1 files changed, 39 insertions, 25 deletions

diff --git a/apps/eq_arm.S b/apps/eq_arm.S
index 85617dc2fb..0c1961d2d3 100644
--- a/apps/eq_arm.S
+++ b/apps/eq_arm.S
@@ -7,7 +7,7 @@
  *                     \/            \/     \/    \/            \/
  * $Id$
  *
- * Copyright (C) 2006 Thom Johansen
+ * Copyright (C) 2006-2007 Thom Johansen
  *
  * All files in this archive are subject to the GNU General Public License.
  * See the file COPYING in the source tree root for full license agreement.
@@ -17,6 +17,15 @@
  *
  ****************************************************************************/
 
+/* uncomment this to make filtering calculate lower bits after shifting.
+ * without this, "shift" of the lower bits will be lost here.
+ */
+/* #define HIGH_PRECISION */
+
+/*
+ * void eq_filter(int32_t **x, struct eqfilter *f, unsigned num,
+ *                unsigned channels, unsigned shift)
+ */
     .text
     .global eq_filter
 eq_filter:
@@ -33,35 +42,40 @@ eq_filter:
     ldr r14, [sp, #8]       @ r14 = numsamples
     ldmia r10, { r0-r3 }    @ load history, r10 should be filter struct addr
     str r10, [sp, #4]       @ save it for loop end
-.loop:
+
     /* r0-r3 = history, r4-r8 = coefs, r9 = x[], r10..r11 = accumulator,
-       r12 = shift amount, r14 = number of samples.
-       See eq_cf.S for explanation of what this loop does. Primary difference
-       is the reordering of the equation we do here, which is done for register
-       reuse reasons, we're pretty short on regs.
+     * r12 = shift amount, r14 = number of samples.
      */
-    smull r10, r11, r6, r1  @ acc = b2*x[i - 2]
-    mov r1, r0              @ fix input history
-    smlal r10, r11, r5, r0  @ acc += b1*x[i - 1]
-    ldr r0, [r9]            @ load input and fix history in same operation
-    smlal r10, r11, r4, r0  @ acc += b0*x[i]
-    smlal r10, r11, r7, r2  @ acc += a1*y[i - 1]
-    smlal r10, r11, r8, r3  @ acc += a2*y[i - 2]
-    mov r3, r2              @ fix output history
-    mov r2, r11, lsl r12    @ get result
-    @ TODO: arm makes it easy to mix in lower bits from r10 for extended
-    @ precision here, but we don't have enough regs to save the shift factor
-    @ we would need (32 - r12).
-    str r2, [r9], #4        @ save result
-    subs r14, r14, #1       @ are we done with this channel?
+.loop:
+    /* Direct form 1 filtering code.
+     * y[n] = b0*x[i] + b1*x[i - 1] + b2*x[i - 2] + a1*y[i - 1] + a2*y[i - 2],
+     * where y[] is output and x[] is input. This is performed out of order to
+     * reuse registers, we're pretty short on regs.
+     */
+    smull r10, r11, r6, r1     @ acc = b2*x[i - 2]
+    mov r1, r0                 @ fix input history
+    smlal r10, r11, r5, r0     @ acc += b1*x[i - 1]
+    ldr r0, [r9]               @ load input and fix history in same operation
+    smlal r10, r11, r4, r0     @ acc += b0*x[i]
+    smlal r10, r11, r7, r2     @ acc += a1*y[i - 1]
+    smlal r10, r11, r8, r3     @ acc += a2*y[i - 2]
+    mov r3, r2                 @ fix output history
+    mov r2, r11, asl r12       @ get upper part of result and shift left
+#ifdef HIGH_PRECISION
+    rsb r11, r12, #32          @ get shift amount for lower part
+    orr r2, r2, r10, lsr r11   @ then mix in correctly shifted lower part
+#endif
+    str r2, [r9], #4           @ save result
+    subs r14, r14, #1          @ are we done with this channel?
     bne .loop
 
-    ldr r10, [sp, #4]       @ load filter struct pointer
-    stmia r10!, { r0-r3 }   @ save back history
-    ldr r11, [sp, #12]      @ load number of channels
-    subs r11, r11, #1       @ all channels processed?
+    ldr r10, [sp, #4]          @ load filter struct pointer
+    stmia r10!, { r0-r3 }      @ save back history
+    ldr r11, [sp, #12]         @ load number of channels
+    subs r11, r11, #1          @ all channels processed?
     strne r11, [sp, #12]
     bne .filterloop
 
-    add sp, sp, #16         @ compensate for temp storage
+    add sp, sp, #16            @ compensate for temp storage
     ldmia sp!, { r4-r11, pc }
+