1 files changed, 177 insertions, 0 deletions
diff --git a/apps/dsp_arm.S b/apps/dsp_arm.S
index c3e5c7cd05..751e0f5130 100644
--- a/apps/dsp_arm.S
+++ b/apps/dsp_arm.S
@@ -18,6 +18,183 @@
 ****************************************************************************/
 /****************************************************************************
+ *  void channels_process_sound_chan_mono(int count, int32_t *buf[])
+ *
+ *  NOTE: The following code processes two samples at once. When count is odd,
+ *        there is an additional obsolete sample processed, which will not be
+ *        used by the calling functions.
+ */
+    .section .icode, "ax", %progbits
+    .align  2
+    .global channels_process_sound_chan_mono
+    .type   channels_process_sound_chan_mono, %function
+channels_process_sound_chan_mono:
+    @ input: r0 = count, r1 = buf
+    stmfd   sp!, {r4-r6, lr}
+    ldmia   r1, {r2-r3}                @ r4 = buf[0], r5 = buf[1]
+.monoloop:
+    ldmia   r2, {r4-r5}
+    ldmia   r3, {r6,lr}
+    mov     r4, r4, asr #1             @ r4 = r4/2
+    add     r4, r4, r6, asr #1         @ r4 = r4 + r6/2 = (buf[0]+buf[1])/2
+    mov     r5, r5, asr #1             @ r5 = r5/2
+    add     r5, r5, lr, asr #1         @ r5 = r5 + lr/2 = (buf[0]+buf[1])/2
+    stmia   r2!, {r4-r5}
+    stmia   r3!, {r4-r5}
+    subs    r0, r0, #2
+    bgt     .monoloop
+    
+    ldmfd   sp!, {r4-r6, pc}
+.monoend:
+    .size   channels_process_sound_chan_mono,.monoend-channels_process_sound_chan_mono 
+ 
+/****************************************************************************
+ *  void channels_process_sound_chan_karaoke(int count, int32_t *buf[])
+ *  NOTE: The following code processes two samples at once. When count is odd,
+ *        there is an additional obsolete sample processed, which will not be
+ *        used by the calling functions.
+ */
+    .section .icode, "ax", %progbits
+    .align  2
+    .global channels_process_sound_chan_karaoke
+    .type   channels_process_sound_chan_karaoke, %function
+channels_process_sound_chan_karaoke:
+    @ input: r0 = count, r1 = buf
+    stmfd   sp!, {r4-r6, lr}
+    ldmia   r1, {r2-r3}                @ r4 = buf[0], r5 = buf[1]
+.karaokeloop:
+    ldmia   r2, {r4-r5}
+    ldmia   r3, {r6,lr} 
+    mov     r6, r6, asr #1             @ r6 = r6/2
+    rsb     r4, r6, r4, asr #1         @ r4 = -r6 + r4/2 = (buf[0]-buf[1])/2
+    rsb     r6, r4, #0                 @ r6 = -r4
+    mov     lr, lr, asr #1             @ lr = lr/2
+    rsb     r5, lr, r5, asr #1         @ r5 = -lr + r5/2 = (buf[0]-buf[1])/2
+    rsb     lr, r5, #0                 @ lr = -r5  
+    stmia   r2!, {r4-r5}
+    stmia   r3!, {r6,lr}
+    subs    r0, r0, #2
+    bgt     .karaokeloop
+    ldmfd   sp!, {r4-r6, pc}
+.karaokeend:
+    .size   channels_process_sound_chan_karaoke,.karaokeend-channels_process_sound_chan_karaoke
+    
+/****************************************************************************
+ *  void sample_output_mono(int count, struct dsp_data *data,
+                               int32_t *src[], int16_t *dst)
+ *  NOTE: The following code processes two samples at once. When count is odd,
+ *        there is an additional obsolete sample processed, which will not be
+ *        used by the calling functions.
