serge	566	mov esi, 16
168	serge	567
378	serge	568	align 4
168	serge	569	.l1:
291	serge	570	mov ecx, esi
		571	mov eax, esi
		572	and ecx, 0x7FFF
		573	shr eax, 15
		574	lea eax, [edx+eax]
168	serge	575
291	serge	576	mov bx, word [eax]
		577	sub bh, 0x80
		578	sub bl, 0x80
		579	movsx eax, bh
		580	shl eax,8
		581	movsx ebp, bl
		582	shl ebp,8
		583	mov ebx, 32768
		584	imul eax, ecx
		585	sub ebx, ecx
		586	imul ebx, ebp
		587	lea ecx, [ebx+eax+16384]
		588	sar ecx, 15
		589	cmp ecx, 32767 ; 00007fffH
		590	jle @f
		591	mov ecx, 32767 ; 00007fffH
		592	jmp .write
168	serge	593	@@:
291	serge	594	cmp ecx, -32768 ; ffff8000H
		595	jge .write
		596	mov ecx, -32768 ; ffff8000H
168	serge	597	.write:
291	serge	598	mov ebx, ecx
		599	shl ebx, 16
		600	mov bx, cx
		601	mov [edi], ebx
		602	add edi, 4
168	serge	603
293	serge	604	add esi, [esp+16]
		605	cmp esi, [esp+24]
291	serge	606	jb .l1
168	serge	607
291	serge	608	mov ebp, esp
		609	sub edi, [dest]
		610	mov eax, edi
		611	ret
168	serge	612	endp
		613
		614	align 4
285	serge	615	proc copy_stream stdcall, dest:dword,src:dword,\
291	serge	616	r_dt:dword, r_size:dword,r_end:dword
168	serge	617
281	serge	618	mov ecx, [r_size]
		619	mov eax, ecx
		620	shr ecx, 2
		621	mov esi, [src]
		622	mov edi, [dest]
575	serge	623	cld
281	serge	624	rep movsd
		625	ret
168	serge	626	endp
		627
		628	align 4
285	serge	629	proc resample_2 stdcall, dest:dword,src:dword,\
291	serge	630	r_dt:dword, r_size:dword,r_end:dword
168	serge	631
285	serge	632	mov edx, [src]
		633	sub edx, 32*4
		634	mov edi, [dest]
		635	mov ebx, [r_dt]
		636	mov eax, 16
		637	emms
168	serge	638
378	serge	639	align 4
168	serge	640	.l1:
285	serge	641	mov ecx, eax
		642	mov esi, eax
		643	and ecx, 0x7FFF
		644	shr esi, 15
		645	lea esi, [edx+esi*4]
168	serge	646
285	serge	647	movq mm0, [esi]
		648	movq mm1, mm0
168	serge	649
285	serge	650	movd mm2, ecx
		651	punpcklwd mm2, mm2
		652	movq mm3, qword [m7] ;0x8000
168	serge	653
285	serge	654	psubw mm3, mm2 ; ;0x8000 - iconst
		655	punpckldq mm3, mm2
168	serge	656
285	serge	657	pmulhw mm0, mm3
		658	pmullw mm1, mm3
168	serge	659
285	serge	660	movq mm4, mm1
		661	punpcklwd mm1, mm0
		662	punpckhwd mm4, mm0
		663	paddd mm1, mm4
		664	psrad mm1, 15
		665	packssdw mm1, mm1
		666	movd [edi], mm1
		667	add edi, 4
168	serge	668
285	serge	669	add eax, ebx
		670	cmp eax, [r_end]
		671	jb .l1
		672	emms
168	serge	673
285	serge	674	sub edi, [dest]
		675	mov eax, edi
		676	ret
168	serge	677	endp
		678
		679	align 4
285	serge	680	proc resample_28 stdcall, dest:dword,src:dword,\
291	serge	681	r_dt:dword, r_size:dword,r_end:dword
168	serge	682
285	serge	683	mov edx, [src]
		684	sub edx, 32*2
		685	mov edi, [dest]
