FS #6848 - fast vector operations for ARM in Tremor.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12902 a1c6a512-1295-4272-9138-f99709370657

apps/codecs/Tremor/asm_arm.h

@@ -95,6 +95,112 @@ static inline void XNPROD31(ogg_int32_t  a, ogg_int32_t  b,
   *y = y1 << 1;
 }
 
+#ifndef _V_VECT_OPS
+#define _V_VECT_OPS
+
+/* asm versions of vector operations for block.c, window.c */
+static inline 
+void vect_add(ogg_int32_t *x, ogg_int32_t *y, int n)
+{
+  while (n>=4) {
+    asm volatile ("ldmia %[x], {r0, r1, r2, r3};"
+                  "ldmia %[y]!, {r4, r5, r6, r7};"
+                  "add r0, r0, r4;"
+                  "add r1, r1, r5;"
+                  "add r2, r2, r6;"
+                  "add r3, r3, r7;"
+                  "stmia %[x]!, {r0, r1, r2, r3};"
+                  : [x] "+r" (x), [y] "+r" (y)
+                  : : "r0", "r1", "r2", "r3",
+                  "r4", "r5", "r6", "r7",
+                  "memory");
+    n -= 4;
+  }
+  /* add final elements */
+  while (n>0) {
+    *x++ += *y++;
+    n--;
+  }
+}
+
+static inline 
+void vect_copy(ogg_int32_t *x, ogg_int32_t *y, int n)
+{
+  while (n>=4) {
+    asm volatile ("ldmia %[y]!, {r0, r1, r2, r3};"
+                  "stmia %[x]!, {r0, r1, r2, r3};"
+                  : [x] "+r" (x), [y] "+r" (y)
+                  : : "r0", "r1", "r2", "r3",
+                  "memory");
+    n -= 4;
+  }
+  /* copy final elements */
+  while (n>0) {
+    *x++ = *y++;
+    n--;
+  }
+}
+
+static inline 
+void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  while (n>=4) {
+    asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
+                  "ldmia %[w]!, {r4, r5, r6, r7};"
+                  "smull r8, r9, r0, r4;"
+                  "mov   r0, r9, lsl #1;"
+                  "smull r8, r9, r1, r5;"
+                  "mov   r1, r9, lsl #1;"
+                  "smull r8, r9, r2, r6;"
+                  "mov   r2, r9, lsl #1;"
+                  "smull r8, r9, r3, r7;"
+                  "mov   r3, r9, lsl #1;"
+                  "stmia %[d]!, {r0, r1, r2, r3};"
+                  : [d] "+r" (data), [w] "+r" (window)
+                  : : "r0", "r1", "r2", "r3",
+                  "r4", "r5", "r6", "r7", "r8", "r9",
+                  "memory", "cc");
+    n -= 4;
+  }
+  while(n>0) {
+    *data = MULT31(*data, *window);
+    data++;
+    window++;
+    n--;
+  }
+}
+
+static inline
+void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  while (n>=4) {
+    asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
+                  "ldmda %[w]!, {r4, r5, r6, r7};"
+                  "smull r8, r9, r0, r7;"
+                  "mov   r0, r9, lsl #1;"
+                  "smull r8, r9, r1, r6;"
+                  "mov   r1, r9, lsl #1;"
+                  "smull r8, r9, r2, r5;"
+                  "mov   r2, r9, lsl #1;"
+                  "smull r8, r9, r3, r4;"
+                  "mov   r3, r9, lsl #1;"
+                  "stmia %[d]!, {r0, r1, r2, r3};"
+                  : [d] "+r" (data), [w] "+r" (window)
+                  : : "r0", "r1", "r2", "r3",
+                  "r4", "r5", "r6", "r7", "r8", "r9",
+                  "memory", "cc");
+    n -= 4;
+  }
+  while(n>0) {
+    *data = MULT31(*data, *window);
+    data++;
+    window--;
+    n--;
+  }
+}
+
+#endif
+
 #endif
 