+ */
+    .section .icode, "ax", %progbits
+    .align  2
+    .global sample_output_mono
+    .type   sample_output_mono, %function
+sample_output_mono:
+    @ input: r0 = count, r1 = data, r2 = src, r3 = dst
+    stmfd   sp!, {r4-r9, lr}
+    ldr     r4, [r2]                   @ r4 = src[0]
+    ldr     r5, [r1]                   @ lr = data->output_scale
+    sub     r1, r5, #1                 @ r1 = r5-1
+    mov     r2, #1
+    mov     r2, r2, asl r1             @ r2 = 1<<r1 = 1 << (scale-1)
+    mvn     r1, #0x8000                @ r1 needed for clipping
+    mov     r8, #0xff00
+    orr     r8, r8, #0xff              @ r8 needed for masking
+.somloop:
+    ldmia   r4!, {r6-r7}
+    add     r6, r6, r2
+    mov     r6, r6, asr r5             @ r6 = (r6 + 1<<(scale-1)) >> scale
+    mov     lr, r6, asr #15
+    teq     lr, lr, asr #31
+    eorne   r6, r1, lr, asr #31        @ Clip (-32768...+32767)
+    add     r7, r7, r2
+    mov     r7, r7, asr r5             @ r7 = (r7 + 1<<(scale-1)) >> scale
+    mov     lr, r7, asr #15
+    teq     lr, lr, asr #31
+    eorne   r7, r1, lr, asr #31        @ Clip (-32768...+32767)
+    
+    and     r6, r6, r8
+    orr     r6, r6, r6, asl #16        @ pack first 2 halfwords into 1 word
+    and     r7, r7, r8
+    orr     r7, r7, r7, asl #16        @ pack last 2 halfwords into 1 word
+    stmia   r3!, {r6-r7}
+    
+    subs    r0, r0, #2
+    bgt     .somloop     
+       
+    ldmfd   sp!, {r4-r9, pc}
+.somend:
+    .size   sample_output_mono,.somend-sample_output_mono
+    
+/****************************************************************************
+ * void sample_output_stereo(int count, struct dsp_data *data,
+                                 int32_t *src[], int16_t *dst)
+ *  NOTE: The following code processes two samples at once. When count is odd,
+ *        there is an additional obsolete sample processed, which will not be
+ *        used by the calling functions.
+ */
+    .section .icode, "ax", %progbits
+    .align  2
+    .global sample_output_stereo
+    .type   sample_output_stereo, %function
+sample_output_stereo:
+    @ input: r0 = count, r1 = data, r2 = src, r3 = dst
+    stmfd   sp!, {r4-r11, lr}
+    ldmia   r2, {r4-r5}                @ r4 = src[0], r5 = src[1]
+    ldr     r6, [r1]                   @ r6 = data->output_scale
+    sub     r1, r6, #1                 @ r1 = r6-1
+    mov     r2, #1
+    mov     r2, r2, asl r1             @ r2 = 1<<r1 = 1 << (scale-1)
+    mvn     r1, #0x8000                @ r1 needed for clipping
+    mov     r11, #0xff00
+    orr     r11, r11, #0xff            @ r11 needed for masking
+.sosloop:
+    ldmia   r4!, {r7-r8}
+    add     r7, r7, r2
+    mov     r7, r7, asr r6             @ r7 = (r7 + 1<<(scale-1)) >> scale
+    mov     lr, r7, asr #15
+    teq     lr, lr, asr #31
+    eorne   r7, r1, lr, asr #31        @ Clip (-32768...+32767)
+    add     r8, r8, r2
+    mov     r8, r8, asr r6             @ r8 = (r8 + 1<<(scale-1)) >> scale
+    mov     lr, r8, asr #15
+    teq     lr, lr, asr #31
+    eorne   r8, r1, lr, asr #31        @ Clip (-32768...+32767)
+    
+    ldmia   r5!, {r9-r10}
+    add     r9, r9, r2
+    mov     r9, r9, asr r6             @ r9 = (r9 + 1<<(scale-1)) >> scale
+    mov     lr, r9, asr #15
+    teq     lr, lr, asr #31
+    eorne   r9, r1, lr, asr #31        @ Clip (-32768...+32767)
+    add     r10, r10, r2
+    mov     r10, r10, asr r6           @ r10 = (r10 + 1<<(scale-1)) >> scale
+    mov     lr, r10, asr #15
+    teq     lr, lr, asr #31
+    eorne   r10, r1, lr, asr #31       @ Clip (-32768...+32767)
+    
+    and     r7, r7, r11
+    orr     r9, r7, r9, asl #16        @ pack first 2 halfwords into 1 word
+    and     r8, r8, r11
+    orr     r10, r8, r10, asl #16      @ pack last 2 halfwords into 1 word
+    stmia   r3!, {r9-r10}
+    subs    r0, r0, #2
+    bgt     .sosloop
+    ldmfd   sp!, {r4-r11, pc}
+.sosend:
+    .size   sample_output_stereo,.sosend-sample_output_stereo
+/****************************************************************************
 * void apply_crossfeed(int count, int32_t* src[])
 */
    .section .text

diff --git a/apps/dsp_arm.S b/apps/dsp_arm.S index c3e5c7cd05..751e0f5130 100644 --- a/apps/dsp_arm.S +++ b/apps/dsp_arm.S
@@ -18,6 +18,183 @@
18	****************************************************************************/	18	****************************************************************************/
19		19
20	/****************************************************************************	20	/****************************************************************************
		21	* void channels_process_sound_chan_mono(int count, int32_t *buf[])
		22	*
		23	* NOTE: The following code processes two samples at once. When count is odd,
		24	* there is an additional obsolete sample processed, which will not be
		25	* used by the calling functions.