		686	mov ebx, [r_dt]
		687	mov eax, 16
		688	emms
		689	movq mm7,[mm80]
		690	movq mm6,[mm_mask]
168	serge	691
378	serge	692	align 4
168	serge	693	.l1:
291	serge	694	mov ecx, eax
		695	mov esi, eax
		696	and ecx, 0x7FFF
		697	shr esi, 15
		698	lea esi, [edx+esi*2]
168	serge	699
291	serge	700	movq mm0, [esi]
		701	psubb mm0,mm7
		702	punpcklbw mm0,mm0
		703	pand mm0,mm6
168	serge	704
291	serge	705	movq mm1, mm0
168	serge	706
291	serge	707	movd mm2, ecx
		708	punpcklwd mm2, mm2
		709	movq mm3, qword [m7] ; // 0x8000
168	serge	710
291	serge	711	psubw mm3, mm2 ; // 0x8000 - iconst
		712	punpckldq mm3, mm2
168	serge	713
291	serge	714	pmulhw mm0, mm3
		715	pmullw mm1, mm3
168	serge	716
291	serge	717	movq mm4, mm1
		718	punpcklwd mm1, mm0
		719	punpckhwd mm4, mm0
		720	paddd mm1, mm4
		721	psrad mm1, 15
		722	packssdw mm1, mm1
		723	movd [edi], mm1
		724	add edi, 4
168	serge	725
291	serge	726	add eax, ebx
		727	cmp eax, [r_end]
		728	jb .l1
		729	emms
168	serge	730
		731
291	serge	732	sub edi, [dest]
		733	mov eax, edi
		734	ret
168	serge	735	endp
		736
		737
285	serge	738	proc m16_stereo stdcall, dest:dword,src:dword,\
291	serge	739	r_dt:dword, r_size:dword,r_end:dword
168	serge	740
291	serge	741	mov esi, [src]
		742	mov edi, [dest]
		743	mov ecx, [r_size]
		744	shr ecx,8
168	serge	745	@@:
291	serge	746	call m16_s_mmx
		747	add edi, 128
		748	add esi, 64
		749	call m16_s_mmx
		750	add edi, 128
		751	add esi, 64
		752	call m16_s_mmx
		753	add edi, 128
		754	add esi, 64
		755	call m16_s_mmx
		756	add edi, 128
		757	add esi, 64
		758	dec ecx
		759	jnz @b
168	serge	760
291	serge	761	mov eax, [r_size]
		762	add eax, eax
		763	ret
168	serge	764	endp
		765
		766	align 4
285	serge	767	proc s8_stereo stdcall, dest:dword,src:dword,\
291	serge	768	r_dt:dword, r_size:dword,r_end:dword
168	serge	769
281	serge	770	mov esi, [src]
		771	mov edi, [dest]
		772	mov ecx, [r_size]
		773	shr ecx, 7
168	serge	774
281	serge	775	movq mm7, [mm80]
		776	movq mm6, [mm_mask]
168	serge	777	@@:
281	serge	778	call s8_s_mmx
		779	add edi, 64
		780	add esi, 32
		781	call s8_s_mmx
		782	add edi, 64
		783	add esi, 32
		784	call s8_s_mmx
		785	add edi, 64
		786	add esi, 32
		787	call s8_s_mmx
		788	add edi, 64
		789	add esi, 32
168	serge	790	dec ecx
281	serge	791	jnz @b
168	serge	792
281	serge	793	mov eax, [r_size]
		794	add eax, eax
		795	ret
168	serge	796	endp
		797
285	serge	798	proc m8_stereo stdcall, dest:dword,src:dword,\
291	serge	799	r_dt:dword, r_size:dword,r_end:dword
168	serge	800
281	serge	801	mov esi, [src]
		802	mov edi, [dest]
		803	mov ecx, [r_size]
		804	shr ecx, 6
168	serge	805
281	serge	806	movq mm7, [mm80]
		807	movq mm6, [mm_mask]
168	serge	808	@@:
281	serge	809	call m8_s_mmx
