Use hand-written constants table on ARMv5+ for JPEG IDCT, and load four 16-bit constants at a time with ldrd. Not useful for ARMv4, since one load per constant would still be needed, and limited range of ldrsh would force multiple copies of table.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@21535 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 86 insertions, 76 deletions

diff --git a/apps/recorder/jpeg_idct_arm.S b/apps/recorder/jpeg_idct_arm.S
index 01b08c4b5a..d84e5e7962 100644
--- a/apps/recorder/jpeg_idct_arm.S
+++ b/apps/recorder/jpeg_idct_arm.S
@@ -113,7 +113,11 @@ jpeg_idct2h:
    results can not be stored merged.
 */
     stmdb  sp!, { r4-r5, lr }
+#if ARM_ARCH < 5
     ldr    r14, =4112
+#else
+    ldrsh  r14, .Lpool4+2
+#endif
 1:
     ldrsh  r12, [r0]
     ldrsh  r4,  [r0, #2]
@@ -140,7 +144,7 @@ jpeg_idct2h:
     ldmia  sp!, { r4-r5, pc }
 #else
     stmdb  sp!, { r4, lr }
-    ldr    r14, =4112
+    ldrsh  r14, .Lpool4+2
 1:
     ldr    r12, [r0]
     sadd16 r12, r12, r14
@@ -198,27 +202,26 @@ jpeg_idct4v:
     ldmia  sp!, { r4-r7, pc }
 #elif ARM_ARCH < 6
     stmdb sp!, { r4-r8, lr }
-    ldr    r8,  =1024
-    ldr    r14, =4433
-    ldr    r12, =3302955134
+    mov    r8,  #1024
+    ldrd   r4,  .Lpool4
 1:
-    ldrsh  r5,  [r0, #48]
+    ldrsh  r14, [r0, #48]
     ldrsh  r3,  [r0, #16]
-    ldrsh  r4,  [r0, #32]
+    ldrsh  r12, [r0, #32]
     ldrsh  r2,  [r0]
-    add    r6,  r3,  r5            /* r6 = z1 = d1 + d3 */
-    add    r7,  r2,  r4            /* r7 = tmp10 >> 2 = d0 + d2 */
-    smlabb r6, r14, r6,  r8        /* z1 *= 4433 */
-    sub    r2,  r2,  r4            /* r2 = tmp12 >> 2= d0 - d2 */
-    smlabb r3,  r12, r3,  r6       /* r3 = tmp2 = z1 + z2 * 6270 */ 
-    smlatb r5,  r12, r5,  r6       /* r5 = tmp0 = z1 - z3 * 15137 */
+    add    r6,  r3,  r14           /* r6  = z1 = d1 + d3 */
+    add    r7,  r2,  r12           /* r7  = tmp10 >> 2 = d0 + d2 */
+    smlabb r6,  r5,  r6,  r8       /* z1 *= 4433 */
+    sub    r2,  r2,  r12           /* r2  = tmp12 >> 2= d0 - d2 */
+    smlatb r3,  r5,  r3,  r6       /* r3  = tmp2 = z1 + z2 * 6270 */ 
+    smlabb r14, r4,  r14, r6       /* r14 = tmp0 = z1 - z3 * 15137 */
     mov    r7,  r7,  lsl #2
     mov    r2,  r2,  lsl #2
-    add    r4,  r7,  r3,  asr #11   /* r4 = o0 */
-    sub    r7,  r7,  r3,  asr #11   /* r7 = o3 */
-    add    r3,  r2,  r5,  asr #11   /* r3 = o1 */
-    sub    r2,  r2,  r5,  asr #11   /* r2 = o2 */
-    strh   r4,  [r0]
+    add    r12, r7,  r3,  asr #11   /* r12 = o0 */
+    sub    r7,  r7,  r3,  asr #11   /* r7  = o3 */
+    add    r3,  r2,  r14, asr #11   /* r3  = o1 */
+    sub    r2,  r2,  r14, asr #11   /* r2  = o2 */
+    strh   r12, [r0]
     strh   r7,  [r0, #48]
     strh   r3,  [r0, #16]
     strh   r2,  [r0, #32]
@@ -228,9 +231,8 @@ jpeg_idct4v:
     ldmia sp!, { r4-r8, pc }
 #else
     stmdb  sp!, { r4-r10, lr }
-    ldr    r2,  =1024
-    ldr    r3,  =4433
-    ldr    r12, =3302955134
+    ldrd   r2,  .Lpool4
+    mov    r12, #1024
 1:
     ldr    r6,  [r0, #32]
     ldr    r4,  [r0]
@@ -247,12 +249,12 @@ jpeg_idct4v:
     /* multiplication expands values beyond 16 bits, so this part needs to be
        split. the values will be merged below so that the rest of the addition
