Faster idct for ARMv6. Overall mpegplayer speedup is quite minimal though.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@21392 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 342 insertions, 1 deletions

diff --git a/apps/plugins/mpegplayer/SOURCES b/apps/plugins/mpegplayer/SOURCES
index 5b3360cc5a..5ca0fcd86e 100644
--- a/apps/plugins/mpegplayer/SOURCES
+++ b/apps/plugins/mpegplayer/SOURCES
@@ -9,7 +9,11 @@ idct_coldfire.S
 motion_comp_coldfire_c.c
 motion_comp_coldfire_s.S
 #elif defined CPU_ARM
+#if ARM_ARCH >= 6
+idct_armv6.S
+#else
 idct_arm.S
+#endif
 motion_comp_arm_c.c
 motion_comp_arm_s.S
 #else  /* other CPU or SIM */
diff --git a/apps/plugins/mpegplayer/decode.c b/apps/plugins/mpegplayer/decode.c
index a19b929bef..91251206e9 100644
--- a/apps/plugins/mpegplayer/decode.c
+++ b/apps/plugins/mpegplayer/decode.c
@@ -35,7 +35,7 @@
 
 #define BUFFER_SIZE (1194 * 1024)
 
-#ifdef CPU_COLDFIRE
+#if defined(CPU_COLDFIRE) || (defined(CPU_ARM) && ARM_ARCH >= 6)
 /* twice as large as on other targets because coldfire uses
  * a secondary, transposed buffer for optimisation */
 static int16_t static_dct_block[128] IBSS_ATTR ATTR_ALIGN(16);
diff --git a/apps/plugins/mpegplayer/idct_armv6.S b/apps/plugins/mpegplayer/idct_armv6.S
new file mode 100644
index 0000000000..73feed4785
--- /dev/null
+++ b/apps/plugins/mpegplayer/idct_armv6.S
@@ -0,0 +1,337 @@
+/***************************************************************************
+ *             __________               __   ___.
+ *   Open      \______   \ ____   ____ |  | _\_ |__   _______  ___
+ *   Source     |       _//  _ \_/ ___\|  |/ /| __ \ /  _ \  \/  /
+ *   Jukebox    |    |   (  <_> )  \___|    < | \_\ (  <_> > <  <
+ *   Firmware   |____|_  /\____/ \___  >__|_ \|___  /\____/__/\_ \
+ *                     \/            \/     \/    \/            \/
+ * $Id$
+ *
+ * Copyright (C) 2009 by Jens Arnold
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
+ * KIND, either express or implied.
+ *
+ ****************************************************************************/
+
+
+    .global     mpeg2_idct_copy
+    .type       mpeg2_idct_copy, %function
+    .global     mpeg2_idct_add
+    .type       mpeg2_idct_add, %function
+
+/* Custom calling convention:
+ * r0 contains block pointer and is non-volatile
+ * all non-volatile c context saved and restored on its behalf
+ */
+.idct:
+    str     lr, [sp, #-4]!      @ lr is used
+    add     r1, r0, #128        @ secondary, transposed temp buffer
+    mov     r14, #8             @ loop counter
+
+.row_loop:
+    ldmia   r0!, {r2, r3, r10, r11} @ fetch f0, f2, f4, f6, f1, f3, f5, f7
+    ldrd    r4, L_W1357         @ load  W1, W3, W5, W7
+
+    smuad   r6, r4, r10         @ b0 = W1 * f1 + W3 * f3