		26	*/
		27	.section .icode, "ax", %progbits
		28	.align 2
		29	.global channels_process_sound_chan_mono
		30	.type channels_process_sound_chan_mono, %function
		31	channels_process_sound_chan_mono:
		32	@ input: r0 = count, r1 = buf
		33	stmfd sp!, {r4-r6, lr}
		34	ldmia r1, {r2-r3} @ r4 = buf[0], r5 = buf[1]
		35
		36	.monoloop:
		37	ldmia r2, {r4-r5}
		38	ldmia r3, {r6,lr}
		39	mov r4, r4, asr #1 @ r4 = r4/2
		40	add r4, r4, r6, asr #1 @ r4 = r4 + r6/2 = (buf[0]+buf[1])/2
		41	mov r5, r5, asr #1 @ r5 = r5/2
		42	add r5, r5, lr, asr #1 @ r5 = r5 + lr/2 = (buf[0]+buf[1])/2
		43	stmia r2!, {r4-r5}
		44	stmia r3!, {r4-r5}
		45	subs r0, r0, #2
		46	bgt .monoloop
		47
		48	ldmfd sp!, {r4-r6, pc}
		49	.monoend:
		50	.size channels_process_sound_chan_mono,.monoend-channels_process_sound_chan_mono
		51
		52	/****************************************************************************
		53	* void channels_process_sound_chan_karaoke(int count, int32_t *buf[])
		54	* NOTE: The following code processes two samples at once. When count is odd,
		55	* there is an additional obsolete sample processed, which will not be
		56	* used by the calling functions.
		57	*/
		58	.section .icode, "ax", %progbits
		59	.align 2
		60	.global channels_process_sound_chan_karaoke
		61	.type channels_process_sound_chan_karaoke, %function
		62	channels_process_sound_chan_karaoke:
		63	@ input: r0 = count, r1 = buf
		64	stmfd sp!, {r4-r6, lr}
		65	ldmia r1, {r2-r3} @ r4 = buf[0], r5 = buf[1]
		66
		67	.karaokeloop:
		68	ldmia r2, {r4-r5}
		69	ldmia r3, {r6,lr}
		70	mov r6, r6, asr #1 @ r6 = r6/2
		71	rsb r4, r6, r4, asr #1 @ r4 = -r6 + r4/2 = (buf[0]-buf[1])/2
		72	rsb r6, r4, #0 @ r6 = -r4
		73	mov lr, lr, asr #1 @ lr = lr/2
		74	rsb r5, lr, r5, asr #1 @ r5 = -lr + r5/2 = (buf[0]-buf[1])/2
		75	rsb lr, r5, #0 @ lr = -r5
		76	stmia r2!, {r4-r5}
		77	stmia r3!, {r6,lr}
		78	subs r0, r0, #2
		79	bgt .karaokeloop
		80
		81	ldmfd sp!, {r4-r6, pc}
		82	.karaokeend:
		83	.size channels_process_sound_chan_karaoke,.karaokeend-channels_process_sound_chan_karaoke
		84
		85	/****************************************************************************
		86	* void sample_output_mono(int count, struct dsp_data *data,
		87	int32_t src[], int16_t dst)
		88	* NOTE: The following code processes two samples at once. When count is odd,
		89	* there is an additional obsolete sample processed, which will not be
		90	* used by the calling functions.
