Initial attempt for SVP64 asm version of vpx_get_mb_ss_svp64_real()

diff --git a/media/video/libvpx/variance_svp64.c b/media/video/libvpx/variance_svp64.c
index 55e3f199716cde265a1e53c915bf7e154d08a1b2..f0fc3f5544d70f60edd683fe9b63155f2aa07c17 100644 (file)
--- a/media/video/libvpx/variance_svp64.c
+++ b/media/video/libvpx/variance_svp64.c
@@ -39,17 +39,6 @@ uint32_t vpx_get4x4sse_cs_svp64(const uint8_t *src_ptr, int src_stride,
   return distortion;
 }
 
-/*
-uint32_t vpx_get_mb_ss_svp64(const int16_t *src_ptr) {
-  unsigned int i, sum = 0;
-
-  for (i = 0; i < 256; ++i) {
-    sum += src_ptr[i] * src_ptr[i];
-  }
-
-  return sum;
-}*/
-
 static void variance_svp64(const uint8_t *src_ptr, int src_stride,
                      const uint8_t *ref_ptr, int ref_stride, int w, int h,
                      uint32_t *sse, int *sum) {

diff --git a/media/video/libvpx/variancefuncs_svp64.s b/media/video/libvpx/variancefuncs_svp64.s
index 6d33d6b9c5c4ad7e825bd46b6ef5233d0edf02ff..865414500cf589458a1f9016ef9cfdf6a9786236 100644 (file)
--- a/media/video/libvpx/variancefuncs_svp64.s
+++ b/media/video/libvpx/variancefuncs_svp64.s
@@ -1,4 +1,12 @@
+.set in, 3
+.set vin, 20
+.set sum, 6
+.set ctr, 9
+
+
+       .machine libresoc
        .file   "variancefuncs_svp64.c"
+       .text
        .abiversion 2
        .section        ".text"
        .align 2
@@ -7,15 +15,19 @@
 vpx_get_mb_ss_svp64_real:
 .LFB0:
        .cfi_startproc
-       addi 10,3,-2
-       li 3,0
-       li 9,256
-       mtctr 9
+       addi 10, in ,-2
+       li in, 0
+       li sum, 0
+       li ctr, 4
+       mtctr ctr
+       setvl 0,0,64,0,1,1                      # Set VL to 64 elements
 .L2:
-       lhau 9,2(10)
-       mullw 9,9,9
-       add 9,9,3
-       rldicl 3,9,0,32
+       # Load 64 ints from (in)
+       sv.lha          *vin, 0(in)
+       # equivalent to: for (i = 0; i < 64; i++) sum += in[i] * in[i];
+       sv.maddld       sum, *vin, *vin, sum
+       addi in, in, 16
+       rldicl in,ctr,0,32
        bdnz .L2
        blr
        .long 0