getting better, get rid of the ctr, group src/ref loads

diff --git a/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s
index 5ec8e0085edf827f00926ce4d211b3a6724cdd74..c607da69c0583416d66a8173c126a099a321a1e8 100644 (file)
--- a/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s
+++ b/media/video/libvpx/vpx_get4x4sse_cs_svp64_real.s
@@ -4,9 +4,9 @@
 .set ref_stride, 6
 .set sum, 7
 .set src, 10
-.set ref, 36
-.set diff, 52
-.set prod, 68
+.set ref, 26
+.set diff, 42
+.set prod, 58
 .set ctr, 9
 
        .machine libresoc
@@ -21,24 +21,25 @@ vpx_get4x4sse_cs_svp64_real:
        .cfi_startproc
        # Set sum to zero
        li sum, 0                               # Set sum to zero
-       li ctr, 4                               # Need 4 iterations of 4 elements
-       mtctr ctr                               # Set counter special register
+       sldi    src_stride, src_stride, 1       # strides are for 16-bit elements
+       sldi    ref_stride, ref_stride, 1       # we need to increase by bytes
        # Load 16 elements from src_ptr and ref_ptr, at groups of 4 with stride
        setvl   0,0,4,0,1,1                     # Set VL to 4 elements
        sv.lha  *src, 0(src_ptr)                # Load 4 ints from (src_ptr)
-       sv.lha  *ref, 0(ref_ptr)                # Load 4 ints from (ref_ptr)
        add     src_ptr, src_ptr, src_stride    # Advance src_ptr by src_stride
-       add     ref_ptr, ref_ptr, ref_stride    # Advance ref_ptr by ref_stride
-       sv.lha  *(src + 4), 0(src_ptr)
-       sv.lha  *(ref + 4), 0(ref_ptr)
+       sv.lha  *src + 4, 0(src_ptr)
        add     src_ptr, src_ptr, src_stride
-       add     ref_ptr, ref_ptr, ref_stride
-       sv.lha  *(src + 8), 0(src_ptr)
-       sv.lha  *(ref + 8), 0(ref_ptr)
+       sv.lha  *src + 8, 0(src_ptr)
        add     src_ptr, src_ptr, src_stride
+       sv.lha  *src + 12, 0(src_ptr)
+       setvl   0,0,4,0,1,1                     # Set VL to 4 elements
+       sv.lha  *ref, 0(ref_ptr)                # Load 4 ints from (ref_ptr)
+       add     ref_ptr, ref_ptr, ref_stride    # Advance ref_ptr by ref_stride
+       sv.lha  *ref + 4, 0(ref_ptr)
+       add     ref_ptr, ref_ptr, ref_stride
+       sv.lha  *ref + 8, 0(ref_ptr)
        add     ref_ptr, ref_ptr, ref_stride
-       sv.lha  *(src + 12), 0(src_ptr)
-       sv.lha  *(ref + 12), 0(ref_ptr)
+       sv.lha  *ref + 12, 0(ref_ptr)
 
        # now our values are in consecutive registers and we can set VL to 16 elements
        setvl           0,0,16,0,1,1