+    smlad   r6, r5, r11, r6     @    + W5 * f5 + W7 * f7
+
+    smultt  r7, r5, r10         @ b1 = -W7 * f3
+    smlabb  r7, r4, r11, r7     @    + -W1 * f5
+    smlabt  r7, r5, r11, r7     @    + -W5 * f7
+    rsb     r7, r7, #0
+    smlatb  r7, r4, r10, r7     @    + W3 * f1
+
+    smulbt  r8, r4, r10         @ b2 = -W1 * f3
+    rsb     r8, r8, #0
+    smlabb  r8, r5, r10, r8     @    + W5 * f1
+    smlatb  r8, r5, r11, r8     @    + W7 * f5
+    smlatt  r8, r4, r11, r8     @    + W3 * f7
+
+    smusdx  r9, r10, r5         @ b3 = f1 * W7 - f3 * W5
+    smlsdx  r9, r11, r4, r9     @    + f5 * W3 - f1 * W1
+
+    ldrd    r4, L_W0246         @ load  W0, W2, W4, W6
+    add     r2, r2, #1          @ f0 += 1
+
+    smulbb  r10, r4, r2         @ a0' = W0 * f0
+    smlabb  r10, r5, r3, r10    @     + W4 * f4
+    smultt  r12, r4, r2         @ a3' = W2 * f2
+    smlatt  r12, r5, r3, r12    @     + W6 * f6
+    add     r10, r10, r12       @ a0  = a0' + a3'
+    sub     r12, r10, r12, lsl #1  @ a3  = a0 - 2 * a3'
+
+    smulbb  r11, r5, r3         @ a1' = -W4 * f4
+    rsb     r11, r11, #0
+    smlabb  r11, r4, r2, r11    @     + W0 * f0
+    smultt  r3, r4, r3          @ a2' = -W2 * f6
+    rsb     r3, r3, #0
+    smlatt  r3, r5, r2, r3      @     + W6 * f2
+    add     r11, r11, r3        @ a1  = a1' + a2'
+    sub     r3, r11, r3, lsl #1 @ a2  = a1 - 2 * a2'
+    
+    sub     r2, r10, r6         @ block[7] = (a0 - b0)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #7*16]
+    sub     r2, r11, r7         @ block[6] = (a1 - b1)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #6*16]
+    sub     r2, r3, r8          @ block[5] = (a2 - b2)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #5*16]
+    sub     r2, r12, r9         @ block[4] = (a3 - b3)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #4*16]
+    add     r2, r12, r9         @ block[3] = (a3 + b3)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #3*16]
+    add     r2, r3, r8          @ block[2] = (a2 + b2)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #2*16]
+    add     r2, r11, r7         @ block[1] = (a1 + b1)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1, #1*16]
+    add     r2, r10, r6         @ block[0] = (a0 + b0)
+    mov     r2, r2, asr #12     @            >> 12
+    strh    r2, [r1], #2        @ advance to next temp column
+    
+    subs    r14, r14, #1
+    bne     .row_loop
+    b       .col_start
+
+    @placed here because of ldrd's offset limit
+L_W1357:
+    .short  2841
+    .short  2408
+    .short  1609
+    .short   565
+
+L_W0246:
+    .short  2048
+    .short  2676
+    .short  2048
+    .short  1108
+
+.col_start:
+    @ r0 now points to the temp buffer, where we need it.