        can be done in parallel */
-    smlabb r9,  r3,  r6,  r2       /* r9 = z1[0] = (d1 * d3) * 4433 + 1024 */
-    smlabt r6,  r3,  r6,  r2       /* r6 = z1[1] = (d1 * d3) * 4433 + 1024 */
-    smlabb r10, r12, r5,  r9       /* r10 = tmp2[0] = z1 + d1 * 6270 */
-    smlatb r14, r12, r7,  r9       /* r14 = tmp0[0] = z1 - d3 * 15137 */
-    smlabt r5,  r12, r5,  r6       /* r5  = tmp2[1] */
-    smlatt r6,  r12, r7,  r6       /* r6  = tmp0[1] */
+    smlabb r9,  r3,  r6,  r12      /* r9 = z1[0] = (d1 * d3) * 4433 + 1024 */
+    smlabt r6,  r3,  r6,  r12      /* r6 = z1[1] = (d1 * d3) * 4433 + 1024 */
+    smlatb r10, r3,  r5,  r9       /* r10 = tmp2[0] = z1 + d1 * 6270 */
+    smlabb r14, r2,  r7,  r9       /* r14 = tmp0[0] = z1 - d3 * 15137 */
+    smlatt r5,  r3,  r5,  r6       /* r5  = tmp2[1] */
+    smlabt r6,  r2,  r7,  r6       /* r6  = tmp0[1] */
     mov    r8,  r8,  lsl #2        /* complete the parallel shift started */
     mov    r4,  r4,  lsl #2        /* with the earlier bic instructions */
     /* tmp2 are in r10, r5; tmp0 are in r14, r6 */
@@ -276,6 +278,17 @@ jpeg_idct4v:
 #endif
     .size jpeg_idct4v, .-jpeg_idct4v
 
+#if ARM_ARCH > 4
+    .align 4
+.Lpool4:
+    .short -15137
+    .short 4112
+    .short 4433
+    .short 6270
+
+    .align 2
+#endif
+
 jpeg_idct4h:
 #if ARM_ARCH < 5
     stmdb  sp!, { r4-r10, lr }
@@ -328,88 +341,85 @@ jpeg_idct4h:
     cmp    r0,  r2
     bcc    1b
     ldmia sp!, { r4-r10, pc }
-#elif ARM_ARCH < 6
-    stmdb  sp!, { r4-r10, lr }
-    ldr    r10, =4433
-    ldr    r14, =4112
-    ldr    r12, =3302955134
+#elif ARM_ARCH < 6 || 1
+    stmdb  sp!, { r4-r9, lr }
+    ldrd   r4,  .Lpool4
 1:
     ldrsh  r7,  [r0, #6]
-    ldrsh  r5,  [r0, #2]
-    ldrsh  r4,  [r0]
+    ldrsh  r14, [r0, #2]
+    ldrsh  r12, [r0]
     ldrsh  r6,  [r0, #4]
-    add    r8,  r5,  r7             /* r8 = z1 = d1 + d3 */
-    add    r4,  r4,  r14
-    smulbb r8, r10, r8              /* z1 *= 4433 */
-    add    r9,  r4,  r6             /* r9 = tmp10 >> 13 = d0 + d2 */
-    smlabb r5,  r12, r5,  r8        /* r5 = tmp2 = z1 + z2 * 6270 */ 
-    smlatb r7,  r12, r7,  r8        /* r7 = tmp0 = z1 - z3 * 15137 */
-    sub    r4,  r4,  r6             /* r4 = tmp12 >> 13 = d0 - d2 */
-    add    r6,  r5,  r9,  lsl #13   /* r6 = o0 */
-    rsb    r9,  r5,  r9,  lsl #13   /* r9 = o3 */
-    add    r5,  r7,  r4,  lsl #13   /* r5 = o1 */
-    rsb    r4,  r7,  r4,  lsl #13   /* r4 = o2 */
+    add    r8,  r14, r7             /* r8 = z1 = d1 + d3 */
+    add    r12, r12, r4,  lsr #16
+    smulbb r8,  r5,  r8             /* z1 *= 4433 */
+    add    r9,  r12, r6             /* r9 = tmp10 >> 13 = d0 + d2 */
+    smlatb r14, r5,  r14, r8        /* r14= tmp2 = z1 + z2 * 6270 */ 
+    smlabb r7,  r4,  r7,  r8        /* r7 = tmp0 = z1 - z3 * 15137 */
+    sub    r12, r12, r6             /* r12= tmp12 >> 13 = d0 - d2 */
+    add    r6,  r14, r9,  lsl #13   /* r6 = o0 */
+    rsb    r9,  r14, r9,  lsl #13   /* r9 = o3 */
+    add    r14, r7,  r12, lsl #13   /* r14= o1 */
+    rsb    r12, r7,  r12, lsl #13   /* r12= o2 */
     mov    r6,  r6,  asr #18
-    mov    r5,  r5,  asr #18
-    mov    r4,  r4,  asr #18
+    mov    r14, r14, asr #18