		810	add edi, 64
		811	add esi, 16
		812	call m8_s_mmx
		813	add edi, 64
		814	add esi, 16
		815	call m8_s_mmx
		816	add edi, 64
		817	add esi, 16
		818	call m8_s_mmx
		819	add edi, 64
		820	add esi, 16
		821	dec ecx
		822	jnz @b
168	serge	823
281	serge	824	mov eax, [r_size]
		825	add eax, eax
		826	add eax, eax
		827	ret
168	serge	828	endp
		829
		830	align 4
		831	proc alloc_mix_buff
		832
281	serge	833	bsf eax, [mix_buff_map]
		834	jnz .find
		835	xor eax, eax
		836	ret
168	serge	837	.find:
281	serge	838	btr [mix_buff_map], eax
		839	shl eax, 9
		840	add eax, [mix_buff]
		841	ret
168	serge	842	endp
		843
293	serge	844	align 4
168	serge	845	proc m16_s_mmx
		846
291	serge	847	movq mm0, [esi]
		848	movq mm1, mm0
		849	punpcklwd mm0, mm0
		850	punpckhwd mm1, mm1
		851	movq [edi], mm0
		852	movq [edi+8], mm1
168	serge	853
291	serge	854	movq mm0, [esi+8]
		855	movq mm1, mm0
		856	punpcklwd mm0, mm0
		857	punpckhwd mm1, mm1
		858	movq [edi+16], mm0
		859	movq [edi+24], mm1
168	serge	860
291	serge	861	movq mm0, [esi+16]
		862	movq mm1, mm0
		863	punpcklwd mm0, mm0
		864	punpckhwd mm1, mm1
		865	movq [edi+32], mm0
		866	movq [edi+40], mm1
168	serge	867
291	serge	868	movq mm0, [esi+24]
		869	movq mm1, mm0
		870	punpcklwd mm0, mm0
		871	punpckhwd mm1, mm1
		872	movq [edi+48], mm0
		873	movq [edi+56], mm1
168	serge	874
291	serge	875	movq mm0, [esi+32]
		876	movq mm1, mm0
		877	punpcklwd mm0, mm0
		878	punpckhwd mm1, mm1
		879	movq [edi+64], mm0
		880	movq [edi+72], mm1
168	serge	881
291	serge	882	movq mm0, [esi+40]
		883	movq mm1, mm0
		884	punpcklwd mm0, mm0
		885	punpckhwd mm1, mm1
		886	movq [edi+80], mm0
		887	movq [edi+88], mm1
168	serge	888
		889
291	serge	890	movq mm0, [esi+48]
		891	movq mm1, mm0
		892	punpcklwd mm0, mm0
		893	punpckhwd mm1, mm1
		894	movq [edi+96], mm0
		895	movq [edi+104], mm1
168	serge	896
291	serge	897	movq mm0, [esi+56]
		898	movq mm1, mm0
		899	punpcklwd mm0, mm0
		900	punpckhwd mm1, mm1
		901	movq [edi+112], mm0
		902	movq [edi+120], mm1
168	serge	903
291	serge	904	ret
168	serge	905	endp
		906
		907	align 4
		908	proc s8_s_mmx
		909
281	serge	910	movq mm0, [esi]
		911	psubb mm0, mm7
		912	movq mm1, mm0
		913	punpcklbw mm0, mm0
		914	pand mm0, mm6
		915	punpckhbw mm1, mm1
		916	pand mm1, mm6
		917	movq [edi], mm0
		918	movq [edi+8], mm1
168	serge	919
281	serge	920	movq mm0, [esi+8]
		921	psubb mm0, mm7
		922	movq mm1, mm0
		923	punpcklbw mm0, mm0
		924	pand mm0, mm6
		925	punpckhbw mm1, mm1
		926	pand mm1, mm6
		927	movq [edi+16], mm0
		928	movq [edi+24], mm1
168	serge	929
281	serge	930	movq mm0, [esi+16]
		931	psubb mm0, mm7
		932	movq mm1, mm0
		933	punpcklbw mm0, mm0
		934	pand mm0, mm6
		935	punpckhbw mm1, mm1