+    sub     r1, r1, #128+16     @ point r1 back to the input block
+    mov     r14, #8             @ loop counter
+
+.col_loop:
+    ldmia   r0!, {r2, r3, r10, r11} @ fetch f0, f2, f4, f6, f1, f3, f5, f7
+    ldrd    r4, L_W1357         @ load  W1, W3, W5, W7
+
+    smuad   r6, r4, r10         @ b0 = W1 * f1 + W3 * f3
+    smlad   r6, r5, r11, r6     @    + W5 * f5 + W7 * f7
+
+    smultt  r7, r5, r10         @ b1 = -W7 * f3
+    smlabb  r7, r4, r11, r7     @    + -W1 * f5
+    smlabt  r7, r5, r11, r7     @    + -W5 * f7
+    rsb     r7, r7, #0
+    smlatb  r7, r4, r10, r7     @    + W3 * f1
+
+    smulbt  r8, r4, r10         @ b2 = -W1 * f3
+    rsb     r8, r8, #0
+    smlabb  r8, r5, r10, r8     @    + W5 * f1
+    smlatb  r8, r5, r11, r8     @    + W7 * f5
+    smlatt  r8, r4, r11, r8     @    + W3 * f7
+
+    smusdx  r9, r10, r5         @ b3 = f1 * W7 - f3 * W5
+    smlsdx  r9, r11, r4, r9     @    + f5 * W3 - f1 * W1
+
+    ldrd    r4, L_W0246         @ load  W0, W2, W4, W6
+    add     r2, r2, #32         @ DC offset: 0.5
+
+    smulbb  r10, r4, r2         @ a0' = W0 * f0
+    smlabb  r10, r5, r3, r10    @     + W4 * f4
+    smultt  r12, r4, r2         @ a3' = W2 * f2
+    smlatt  r12, r5, r3, r12    @     + W6 * f6
+    add     r10, r10, r12       @ a0  = a0' + a3'
+    sub     r12, r10, r12, lsl #1  @ a3  = a0 - 2 * a3'
+
+    smulbb  r11, r5, r3         @ a1' = -W4 * f4
+    rsb     r11, r11, #0
+    smlabb  r11, r4, r2, r11    @     + W0 * f0
+    smultt  r3, r4, r3          @ a2' = -W2 * f6
+    rsb     r3, r3, #0
+    smlatt  r3, r5, r2, r3      @     + W6 * f2
+    add     r11, r11, r3        @ a1  = a1' + a2'
+    sub     r3, r11, r3, lsl #1 @ a2  = a1 - 2 * a2'
+    
+    sub     r2, r10, r6         @ block[7] = (a0 - b0)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #7*16]
+    sub     r2, r11, r7         @ block[6] = (a1 - b1)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #6*16]
+    sub     r2, r3, r8          @ block[5] = (a2 - b2)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #5*16]
+    sub     r2, r12, r9         @ block[4] = (a3 - b3)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #4*16]
+    add     r2, r12, r9         @ block[3] = (a3 + b3)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #3*16]
+    add     r2, r3, r8          @ block[2] = (a2 + b2)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #2*16]
+    add     r2, r11, r7         @ block[1] = (a1 + b1)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1, #1*16]
+    add     r2, r10, r6         @ block[0] = (a0 + b0)
+    mov     r2, r2, asr #17     @            >> 17
+    strh    r2, [r1], #2        @ advance to next column
+
+    subs    r14, r14, #1
+    bne     .col_loop
+
+    sub     r0, r0, #256        @ point r0 back to the input block
+    ldr     pc, [sp], #4
+
+
+mpeg2_idct_copy:
+    stmfd  sp!, {r1-r2, r4-r12, lr}
+    bl     .idct
+    ldmfd  sp!, {r1-r2}
+
+    add    r12, r0, #128
+    ldrd   r4, [r0]
+    mov    r8, #0
+    mov    r9, #0
+    mov    r10, #0
+    mov    r11, #0
+1:
+    ldrd   r6, [r0, #8]
+    usat16 r4, #8, r4