		91	*/
		92	.section .icode, "ax", %progbits
		93	.align 2
		94	.global sample_output_mono
		95	.type sample_output_mono, %function
		96	sample_output_mono:
		97	@ input: r0 = count, r1 = data, r2 = src, r3 = dst
		98	stmfd sp!, {r4-r9, lr}
		99
		100	ldr r4, [r2] @ r4 = src[0]
		101	ldr r5, [r1] @ lr = data->output_scale
		102	sub r1, r5, #1 @ r1 = r5-1
		103	mov r2, #1
		104	mov r2, r2, asl r1 @ r2 = 1<<r1 = 1 << (scale-1)
		105	mvn r1, #0x8000 @ r1 needed for clipping
		106	mov r8, #0xff00
		107	orr r8, r8, #0xff @ r8 needed for masking
		108
		109	.somloop:
		110	ldmia r4!, {r6-r7}
		111	add r6, r6, r2
		112	mov r6, r6, asr r5 @ r6 = (r6 + 1<<(scale-1)) >> scale
		113	mov lr, r6, asr #15
		114	teq lr, lr, asr #31
		115	eorne r6, r1, lr, asr #31 @ Clip (-32768...+32767)
		116	add r7, r7, r2
		117	mov r7, r7, asr r5 @ r7 = (r7 + 1<<(scale-1)) >> scale
		118	mov lr, r7, asr #15
		119	teq lr, lr, asr #31
		120	eorne r7, r1, lr, asr #31 @ Clip (-32768...+32767)
		121
		122	and r6, r6, r8
		123	orr r6, r6, r6, asl #16 @ pack first 2 halfwords into 1 word
		124	and r7, r7, r8
		125	orr r7, r7, r7, asl #16 @ pack last 2 halfwords into 1 word
		126	stmia r3!, {r6-r7}
		127
		128	subs r0, r0, #2
		129	bgt .somloop
		130
		131	ldmfd sp!, {r4-r9, pc}
		132	.somend:
		133	.size sample_output_mono,.somend-sample_output_mono
		134
		135	/****************************************************************************
		136	* void sample_output_stereo(int count, struct dsp_data *data,
		137	int32_t src[], int16_t dst)
		138	* NOTE: The following code processes two samples at once. When count is odd,
		139	* there is an additional obsolete sample processed, which will not be
		140	* used by the calling functions.
		141	*/
		142	.section .icode, "ax", %progbits
		143	.align 2
		144	.global sample_output_stereo
		145	.type sample_output_stereo, %function
		146	sample_output_stereo:
		147	@ input: r0 = count, r1 = data, r2 = src, r3 = dst
		148	stmfd sp!, {r4-r11, lr}
		149
		150	ldmia r2, {r4-r5} @ r4 = src[0], r5 = src[1]
		151	ldr r6, [r1] @ r6 = data->output_scale
		152	sub r1, r6, #1 @ r1 = r6-1
		153	mov r2, #1
		154	mov r2, r2, asl r1 @ r2 = 1<<r1 = 1 << (scale-1)
		155	mvn r1, #0x8000 @ r1 needed for clipping
		156	mov r11, #0xff00
		157	orr r11, r11, #0xff @ r11 needed for masking
		158
		159	.sosloop:
		160	ldmia r4!, {r7-r8}
		161	add r7, r7, r2
		162	mov r7, r7, asr r6 @ r7 = (r7 + 1<<(scale-1)) >> scale
		163	mov lr, r7, asr #15
		164	teq lr, lr, asr #31
		165	eorne r7, r1, lr, asr #31 @ Clip (-32768...+32767)
		166	add r8, r8, r2
		167	mov r8, r8, asr r6 @ r8 = (r8 + 1<<(scale-1)) >> scale
		168	mov lr, r8, asr #15
		169	teq lr, lr, asr #31
		170	eorne r8, r1, lr, asr #31 @ Clip (-32768...+32767)
		171
		172	ldmia r5!, {r9-r10}
		173	add r9, r9, r2
		174	mov r9, r9, asr r6 @ r9 = (r9 + 1<<(scale-1)) >> scale
		175	mov lr, r9, asr #15
		176	teq lr, lr, asr #31
		177	eorne r9, r1, lr, asr #31 @ Clip (-32768...+32767)
		178	add r10, r10, r2
		179	mov r10, r10, asr r6 @ r10 = (r10 + 1<<(scale-1)) >> scale
		180	mov lr, r10, asr #15
		181	teq lr, lr, asr #31
		182	eorne r10, r1, lr, asr #31 @ Clip (-32768...+32767)
		183
		184	and r7, r7, r11
		185	orr r9, r7, r9, asl #16 @ pack first 2 halfwords into 1 word
		186	and r8, r8, r11
		187	orr r10, r8, r10, asl #16 @ pack last 2 halfwords into 1 word
		188	stmia r3!, {r9-r10}
		189
		190	subs r0, r0, #2
		191	bgt .sosloop
		192
		193	ldmfd sp!, {r4-r11, pc}
		194	.sosend:
		195	.size sample_output_stereo,.sosend-sample_output_stereo
		196
		197	/****************************************************************************
21	* void apply_crossfeed(int count, int32_t* src[])	198	* void apply_crossfeed(int count, int32_t* src[])
22	*/	199	*/
23	.section .text	200	.section .text