Adapt ARM crossfeed assembler to work like the Coldfire one. Remove SWITCHPARAM cruft as it's no longer needed.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12507 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 60 insertions, 77 deletions

diff --git a/apps/dsp.c b/apps/dsp.c
index d27df0500c..cca94ae075 100644
--- a/apps/dsp.c
+++ b/apps/dsp.c
@@ -112,7 +112,7 @@ struct crossfeed_data
     int32_t coefs[3];       /* 04h - Coefficients for the shelving filter */
     int32_t history[4];     /* 10h - Format is x[n - 1], y[n - 1] for both channels */
     int32_t delay[13][2];   /* 20h */
-    int     index;          /* 88h - Current index/pointer into the delay line */
+    int32_t *index;         /* 88h - Current pointer into the delay line */
                             /* 8ch */
 };
 
@@ -138,10 +138,6 @@ typedef int (*resample_fn_type)(int count, struct dsp_data *data,
                                 int32_t *src[], int32_t *dst[]);
 typedef void (*sample_output_fn_type)(int count, struct dsp_data *data,
                                       int32_t *src[], int16_t *dst);
-/* If ACF_SWITCHPARAM is no longer needed, make apply_crossfeed of type
-   channels_process_fn_type since it is really just that */
-typedef void (*apply_crossfeed_fn_type)(ACF_SWITCHPARAM(int count,
-                                                        int32_t *buf[]));
 typedef void (*channels_process_fn_type)(int count, int32_t *buf[]);
 
 /*
@@ -164,7 +160,7 @@ struct dsp_config
     sample_output_fn_type       output_samples;
     /* These will be NULL for the voice codec and is more economical that
        way */
-    apply_crossfeed_fn_type     apply_crossfeed;
+    channels_process_fn_type    apply_crossfeed;
     channels_process_fn_type    channels_process;
 };
 
@@ -175,14 +171,10 @@ static struct dither_data dither_data[2] IBSS_ATTR; /* 0=left, 1=right */
 static long   dither_mask IBSS_ATTR;
 static long   dither_bias IBSS_ATTR;
 /* Crossfeed */
-#ifdef DSP_CROSSFEED_DELAY_PTR
 struct crossfeed_data crossfeed_data IDATA_ATTR =    /* A */
 {
-    .index = (intptr_t)crossfeed_data.delay
+    .index = (int32_t *)crossfeed_data.delay
 };
-#else
-struct crossfeed_data crossfeed_data IBSS_ATTR;     /* A */
-#endif
 
 /* Equalizer */
 static struct eq_state eq_data;                     /* A/V */
@@ -719,7 +711,7 @@ static void apply_crossfeed(int count, int32_t *buf[])
     int32_t *delay = &crossfeed_data.delay[0][0];
     int32_t *coefs = &crossfeed_data.coefs[0];
     int32_t gain = crossfeed_data.gain;
-    int di = crossfeed_data.index;
+    int32_t *di = crossfeed_data.index;
  
     int32_t acc;
     int32_t left, right;
@@ -731,28 +723,28 @@ static void apply_crossfeed(int count, int32_t *buf[])
         right = buf[1][i];
 
         /* Filter delayed sample from left speaker */
-        ACC_INIT(acc, delay[di*2], coefs[0]);
+        ACC_INIT(acc, *di, coefs[0]);
         ACC(acc, hist_l[0], coefs[1]);
         ACC(acc, hist_l[1], coefs[2]);
         /* Save filter history for left speaker */
         hist_l[1] = GET_ACC(acc);
-        hist_l[0] = delay[di*2];
+        hist_l[0] = *di;
+        *di++ = left;
         /* Filter delayed sample from right speaker */
-        ACC_INIT(acc, delay[di*2 + 1], coefs[0]);
+        ACC_INIT(acc, *di, coefs[0]);
         ACC(acc, hist_r[0], coefs[1]);
         ACC(acc, hist_r[1], coefs[2]);
         /* Save filter history for right speaker */
         hist_r[1] = GET_ACC(acc);
-        hist_r[0] = delay[di*2 + 1];
-        delay[di*2] = left;
-        delay[di*2 + 1] = right;
+        hist_r[0] = *di;
+        *di++ = right;
         /* Now add the attenuated direct sound and write to outputs */
         buf[0][i] = FRACMUL(left, gain) + hist_r[1];
         buf[1][i] = FRACMUL(right, gain) + hist_l[1];
  