 #ifndef _V_CLIP_MATH

apps/codecs/Tremor/asm_mcf5249.h

@@ -132,10 +132,13 @@ void XNPROD31(ogg_int32_t  a, ogg_int32_t  b,
                   [t] "r" (_t), [v] "r" (_v) \
                 : "cc");
 
+#ifndef _V_VECT_OPS
+#define _V_VECT_OPS
+
 /* asm versions of vector operations for block.c, window.c */
 /* assumes MAC is initialized & accumulators cleared */
 static inline 
-void mcf5249_vect_add(ogg_int32_t *x, ogg_int32_t *y, int n)
+void vect_add(ogg_int32_t *x, ogg_int32_t *y, int n)
 {
   /* align to 16 bytes */
   while(n>0 && (int)x&16) {
@@ -169,7 +172,7 @@ void mcf5249_vect_add(ogg_int32_t *x, ogg_int32_t *y, int n)
 }
 
 static inline 
-void mcf5249_vect_copy(ogg_int32_t *x, ogg_int32_t *y, int n)
+void vect_copy(ogg_int32_t *x, ogg_int32_t *y, int n)
 {
   /* align to 16 bytes */
   while(n>0 && (int)x&16) {
@@ -196,7 +199,7 @@ void mcf5249_vect_copy(ogg_int32_t *x, ogg_int32_t *y, int n)
 
 
 static inline 
-void mcf5249_vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
+void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 {
   /* ensure data is aligned to 16-bytes */
   while(n>0 && (int)data%16) {
@@ -250,7 +253,7 @@ void mcf5249_vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 }
 
 static inline 
-void mcf5249_vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
+void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
 {
   /* ensure at least data is aligned to 16-bytes */
   while(n>0 && (int)data%16) {
@@ -338,6 +341,8 @@ void mcf5249_vect_zero(ogg_int32_t *ptr, int n)
 
 #endif
 
+#endif
+
 #ifndef _V_CLIP_MATH
 #define _V_CLIP_MATH
 

apps/codecs/Tremor/block.c

@@ -262,11 +262,7 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
   vorbis_info *vi=v->vi;
   codec_setup_info *ci=(codec_setup_info *)vi->codec_setup;
   private_state *b=v->backend_state;
-#ifdef CPU_COLDFIRE
   int j;
-#else
-  int i,j;
-#endif
 
   if(v->pcm_current>v->pcm_returned  && v->pcm_returned!=-1)return(OV_EINVAL);
 
@@ -312,47 +308,25 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
           /* large/large */
           ogg_int32_t *pcm=v->pcm[j]+prevCenter;
           ogg_int32_t *p=vb->pcm[j];
-#ifdef CPU_COLDFIRE
+          vect_add(pcm, p, n1);
-          mcf5249_vect_add(pcm, p, n1);
-#else
-          for(i=0;i<n1;i++)
-            pcm[i]+=p[i]; 
-#endif
         }else{
           /* large/small */
           ogg_int32_t *pcm=v->pcm[j]+prevCenter+n1/2-n0/2;
           ogg_int32_t *p=vb->pcm[j];
-#ifdef CPU_COLDFIRE
+          vect_add(pcm, p, n0);
-          mcf5249_vect_add(pcm, p, n0);
-#else
-          for(i=0;i<n0;i++)
-            pcm[i]+=p[i];
-#endif
         }
       }else{
         if(v->W){
           /* small/large */
           ogg_int32_t *pcm=v->pcm[j]+prevCenter;
           ogg_int32_t *p=vb->pcm[j]+n1/2-n0/2;
-#ifdef CPU_COLDFIRE
+          vect_add(pcm, p, n0);
-          mcf5249_vect_add(pcm, p, n0);
+          vect_copy(&pcm[n0], &p[n0], n1/2-n0/2);
-          mcf5249_vect_copy(&pcm[n0], &p[n0], n1/2-n0/2);
-#else
-          for(i=0;i<n0;i++)
-            pcm[i]+=p[i];
-          for(;i<n1/2+n0/2;i++)
-            pcm[i]=p[i];
-#endif
         }else{
           /* small/small */
           ogg_int32_t *pcm=v->pcm[j]+prevCenter;
           ogg_int32_t *p=vb->pcm[j];
-#ifdef CPU_COLDFIRE
+          vect_add(pcm, p, n0);
-          mcf5249_vect_add(pcm, p, n0);
-#else 
-          for(i=0;i<n0;i++)
-            pcm[i]+=p[i];
-#endif
         }
       }
       
