From f588d073b576aa30c9cad76b2e77160439ad2da5 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <konstantinos.margaritis@vectorcamp.gr>
Date: Wed, 21 Sep 2022 13:06:13 +0000
Subject: [PATCH] First form of fully working SVP64 version

---
 ...ncs_svp64.s => vpx_get_mb_ss_svp64_real.s} | 27 ++++++++++---------
 1 file changed, 15 insertions(+), 12 deletions(-)
 rename media/video/libvpx/{variancefuncs_svp64.s => vpx_get_mb_ss_svp64_real.s} (56%)

diff --git a/media/video/libvpx/variancefuncs_svp64.s b/media/video/libvpx/vpx_get_mb_ss_svp64_real.s
similarity index 56%
rename from media/video/libvpx/variancefuncs_svp64.s
rename to media/video/libvpx/vpx_get_mb_ss_svp64_real.s
index 86541450..c8cae746 100644
--- a/media/video/libvpx/variancefuncs_svp64.s
+++ b/media/video/libvpx/vpx_get_mb_ss_svp64_real.s
@@ -1,6 +1,7 @@
 .set in, 3
-.set vin, 20
-.set sum, 6
+.set src, 10
+.set prod, 50
+.set sum, 4
 .set ctr, 9
 
 
@@ -15,20 +16,22 @@
 vpx_get_mb_ss_svp64_real:
 .LFB0:
 	.cfi_startproc
-	addi 10, in ,-2
-	li in, 0
 	li sum, 0
-	li ctr, 4
+	li ctr, 8
 	mtctr ctr
-	setvl 0,0,64,0,1,1			# Set VL to 64 elements
+	setvl 0,0,32,0,1,1			# Set VL to 64 elements
 .L2:
-	# 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
+	# Load 32 ints from (in)
+	sv.lha	 	*src, 0(in)
+	# equivalent to: for (i = 0; i < 32; i++) vprod[i] = src[i] * src[i];
+	sv.mulld 	*prod, *src, *src
+	# equivalent to: for (i = 0; i < 32; i++) sum += prod[i];
+	sv.add/mr	sum, *prod, sum
+	addi in, in, 64
+#	rldicl in,ctr,0,32
 	bdnz .L2
+	li in, 0
+	addi in, sum, 0
 	blr
 	.long 0
 	.byte 0,0,0,0,0,0,0,0
-- 
2.30.2