From d428588fe3e1c31b968356993570f832253935ce Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <konstantinos.margaritis@vectorcamp.gr>
Date: Wed, 21 Sep 2022 08:49:11 +0000
Subject: [PATCH] Initial attempt for SVP64 asm version of
 vpx_get_mb_ss_svp64_real()

---
 media/video/libvpx/variance_svp64.c      | 11 ----------
 media/video/libvpx/variancefuncs_svp64.s | 28 +++++++++++++++++-------
 2 files changed, 20 insertions(+), 19 deletions(-)

diff --git a/media/video/libvpx/variance_svp64.c b/media/video/libvpx/variance_svp64.c
index 55e3f199..f0fc3f55 100644
--- 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 6d33d6b9..86541450 100644
--- 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
-- 
2.30.2