		936	pand mm1, mm6
		937	movq [edi+32], mm0
		938	movq [edi+40], mm1
168	serge	939
281	serge	940	movq mm0, [esi+24]
		941	psubb mm0, mm7
		942	movq mm1, mm0
		943	punpcklbw mm0, mm0
		944	pand mm0, mm6
		945	punpckhbw mm1, mm1
		946	pand mm1, mm6
		947	movq [edi+48], mm0
		948	movq [edi+56], mm1
168	serge	949
281	serge	950	ret
168	serge	951
		952	endp
		953
		954	align 4
		955	proc m8_s_mmx
		956
281	serge	957	movq mm0, [esi]
		958	psubb mm0, mm7
		959	movq mm1, mm0
		960	punpcklbw mm0, mm0
		961	pand mm0, mm6
		962	punpckhbw mm1, mm1
		963	pand mm1, mm6
		964	movq mm2, mm0
		965	punpcklwd mm0, mm0
		966	punpckhwd mm2, mm2
168	serge	967
281	serge	968	movq mm3, mm1
		969	punpcklwd mm1, mm1
		970	punpckhwd mm3, mm3
168	serge	971
281	serge	972	movq [edi], mm0
		973	movq [edi+8], mm2
		974	movq [edi+16], mm1
		975	movq [edi+24], mm3
168	serge	976
281	serge	977	movq mm0, [esi+8]
		978	psubb mm0, mm7
		979	movq mm1, mm0
		980	punpcklbw mm0, mm0
		981	pand mm0, mm6
		982	punpckhbw mm1, mm1
		983	pand mm1, mm6
		984	movq mm2, mm0
		985	punpcklwd mm0, mm0
		986	punpckhwd mm2, mm2
168	serge	987
281	serge	988	movq mm3, mm1
		989	punpcklwd mm1, mm1
		990	punpckhwd mm3, mm3
168	serge	991
281	serge	992	movq [edi+32], mm0
		993	movq [edi+40], mm2
		994	movq [edi+48], mm1
		995	movq [edi+56], mm3
168	serge	996
281	serge	997	ret
168	serge	998	endp
		999
		1000	align 4
		1001	proc mix_2_1 stdcall, output:dword, str0:dword, str1:dword
		1002
281	serge	1003	mov edi, [output]
293	serge	1004	mov eax, [str0]
		1005	mov ebx, [str1]
		1006	mov esi, 128
		1007	call [mix_2_core] ;edi, eax, ebx
168	serge	1008
293	serge	1009	add edi, esi
		1010	add eax, esi
		1011	add ebx, esi
		1012	call [mix_2_core] ;edi, eax, ebx
168	serge	1013
293	serge	1014	add edi, esi
		1015	add eax, esi
		1016	add ebx, esi
		1017	call [mix_2_core] ;edi, eax, ebx
		1018
		1019	add edi, esi
		1020	add eax, esi
		1021	add ebx, esi
		1022	call [mix_2_core] ;edi, eax, ebx
281	serge	1023	ret
168	serge	1024	endp
		1025
		1026	align 4
		1027	proc mix_3_1 stdcall, output:dword, str0:dword, str1:dword, str2:dword
		1028
291	serge	1029	mov edi, [output]
293	serge	1030	mov eax, [str0]
		1031	mov ebx, [str1]
		1032	mov ecx, [str2]
		1033	mov esi, 128
		1034	call [mix_3_core]
168	serge	1035
293	serge	1036	add edi, esi
		1037	add eax, esi
		1038	add ebx, esi
		1039	add ecx, esi
		1040	call [mix_3_core]
168	serge	1041
293	serge	1042	add edi, esi
		1043	add eax, esi
		1044	add ebx, esi
		1045	add ecx, esi
		1046	call [mix_3_core]
		1047
		1048	add edi, esi
		1049	add eax, esi
		1050	add ebx, esi
		1051	add ecx, esi
		1052	call [mix_3_core]
291	serge	1053	ret
168	serge	1054	endp
		1055
		1056	align 4