+    strb   r4, [r1, #0]
+    mov    r4, r4, lsr #16
+    strb   r4, [r1, #1]
+    usat16 r5, #8, r5
+    strb   r5, [r1, #2]
+    mov    r5, r5, lsr #16
+    strb   r5, [r1, #3]
+    ldrd   r4, [r0, #16]
+    usat16 r6, #8, r6
+    strb   r6, [r1, #4]
+    mov    r6, r6, lsr #16
+    strb   r6, [r1, #5]
+    usat16 r7, #8, r7
+    strb   r7, [r1, #6]
+    mov    r7, r7, lsr #16
+    strb   r7, [r1, #7]
+    stmia  r0!, {r8-r11}
+    add    r1, r1, r2
+    cmp    r0, r12
+    blo    1b
+
+    ldmfd  sp!, {r4-r12, pc}
+
+mpeg2_idct_add:
+    cmp    r0, #129
+    mov    r0, r1
+    ldreqsh r1, [r0, #0]
+    bne    1f
+    and    r1, r1, #0x70
+    cmp    r1, #0x40
+    bne    3f
+1:
+    stmfd  sp!, {r2-r12, lr}
+    bl     .idct
+    ldmfd  sp!, {r1-r2}
+    mov    r11, #0
+    add    r12, r0, #128
+2:
+    ldmia  r0, {r3-r6}
+    ldrb   r7, [r1, #0]
+    ldrb   r8, [r1, #1]
+    ldrb   r9, [r1, #2]
+    ldrb   r10, [r1, #3]
+    str    r11, [r0], #4        
+    orr    r7, r7, r8, lsl #16
+    sadd16 r3, r3, r7
+    usat16 r3, #8, r3
+    strb   r3, [r1, #0]
+    mov    r3, r3, lsr #16
+    strb   r3, [r1, #1]
+    str    r11, [r0], #4     
+    orr    r9, r9, r10, lsl #16
+    sadd16 r4, r4, r9
+    usat16 r4, #8, r4
+    strb   r4, [r1, #2]
+    mov    r4, r4, lsr #16
+    strb   r4, [r1, #3]
+    ldrb   r7, [r1, #4]
+    ldrb   r8, [r1, #5]
+    ldrb   r9, [r1, #6]
+    ldrb   r10, [r1, #7]
+    str    r11, [r0], #4
+    orr    r7, r7, r8, lsl #16
+    sadd16 r5, r5, r7
+    usat16 r5, #8, r5
+    strb   r5, [r1, #4]
+    mov    r5, r5, lsr #16
+    strb   r5, [r1, #5]
+    str    r11, [r0], #4
+    orr    r9, r9, r10, lsl #16
+    sadd16 r6, r6, r9
+    usat16 r6, #8, r6
+    strb   r6, [r1, #6]
+    mov    r6, r6, lsr #16
+    strb   r6, [r1, #7]
+    add    r1, r1, r2
+    cmp    r0, r12
+    blo    2b
+    ldmfd  sp!, {r4-r12, pc}
+
+3:
+    stmfd  sp!, {r4-r7}
+    ldrsh  r1, [r0, #0]           /* r1 = block[0] */
+    mov    r11, #0
+    strh   r11, [r0, #0]          /* block[0] = 0 */
+    strh   r11, [r0, #126]        /* block[63] = 0 */
+    add    r1, r1, #64            /* r1 = DC << 7 */
+    add    r0, r2, r3, asl #3
+4:
+    ldrb   r4, [r2, #0]
+    ldrb   r5, [r2, #1]
+    ldrb   r6, [r2, #2]
+    ldrb   r7, [r2, #3]
+    add    r4, r4, r1, asr #7
+    usat   r4, #8, r4
+    strb   r4, [r2, #0]
+    add    r5, r5, r1, asr #7
+    usat   r5, #8, r5
+    strb   r5, [r2, #1]
+    add    r6, r6, r1, asr #7
+    usat   r6, #8, r6
+    strb   r6, [r2, #2]
+    add    r7, r7, r1, asr #7
+    usat   r7, #8, r7
+    strb   r7, [r2, #3]
+    ldrb   r4, [r2, #4]
+    ldrb   r5, [r2, #5]
+    ldrb   r6, [r2, #6]
+    ldrb   r7, [r2, #7]
+    add    r4, r4, r1, asr #7
+    usat   r4, #8, r4
+    strb   r4, [r2, #4]
+    add    r5, r5, r1, asr #7
+    usat   r5, #8, r5
+    strb   r5, [r2, #5]
+    add    r6, r6, r1, asr #7
+    usat   r6, #8, r6
+    strb   r6, [r2, #6]
+    add    r7, r7, r1, asr #7
+    usat   r7, #8, r7
+    strb   r7, [r2, #7]
+    add    r2, r2, r3
+    cmp    r2, r0
+    blo    4b
+    ldmfd  sp!, {r4-r7}
+    bx     lr