+    mov    r12, r12, asr #18
     mov    r9,  r9,  asr #18
     cmp    r6,  #255
     mvnhi  r6,  r6,  asr #31
-    cmp    r5,  #255
-    mvnhi  r5,  r5,  asr #31
-    cmp    r4,  #255
-    mvnhi  r4,  r4,  asr #31
+    cmp    r14, #255
+    mvnhi  r14, r14, asr #31
+    cmp    r12, #255
+    mvnhi  r12, r12, asr #31
     cmp    r9,  #255
     mvnhi  r9,  r9,  asr #31
 #ifdef HAVE_LCD_COLOR
     strb   r6,  [r1]
-    strb   r5,  [r1, #4]
-    strb   r4,  [r1, #8]
+    strb   r14, [r1, #4]
+    strb   r12, [r1, #8]
     strb   r9,  [r1, #12]
 #else
     strb   r6,  [r1]
-    strb   r5,  [r1, #1]
-    strb   r4,  [r1, #2]
+    strb   r14, [r1, #1]
+    strb   r12, [r1, #2]
     strb   r9,  [r1, #3]
 #endif
     add    r0,  r0,  #16
     add    r1,  r1,  r3
     cmp    r0,  r2
     bcc    1b
-    ldmia sp!, { r4-r10, pc }
+    ldmia sp!, { r4-r9, pc }
 #else
     stmdb sp!, { r4-r9, lr }
-    ldr    r9,  =4433
-    ldr    r14, =4112
-    ldr    r12, =3302955134
+    ldrd   r4,  .Lpool4
+    mov    r9,  r4,  lsr #16
 1:
-    ldmia  r0,  { r4-r5 }
-    sadd16 r4,  r4,  r14
-    sadd16 r6,  r4,  r5             /* r6lo = d0 + d2, r6hi = d1 + d3 */
-    ssub16 r7,  r4,  r5             /* r7lo = d0 - d2 */
-    smulbt r8,  r9,  r6
+    ldmia  r0,  { r12, r14 }
+    sadd16 r12, r12, r9
+    sadd16 r6,  r12, r14            /* r6lo = d0 + d2, r6hi = d1 + d3 */
+    ssub16 r7,  r12, r14            /* r7lo = d0 - d2 */
+    smulbt r8,  r5,  r6
     sxth   r6,  r6
-    smlabt r4,  r12, r4,  r8        /* r4 = tmp2 = z1 + z2 * 6270 */
-    smlatt r5,  r12, r5,  r8        /* r5 = tmp0 = z1 - z3 * 15137 */
+    smlatt r12, r5,  r12, r8        /* r12= tmp2 = z1 + z2 * 6270 */
+    smlabt r14, r4,  r14, r8        /* r14= tmp0 = z1 - z3 * 15137 */
     sxth   r7,  r7
-    add    r8,  r4,  r6,  lsl #13   /* r8 = o0 */
-    rsb    r6,  r4,  r6,  lsl #13   /* r6 = o3 */
-    add    r4,  r5,  r7,  lsl #13   /* r4 = o1 */
-    rsb    r5,  r5,  r7,  lsl #13   /* r5 = o2 */
+    add    r8,  r12, r6,  lsl #13   /* r8 = o0 */
+    rsb    r6,  r12, r6,  lsl #13   /* r6 = o3 */
+    add    r12, r14, r7,  lsl #13   /* r12= o1 */
+    rsb    r14, r14, r7,  lsl #13   /* r14= o2 */
     usat  r8,  #8,  r8,  asr #18
     usat  r6,  #8,  r6,  asr #18
-    usat  r4,  #8,  r4,  asr #18
-    usat  r5,  #8,  r5,  asr #18
+    usat  r12, #8,  r12, asr #18
+    usat  r14, #8,  r14, asr #18
 #ifdef HAVE_LCD_COLOR
     strb   r8,  [r1]
     strb   r6,  [r1, #12]
-    strb   r4,  [r1, #4]
-    strb   r5,  [r1, #8]
+    strb   r12, [r1, #4]
+    strb   r14, [r1, #8]
 #else
     strb   r8,  [r1]
     strb   r6,  [r1, #3]
-    strb   r4,  [r1, #1]
-    strb   r5,  [r1, #2]
+    strb   r12, [r1, #1]
+    strb   r14, [r1, #2]
 #endif
     add   r0,  r0,  #16
     add   r1,  r1,  r3
@@ -450,7 +460,7 @@ jpeg_idct8v:
     mov    r11, r11, asr #16        /* r11 = z3 = d6 */
     add    r8,  r8,  #8192
     add    r9,  r10, r11
-    mov    r8,  r8,  asr #3         /* r8  = z4 = (d0 + 4112) << 13 */
+    mov    r8,  r8,  asr #3         /* r8  = z4 = (d0 << 13) + 1024 */
     mul    r9,  r14, r9             /* r9  = z1 = (z2 + z3) * 4433 */
     ldr    r14, =6270
     mla    r11, r12, r11, r9         /* r11 = tmp2 = z1 - z3 * 15137 */