         /* Wrap delay line index if bigger than delay line size */
-        if (++di > 12)
-            di = 0;
+        if (di >= delay + 13*2)
+            di = delay;
     }
     /* Write back local copies of data we've modified */
     crossfeed_data.index = di;
@@ -1127,7 +1119,7 @@ int dsp_process(char *dst, const char *src[], int count)
         if ((samples = resample(samples, tmp)) <= 0)
             break; /* I'm pretty sure we're downsampling here */
         if (dsp->apply_crossfeed)
-            dsp->apply_crossfeed(ACF_SWITCHPARAM(samples, tmp));
+            dsp->apply_crossfeed(samples, tmp);
         /* TODO: EQ and tone controls need separate structs for audio and voice
          * DSP processing thanks to filter history. isn't really audible now, but
          * might be the day we start handling voice more delicately.
diff --git a/apps/dsp_arm.S b/apps/dsp_arm.S
index 1abfd34983..27669203f1 100644
--- a/apps/dsp_arm.S
+++ b/apps/dsp_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.
@@ -17,63 +17,64 @@
  *
  ****************************************************************************/
 
+/*
+ * void apply_crossfeed(int count, int32_t* src[])
+ */
     .section .text
     .global apply_crossfeed 
 apply_crossfeed:
     @ unfortunately, we ended up in a bit of a register squeeze here, and need
     @ to keep both the count and the delay line index on the stack :/
-    stmdb sp!, { r4-r11, lr }           @ stack modified regs
-    ldmia r0, { r2-r3 }                 @ r2 = src[0], r3 = src[1]
+    stmdb   sp!, { r4-r11, lr }        @ stack modified regs
+    ldmia   r1, { r2-r3 }              @ r2 = src[0], r3 = src[1]
     
-    ldr r0, =crossfeed_data
-    ldmia r0!, { r4-r11 }               @ load direct gain and filter data
-    ldr r12, [r0, #13*4*2]              @ fetch delay line index
-    add r0, r0, r12, lsl #3             @ r0 = &delay[index][0]
-    stmdb sp!, { r1, r12 }              @ stack count and delay line index
+    ldr     r1, =crossfeed_data
+    ldmia   r1!, { r4-r11 }            @ load direct gain and filter data
+    add     r12, r1, #13*4*2           @ calculate end of delay
+    stmdb   sp!, { r0, r12 }           @ stack count and end of delay adr
+    ldr     r0, [r1, #13*4*2]          @ fetch current delay line address
+
     /* Register usage in loop:
      * r0 = &delay[index][0], r1 = accumulator high, r2 = src[0], r3 = src[1],
      * r4 = direct gain, r5-r7 = b0, b1, a1 (filter coefs),
      * r8-r11 = filter history, r12 = temp, r14 = accumulator low
      */
 .cfloop:
-    smull r14, r1, r6, r8               @ acc = b1*dr[n - 1]
-    smlal r14, r1, r7, r9               @ acc += a1*y_l[n - 1]
-    ldr r8, [r0, #4]                    @ r8 = dr[n]
-    smlal r14, r1, r5, r8               @ acc += b0*dr[n]
-    mov r9, r1, lsl #1                  @ fix format for filter history
-    ldr r12, [r2]                       @ load left input
-    smlal r14, r1, r4, r12              @ acc += gain*x_l[n] 
-    mov r1, r1, lsl #1                  @ fix format
-    str r1, [r2], #4                    @ save result
-    
-    smull r14, r1, r6, r10              @ acc = b1*dl[n - 1]
-    smlal r14, r1, r7, r11              @ acc += a1*y_r[n - 1]
-    ldr r10, [r0]                       @ r10 = dl[n]
-    str r12, [r0], #4                   @ save left input to delay line
-    smlal r14, r1, r5, r10              @ acc += b0*dl[n]
-    mov r11, r1, lsl #1                 @ fix format for filter history
-    ldr r12, [r3]                       @ load right input
-    smlal r14, r1, r4, r12              @ acc += gain*x_r[n]
-    str r12, [r0], #4                   @ save right input to delay line
-    mov r1, r1, lsl #1                  @ fix format
-    str r1, [r3], #4                    @ save result
+    smull   r14, r1, r6, r8            @ acc = b1*dr[n - 1]
+    smlal   r14, r1, r7, r9            @ acc += a1*y_l[n - 1]
+    ldr     r8, [r0, #4]               @ r8 = dr[n]
+    smlal   r14, r1, r5, r8            @ acc += b0*dr[n]
+    mov     r9, r1, lsl #1             @ fix format for filter history
+    ldr     r12, [r2]                  @ load left input
+    smlal   r14, r1, r4, r12           @ acc += gain*x_l[n] 
+    mov     r1, r1, lsl #1             @ fix format
+    str     r1, [r2], #4               @ save result
 
