Add 4-tap version of 2nd-pass ARMv6 MC filter.

The existing code applied a 6-tap filter with 0's on either end.
We're already paying the branch penalty to avoid computing the two
 extra columns needed as input to this filter.
We might as well save time computing the filter as well.
This reduces the inner loop from 21 instructions to 16, the number
 of loads per iteration from 4 to 1, and the number of multiplies
 from 7 to 4.
The gain in overall decoding performance, however, is small (less
 than 1%).

This change also means we now valgrind clean on ARMv6, which is
 its real purpose.
The errors reported here were valgrind's fault (it does not detect
 that 0 times an uninitialized value is initialized), but Julian
 Seward says it would slow down valgrind considerably to make such
 checks.
Speeding up libvpx rather, even by a small amount, seems a much
 better idea if only to enable proper valgrind checking of the
 rest of the codec.

Change-Id: Ifb376ea195e086b60f61daf1097d8910c4d8ff16

diff --git a/vp8/common/arm/armv6/filter_v6.asm b/vp8/common/arm/armv6/filter_v6.asm
index 8bc6d77..03b5bcc 100644 (file)
--- a/vp8/common/arm/armv6/filter_v6.asm
+++ b/vp8/common/arm/armv6/filter_v6.asm
@@ -11,6 +11,7 @@
 
     EXPORT  |vp8_filter_block2d_first_pass_armv6|
     EXPORT  |vp8_filter_block2d_second_pass_armv6|
+    EXPORT  |vp8_filter4_block2d_second_pass_armv6|
     EXPORT  |vp8_filter_block2d_first_pass_only_armv6|
     EXPORT  |vp8_filter_block2d_second_pass_only_armv6|
 
@@ -192,6 +193,64 @@
 
     ENDP
 
+;---------------------------------
+; r0    short         *src_ptr,
+; r1    unsigned char *output_ptr,
+; r2    unsigned int output_pitch,
+; r3    unsigned int cnt,
+; stack const short *vp8_filter
+;---------------------------------
+|vp8_filter4_block2d_second_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #36]                  ; vp8_filter address
+    mov     r7, r3, lsl #16                 ; height is top part of counter
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    add     lr, r1, r3                      ; save final destination pointer
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    pkhbt   r12, r5, r4                     ; pack the filter differently
+    pkhbt   r11, r6, r5
+    mov     r4, #0x40                       ; rounding factor (for smlad{x})
+
+|height_loop_2nd_4|
+    ldrd    r8, [r0, #-4]                   ; load the data
+    orr     r7, r7, r3, lsr #1              ; loop counter
+
+|width_loop_2nd_4|
+    ldr     r10, [r0, #4]!
+    smladx  r6, r9, r12, r4                 ; apply filter
+    pkhbt   r8, r9, r8
+    smlad   r5, r8, r12, r4
+    pkhbt   r8, r10, r9
+    smladx  r6, r10, r11, r6
+    sub     r7, r7, #1
+    smlad   r5, r8, r11, r5
+
+    mov     r8, r9                          ; shift the data for the next loop
+    mov     r9, r10
+
+    usat    r6, #8, r6, asr #7              ; shift and clamp
+    usat    r5, #8, r5, asr #7
+
+    strb    r5, [r1], r2                    ; the result is transposed back and stored
+    tst     r7, #0xff
+    strb    r6, [r1], r2
+
+    bne     width_loop_2nd_4
+
+    subs    r7, r7, #0x10000
+    add     r0, r0, #16                     ; update src for next loop
+    sub     r1, lr, r7, lsr #16             ; update dst for next loop
+
+    bne     height_loop_2nd_4
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
 ;------------------------------------
 ; r0    unsigned char *src_ptr
 ; r1    unsigned char *output_ptr,

diff --git a/vp8/common/arm/filter_arm.c b/vp8/common/arm/filter_arm.c
index 5ed4f80..4bfa3ab 100644 (file)
--- a/vp8/common/arm/filter_arm.c
+++ b/vp8/common/arm/filter_arm.c
@@ -50,6 +50,15 @@ extern void vp8_filter_block2d_second_pass_armv6
     const short *vp8_filter
 );
 
+extern void vp8_filter4_block2d_second_pass_armv6
+(
+    short         *src_ptr,
+    unsigned char *output_ptr,
+    unsigned int output_pitch,
+    unsigned int cnt,
+    const short *vp8_filter
+);
+
 extern void vp8_filter_block2d_first_pass_only_armv6
 (
     unsigned char *src_ptr,
@@ -107,12 +116,16 @@ void vp8_sixtap_predict_armv6
     {
         // Vfilter is a 4 tap filter
         if (yoffset & 0x1)
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 4, 7, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 4, VFilter);
+        }
         // Vfilter is 6 tap filter
         else
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 4, 9, HFilter);
-
-        vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 4, VFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 4, VFilter);
+        }
     }
 }
 
@@ -186,11 +199,15 @@ void vp8_sixtap_predict8x8_armv6
     else
     {
         if (yoffset & 0x1)
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 8, 11, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 8, VFilter);
+        }
         else
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 8, 13, HFilter);
-
-        vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 8, VFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 8, VFilter);
+        }
     }
 }
 
@@ -224,11 +241,15 @@ void vp8_sixtap_predict16x16_armv6
     else
     {
         if (yoffset & 0x1)
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 16, 19, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 16, VFilter);
+        }
         else
+        {
             vp8_filter_block2d_first_pass_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 16, 21, HFilter);
-
-        vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 16, VFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 16, VFilter);
+        }
     }
 
 }