@@ -360,12 +334,7 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
       {
         ogg_int32_t *pcm=v->pcm[j]+thisCenter;
         ogg_int32_t *p=vb->pcm[j]+n;
-#ifdef CPU_COLDFIRE
+        vect_copy(pcm, p, n);
-        mcf5249_vect_copy(pcm, p, n);
-#else
-        for(i=0;i<n;i++)
-          pcm[i]=p[i];
-#endif
       }
     }
     

apps/codecs/Tremor/misc.h

@@ -151,6 +151,51 @@ static inline void XNPROD31(ogg_int32_t  a, ogg_int32_t  b,
   *y = MULT31(b, t) + MULT31(a, v);
 }
 #endif
+
+#ifndef _V_VECT_OPS
+#define _V_VECT_OPS
+
+static inline 
+void vect_add(ogg_int32_t *x, ogg_int32_t *y, int n)
+{
+  while (n>0) {
+    *x++ += *y++;
+    n--;
+  }
+}
+
+static inline 
+void vect_copy(ogg_int32_t *x, ogg_int32_t *y, int n)
+{
+  while (n>0) {
+    *x++ = *y++;
+    n--;
+  }
+}
+
+static inline 
+void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  while(n>0) {
+    *data = MULT31(*data, *window);
+    data++;
+    window++;
+    n--;
+  }
+}
+
+static inline
+void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  while(n>0) {
+    *data = MULT31(*data, *window);
+    data++;
+    window--;
+    n--;
+  }
+}
+#endif
+
 #endif
 
 #ifndef _V_CLIP_MATH

apps/codecs/Tremor/window.c

@@ -68,27 +68,11 @@ void _vorbis_apply_window(ogg_int32_t *d,const void *window_p[2],
   long rightbegin=n/2+n/4-rn/4;
   long rightend=rightbegin+rn/2;
 
-#ifdef CPU_COLDFIRE
   memset((void *)&d[0], 0, sizeof(ogg_int32_t)*leftbegin);
   /* mcf5249_vect_zero(&d[0], leftbegin); */
-  mcf5249_vect_mult_fw(&d[leftbegin], &window[lW][0], leftend-leftbegin);
+  vect_mult_fw(&d[leftbegin], &window[lW][0], leftend-leftbegin);
-  mcf5249_vect_mult_bw(&d[rightbegin], &window[nW][rn/2-1], rightend-rightbegin);
+  vect_mult_bw(&d[rightbegin], &window[nW][rn/2-1], rightend-rightbegin);
   memset((void *)&d[rightend], 0, sizeof(ogg_int32_t)*(n-rightend));
   /* mcf5249_vect_zero(&d[rightend], n-rightend); */
-#else  
-  int i,p;
-
-  for(i=0;i<leftbegin;i++)
-    d[i]=0;
-
-  for(p=0;i<leftend;i++,p++)
-    d[i]=MULT31(d[i],window[lW][p]);
-
-  for(i=rightbegin,p=rn/2-1;i<rightend;i++,p--)
-    d[i]=MULT31(d[i],window[nW][p]);
-
-  for(;i<n;i++)
-    d[i]=0;
-#endif
 }