-    ldr r12, [sp, #4]                   @ fetch delay line index from stack
-    add r12, r12, #1                    @ increment index
-    cmp r12, #13                        @ do we need to wrap to start of delay?
-    moveq r12, #0                       @ yes, wrap index to 0
-    subeq r0, r0, #13*4*2               @ also wrap back delay line ptr to start
-    str r12, [sp, #4]                   @ stack delay line index again
-    
-    ldr r1, [sp]                        @ fetch count from stack
-    subs r1, r1, #1                     @ are we finished?
-    strne r1, [sp]                      @ nope, save count back to stack
-    bne .cfloop
+    smull   r14, r1, r6, r10           @ acc = b1*dl[n - 1]
+    smlal   r14, r1, r7, r11           @ acc += a1*y_r[n - 1]
+    ldr     r10, [r0]                  @ r10 = dl[n]
+    str     r12, [r0], #4              @ save left input to delay line
+    smlal   r14, r1, r5, r10           @ acc += b0*dl[n]
+    mov     r11, r1, lsl #1            @ fix format for filter history
+    ldr     r12, [r3]                  @ load right input
+    smlal   r14, r1, r4, r12           @ acc += gain*x_r[n]
+    str     r12, [r0], #4              @ save right input to delay line
+    mov     r1, r1, lsl #1             @ fix format
+    str     r1, [r3], #4               @ save result
+
+    ldr     r12, [sp, #4]              @ fetch delay line end addr from stack
+    cmp     r0, r12                    @ need to wrap to start of delay?
+    subeq   r0, r0, #13*4*2            @ wrap back delay line ptr to start
+ 
+    ldr     r1, [sp]                   @ fetch count from stack
+    subs    r1, r1, #1                 @ are we finished?
+    strne   r1, [sp]                   @ nope, save count back to stack
+    bne     .cfloop
     
     @ save data back to struct
-    ldr r0, =crossfeed_data + 4*4
-    stmia r0, { r8-r11 }                @ save filter history
-    str r12, [r0, #30*4]                @ save delay line index
-    add sp, sp, #8                      @ remove temp variables from stack
-    ldmia sp!, { r4-r11, pc }
+    ldr     r12, =crossfeed_data + 4*4
+    stmia   r12, { r8-r11 }            @ save filter history
+    str     r0, [r12, #30*4]            @ save delay line index
+    add     sp, sp, #8                 @ remove temp variables from stack
+    ldmia   sp!, { r4-r11, pc }
 
diff --git a/apps/dsp_asm.h b/apps/dsp_asm.h
index a9e7fac6b0..ee90f5763e 100644
--- a/apps/dsp_asm.h
+++ b/apps/dsp_asm.h
@@ -22,21 +22,11 @@
 #ifndef _DSP_ASM_H
 #define _DSP_ASM_H
 
-#define ACF_SWITCHPARAM(count, buf)     count, buf
-
 #ifndef SIMULATOR
 
 #if defined(CPU_COLDFIRE) || defined(CPU_ARM)
 #define DSP_HAVE_ASM_CROSSFEED
-#if defined(CPU_COLDFIRE)
-/* ACF_SWITCHPARAM can be stripped out if all have the same parameter
-   order - DSP_CROSSFEED_DELAY_PTR if all use a pointer instead of index */
-#define DSP_CROSSFEED_DELAY_PTR
-#else
-#undef ACF_SWITCHPARAM
-#define ACF_SWITCHPARAM(count, buf)     buf, count
-#endif
-void apply_crossfeed(ACF_SWITCHPARAM(int count, int32_t *buf[]));
+void apply_crossfeed(int count, int32_t *buf[]);
 #endif /* defined(CPU_COLDFIRE) || defined(CPU_ARM) */
 
 #if defined (CPU_COLDFIRE)