		1057	proc mix_4_1 stdcall, str0:dword, str1:dword,\
291	serge	1058	str2:dword, str3:dword
168	serge	1059
291	serge	1060	local output:DWORD
168	serge	1061
291	serge	1062	call alloc_mix_buff
		1063	and eax, eax
		1064	jz .err
168	serge	1065
378	serge	1066	mov [output], eax
		1067
291	serge	1068	mov edi, eax
293	serge	1069	mov eax, [str0]
		1070	mov ebx, [str1]
		1071	mov ecx, [str2]
		1072	mov edx, [str3]
		1073	mov esi, 128
		1074	call [mix_4_core] ;edi, eax, ebx, ecx, edx
168	serge	1075
293	serge	1076	add edi, esi
		1077	add eax, esi
		1078	add ebx, esi
		1079	add ecx, esi
		1080	add edx, esi
		1081	call [mix_4_core] ;edi, eax, ebx, ecx, edx
		1082
		1083	add edi, esi
		1084	add eax, esi
		1085	add ebx, esi
		1086	add ecx, esi
		1087	add edx, esi
		1088	call [mix_4_core] ;edi, eax, ebx, ecx, edx
		1089
		1090	add edi, esi
		1091	add eax, esi
		1092	add ebx, esi
		1093	add ecx, esi
		1094	add edx, esi
		1095	call [mix_4_core] ;edi, eax, ebx, ecx, edx
291	serge	1096	mov eax, [output]
		1097	ret
168	serge	1098	.err:
291	serge	1099	xor eax, eax
		1100	ret
168	serge	1101	endp
		1102
		1103
		1104	align 4
		1105	proc final_mix stdcall, output:dword, str0:dword, str1:dword,\
291	serge	1106	str2:dword, str3:dword
168	serge	1107
291	serge	1108	mov edi, [output]
168	serge	1109
281	serge	1110	mov eax, [str0]
		1111	mov ebx, [str1]
		1112	mov ecx, [str2]
293	serge	1113	mov edx, [str3]
		1114	mov esi, 128
		1115	call [mix_4_core] ;edi, eax, ebx, ecx, edx
168	serge	1116
293	serge	1117	add edi, esi
		1118	add eax, esi
		1119	add ebx, esi
		1120	add ecx, esi
		1121	add edx, esi
		1122	call [mix_4_core] ;edi, eax, ebx, ecx, edx
168	serge	1123
293	serge	1124	add edi, esi
		1125	add eax, esi
		1126	add ebx, esi
		1127	add ecx, esi
		1128	add edx, esi
		1129	call [mix_4_core] ;edi, eax, ebx, ecx, edx
168	serge	1130
293	serge	1131	add edi, esi
		1132	add eax, esi
		1133	add ebx, esi
		1134	add ecx, esi
		1135	add edx, esi
		1136	call [mix_4_core] ;edi, eax, ebx, ecx, edx
281	serge	1137	ret
168	serge	1138	endp
		1139
		1140	align 4
		1141	proc copy_mem stdcall, output:dword, input:dword
		1142
291	serge	1143	mov edi, [output]
		1144	mov esi, [input]
		1145	mov ecx, 0x80
168	serge	1146	.l1:
291	serge	1147	mov eax, [esi]
		1148	mov [edi], eax
		1149	add esi, 4
		1150	add edi, 4
		1151	loop .l1
168	serge	1152
291	serge	1153	ret
168	serge	1154	endp
		1155
		1156	proc memcpy
		1157	@@:
291	serge	1158	mov eax, [esi]
		1159	mov [edi], eax
		1160	add esi, 4
		1161	add edi, 4
		1162	dec ecx
		1163	jnz @B
		1164	ret
168	serge	1165	endp
		1166
378	serge	1167	if 0
168	serge	1168
378	serge	1169	align 4
		1170	proc new_mix stdcall, output:dword
		1171	locals
		1172	mixCounter dd ?
