From 3d2b1cfa6e3307d3a97a055776ea342cd62f683e Mon Sep 17 00:00:00 2001
From: Nils Wallménius <nils@rockbox.org>
Date: Thu, 17 Jun 2010 16:49:39 +0000
Subject: ARMv6 vector mutiplication asm, speeds up vorbis decoding about
 0.1MHz on gigabeat S.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@26892 a1c6a512-1295-4272-9138-f99709370657
---
 apps/codecs/libtremor/asm_arm.h | 50 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)

(limited to 'apps/codecs/libtremor/asm_arm.h')

diff --git a/apps/codecs/libtremor/asm_arm.h b/apps/codecs/libtremor/asm_arm.h
index ada0604a3a..683de7bf64 100644
--- a/apps/codecs/libtremor/asm_arm.h
+++ b/apps/codecs/libtremor/asm_arm.h
@@ -112,6 +112,32 @@ void vect_add_left_right(ogg_int32_t *x, const ogg_int32_t *y, int n)
   } while (n);
 }
 
+#if ARM_ARCH >= 6
+static inline 
+void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  /* Note, mult_fw uses MULT31 */
+  do{
+    asm volatile (
+                  "ldmia %[d], {r0, r1, r2, r3};"
+                  "ldmia %[w]!, {r4, r5, r6, r7};"
+                  "smmul r0, r4, r0;"
+                  "smmul r1, r5, r1;"
+                  "smmul r2, r6, r2;"
+                  "smmul r3, r7, r3;"
+                  "mov   r0, r0, lsl #1;"
+                  "mov   r1, r1, lsl #1;"
+                  "mov   r2, r2, lsl #1;"
+                  "mov   r3, r3, lsl #1;"
+                  "stmia %[d]!, {r0, r1, r2, r3};"
+                  : [d] "+r" (data), [w] "+r" (window)
+                  : : "r0", "r1", "r2", "r3",
+                  "r4", "r5", "r6", "r7",
+                  "memory" );
+    n -= 4;
+  } while (n);
+}
+#else
 static inline 
 void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 {
@@ -136,7 +162,30 @@ void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
     n -= 4;
   } while (n);
 }
+#endif
 
+#if ARM_ARCH >= 6
+static inline
+void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
+{
+  /* NOTE mult_bw uses MULT_32 i.e. doesn't shift result left at end */
+  /* On ARM, we can do the shift at the same time as the overlap-add */
+  do{
+    asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
+                  "ldmda %[w]!, {r4, r5, r6, r7};"
+                  "smmul r0, r7, r0;"
+                  "smmul r1, r6, r1;"
+                  "smmul r2, r5, r2;"
+                  "smmul r3, r4, r3;"
+                  "stmia %[d]!, {r0, r1, r2, r3};"
+                  : [d] "+r" (data), [w] "+r" (window)
+                  : : "r0", "r1", "r2", "r3",
+                  "r4", "r5", "r6", "r7",
+                  "memory" );
+    n -= 4;
+  } while (n);
+}
+#else
 static inline
 void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
 {
@@ -157,6 +206,7 @@ void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
     n -= 4;
   } while (n);
 }
+#endif
 
 static inline void vect_copy(ogg_int32_t *x, const ogg_int32_t *y, int n)
 {
-- 
cgit v1.2.3