From aaf3c81405e19ae5cda20ba12fc94646c07ec975 Mon Sep 17 00:00:00 2001
From: Konstantinos Margaritis <konstantinos.margaritis@vectorcamp.gr>
Date: Sun, 25 Sep 2022 16:56:08 +0000
Subject: [PATCH] Fixed SVP64 implentation

---
 media/video/libvpx/variance_svp64_real.s | 45 ++++++++++++------------
 1 file changed, 22 insertions(+), 23 deletions(-)

diff --git a/media/video/libvpx/variance_svp64_real.s b/media/video/libvpx/variance_svp64_real.s
index 4242f68e..046f68be 100644
--- a/media/video/libvpx/variance_svp64_real.s
+++ b/media/video/libvpx/variance_svp64_real.s
@@ -13,9 +13,9 @@
 .set ref_col, 15
 .set row, 16
 .set src, 20
-.set ref, 36
-.set diff, 52
-.set prod, 68
+.set ref, 24
+.set diff, 28
+.set prod, 32
 
 	.machine libresoc
 	.file	"variance_svp64_real.c"
@@ -27,45 +27,44 @@
 variance_svp64_real:
 .LFB0:
 	.cfi_startproc
-	# Set sum to zero
-	li sum, 0				# Set sum to zero
-	li sse, 0				# Set sse to zero
-	mr row, height				# Set row to height
+	li	sum, 0				# Set sum to zero
+	li	sse, 0				# Set sse to zero
+	li	row, 0				# Set row to zero
 	sldi	src_stride, src_stride, 1	# strides are for 16-bit elements
 	sldi	ref_stride, ref_stride, 1	# we need to increase by bytes
-	srdi	ctr, width, 2
-	mtctr	ctr
+	srdi	width, width, 2			# We load groups of 4
 	setvl	0,0,4,0,1,1			# Set VL to 4 elements
 
 .L1:	# outer loop: for (r=0; r < h; r++)
-
-.L2:	# inner loop: for (c=0; c < w; c += 4)
-	# Load 4 elements from src_ptr and ref_ptr, at groups of 4
 	mr	src_col, src_ptr		# Temporary variables
 	mr	ref_col, ref_ptr
+	mr	ctr, width			# Set up CTR to width/4 -1 on each row
+	mtctr	ctr				# Set up counter
+.L2:	# inner loop: for (c=0; c < w; c += 4)
+	# Load 4 elements from src_ptr and ref_ptr
 	sv.lha	*src, 0(src_col)		# Load 4 ints from (src_ptr)
 	sv.lha	*ref, 0(ref_col)		# Load 4 ints from (ref_ptr)
-	addi	src_col, src_col, 8		# Increment src, ref by 8 bytes
-	addi	ref_col, ref_col, 8
 
 	# equivalent to: for (i = 0; i < 4; i++) diff[i] = src[i] - ref[i];
-	sv.subf		*diff, *src, *ref
+	sv.subf		*diff, *ref, *src
 	# equivalent to: for (i = 0; i < 4; i++) prod[i] = diff[i] * diff[i];
 	sv.mulld 	*prod, *diff, *diff
 	# equivalent to: for (i = 0; i < 4; i++) sum += diff[i];
 	sv.add/mr	sum, *diff, sum
-	# equivalent to: for (i = 0; i < 4; i++) sum += diff[i];
+	# equivalent to: for (i = 0; i < 4; i++) sse += diff[i]*diff[i];
 	sv.add/mr	sse, *prod, sse
 
-	bdnz .L2				# Loop until CTR is zero
+	addi	src_col, src_col, 8		# Increment src, ref by 8 bytes
+	addi	ref_col, ref_col, 8
+	bdnz	.L2				# Loop until CTR is zero
+
 	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
-
-	subi row, row, 1			# Subtract 1 from row
-	cmpwi cr1, row, 0			# Is row zero?
-	bne cr1, .L1				# Go back to L1 if not done
-	std sum, 0(sum_ptr)			# Set (sum_ptr) to sum
-	std sse, 0(sse_ptr)			# Set (sum_ptr) to sum
+	addi	row, row, 1			# Add 1 to row
+	cmpw	cr1, row, height		# Is row equal to height?
+	bne	cr1, .L1			# Go back to L1 if not done
+	std	sum, 0(sum_ptr)			# Set (sum_ptr) to sum
+	std	sse, 0(sse_ptr)			# Set (sum_ptr) to sum
 	blr
 	.long 0
 	.byte 0,0,0,0,0,3,0,0
-- 
2.30.2