		1173	mixIndex dd ?
		1174	streamIndex dd ?
		1175	inputCount dd ?
		1176	main_count dd ?
		1177	blockCount dd ?
		1178	mix_out dd ?
		1179	endl
		1180
		1181	call prepare_playlist
		1182
		1183	cmp [play_count], 0
		1184	je .exit
		1185	call FpuSave
		1186	mov [main_count], 32;
		1187	.l00:
		1188	mov [mix_buff_map], 0x0000FFFF;
		1189	xor eax, eax
		1190	mov [mixCounter], eax
		1191	mov [mixIndex],eax
		1192	mov [streamIndex], eax;
		1193	mov ebx, [play_count]
		1194	mov [inputCount], ebx
		1195	.l0:
		1196	mov ecx, 4
		1197	.l1:
		1198	mov ebx, [streamIndex]
		1199	mov esi, [play_list+ebx*4]
		1200	mov eax, [esi+STREAM.work_read]
		1201	add [esi+STREAM.work_read], 512
		1202
		1203	mov ebx, [mixIndex]
		1204	mov [mix_input+ebx*4], eax
		1205	inc [mixCounter]
		1206	inc [mixIndex]
		1207	inc [streamIndex]
		1208	dec [inputCount]
		1209	jz .m2
		1210
		1211	dec ecx
		1212	jnz .l1
		1213
		1214	cmp [mixCounter], 4
		1215	jnz .m2
		1216
		1217	stdcall mix_4_1, [mix_input],[mix_input+4],[mix_input+8],[mix_input+12]
		1218	sub [mixIndex],4
		1219	mov ebx, [mixIndex]
		1220	mov [mix_input+ebx*4], eax
		1221	inc [mixIndex]
		1222	mov [mixCounter], 0
		1223
		1224	cmp [inputCount], 0
		1225	jnz .l0
		1226	.m2:
		1227	cmp [mixIndex], 1
		1228	jne @f
		1229	stdcall copy_mem, [output], [mix_input]
		1230	jmp .m3
		1231	@@:
		1232	cmp [mixIndex], 2
		1233	jne @f
		1234	stdcall mix_2_1, [output], [mix_input], [mix_input+4]
		1235	jmp .m3
		1236	@@:
		1237	cmp [mixIndex], 3
		1238	jne @f
		1239	stdcall mix_3_1, [output],[mix_input],[mix_input+4],[mix_input+8]
		1240	jmp .m3
		1241	@@:
		1242	stdcall final_mix, [output],[mix_input],[mix_input+4],[mix_input+8], [mix_input+12]
		1243	.m3:
		1244	add [output],512
		1245
		1246	dec [main_count]
		1247	jnz .l00
		1248
		1249	call update_stream
		1250	emms
		1251	call FpuRestore
		1252	ret
		1253	.exit:
		1254	mov edi, [output]
		1255	mov ecx, 0x1000
		1256	xor eax, eax
		1257	cld
		1258	rep stosd
		1259	ret
		1260	endp
		1261
		1262	end if

Subversion Repositories Kolibri OS

(root)/kernel/trunk/drivers/mixer.asm – Rev 1058