From 257b04f91c1cfca3e05d83ed79b672781ead3e59 Mon Sep 17 00:00:00 2001
From: yuanhecai <yuanhecai@loongson.cn>
Date: Thu, 10 Oct 2024 20:37:43 +0800
Subject: [PATCH] loongarch: fix argon tests failure
---
src/loongarch/itx.S | 73 ++++++++++++++++++++++++++++++++-------------
1 file changed, 52 insertions(+), 21 deletions(-)
diff --git a/src/loongarch/itx.S b/src/loongarch/itx.S
index e8a6a5e8..9cd59686 100644
--- a/src/loongarch/itx.S
+++ b/src/loongarch/itx.S
@@ -545,15 +545,21 @@ endconst
vmulev_vmaddod_lsx vr5, vr18, vr20, vr20, vr16, \out3, \sz
vneg.h vr21, vr20
vmulev_vmaddod_lsx vr5, vr18, vr20, vr21, vr17, \out4, \sz
- vssrarni.h.w \out3, vr16, 12 // out3
+ vsrari.w vr16, vr16, 12
+ vsrari.w \out3, \out3, 12
+ vneg.w vr16, vr16
+ vneg.w \out3, \out3
+ vssrarni.h.w \out3, vr16, 0 // out3
vssrarni.h.w \out4, vr17, 12 // out4
- vneg.h \out3, \out3
vmulev_vmaddod_lsx vr19, vr23, vr20, vr20, vr16, \out2, \sz
vmulev_vmaddod_lsx vr19, vr23, vr20, vr21, vr17, \out5, \sz
- vssrarni.h.w \out5, vr17, 12 // out5
vssrarni.h.w \out2, vr16, 12 // out2
- vneg.h \out5, \out5
+ vsrari.w vr17, vr17, 12
+ vsrari.w \out5, \out5, 12
+ vneg.w vr17, vr17
+ vneg.w \out5, \out5
+ vssrarni.h.w \out5, vr17, 0 // out5
.endm
functionl inv_adst_8h_x8_lsx
@@ -1512,24 +1518,38 @@ endconst
vmulev_vmaddod_lsx vr17, vr7, vr20, vr20, vr6, vr10, \sz
vneg.h vr21, vr20
vmulev_vmaddod_lsx vr17, vr7, vr20, vr21, vr16, vr1, \sz
- vssrarni.h.w vr10, vr6, 12 // out[7]
vssrarni.h.w vr1, vr16, 12 // out[8]
- vneg.h vr10, vr10
+ vsrari.w vr6, vr6, 12
+ vsrari.w vr10, vr10, 12
+ vneg.w vr6, vr6
+ vneg.w vr10, vr10
+ vssrarni.h.w vr10, vr6, 0 // out[7]
vmulev_vmaddod_lsx vr0, vr8, vr20, vr21, vr16, vr17, \sz
vmulev_vmaddod_lsx vr0, vr8, vr20, vr20, vr6, vr7, \sz
- vssrarni.h.w vr17, vr16, 12 // out[11]
vssrarni.h.w vr7, vr6, 12 // out[4]
- vneg.h vr17, vr17
+ vsrari.w vr16, vr16, 12
+ vsrari.w vr17, vr17, 12
+ vneg.w vr16, vr16
+ vneg.w vr17, vr17
+ vssrarni.h.w vr17, vr16, 0 // out[11]
+
vmulev_vmaddod_lsx vr4, vr19, vr20, vr21, vr16, vr0, \sz
vmulev_vmaddod_lsx vr4, vr19, vr20, vr20, vr6, vr8, \sz
- vssrarni.h.w vr0, vr16, 12 // out[9]
vssrarni.h.w vr8, vr6, 12 // out[6]
- vneg.h vr0, vr0
+ vsrari.w vr16, vr16, 12
+ vsrari.w vr0, vr0, 12
+ vneg.w vr16, vr16
+ vneg.w vr0, vr0
+ vssrarni.h.w vr0, vr16, 0 // out[9]
+
vmulev_vmaddod_lsx vr11, vr12, vr20, vr20, vr6, vr4, \sz
vmulev_vmaddod_lsx vr11, vr12, vr20, vr21, vr16, vr19, \sz
- vssrarni.h.w vr4, vr6, 12 // out[5]
vssrarni.h.w vr19, vr16, 12 // out[10]
- vneg.h vr4, vr4
+ vsrari.w vr6, vr6, 12
+ vsrari.w vr4, vr4, 12
+ vneg.w vr6, vr6
+ vneg.w vr4, vr4
+ vssrarni.h.w vr4, vr6, 0 // out[5]
.ifc \txfm, adst
vor.v vr12, vr3, vr3
@@ -4664,9 +4684,12 @@ endfunc
xvilvl.w xr7, xr1, xr16
xvilvh.w xr10, xr10, xr6
xvilvh.w xr1, xr1, xr16
- xvssrarni.h.w xr10, xr17, 12 // out[7]
xvssrarni.h.w xr1, xr7, 12 // out[8]
- xvneg.h xr10, xr10
+ xvsrari.w xr17, xr17, 12
+ xvsrari.w xr10, xr10, 12
+ xvneg.w xr17, xr17
+ xvneg.w xr10, xr10
+ xvssrarni.h.w xr10, xr17, 0 // out[7]
xvmulev_xvmaddod_lasx xr0, xr8, xr20, xr21, xr16, xr17
xvmulev_xvmaddod_lasx xr0, xr8, xr20, xr20, xr6, xr7
@@ -4674,9 +4697,12 @@ endfunc
xvilvl.w xr8, xr7, xr6
xvilvh.w xr17, xr17, xr16
xvilvh.w xr7, xr7, xr6
- xvssrarni.h.w xr17, xr0, 12 // out[11]
xvssrarni.h.w xr7, xr8, 12 // out[4]
- xvneg.h xr17, xr17
+ xvsrari.w xr0, xr0, 12
+ xvsrari.w xr17, xr17, 12
+ xvneg.w xr0, xr0
+ xvneg.w xr17, xr17
+ xvssrarni.h.w xr17, xr0, 0 // out[11]
xvmulev_xvmaddod_lasx xr4, xr19, xr20, xr21, xr16, xr0
xvmulev_xvmaddod_lasx xr4, xr19, xr20, xr20, xr6, xr8
@@ -4684,19 +4710,24 @@ endfunc
xvilvl.w xr19, xr8, xr6
xvilvh.w xr0, xr0, xr16
xvilvh.w xr8, xr8, xr6
- xvssrarni.h.w xr0, xr4, 12 // out[9]
xvssrarni.h.w xr8, xr19, 12 // out[6]
- xvneg.h xr0, xr0
-
+ xvsrari.w xr4, xr4, 12
+ xvsrari.w xr0, xr0, 12
+ xvneg.w xr4, xr4
+ xvneg.w xr0, xr0
+ xvssrarni.h.w xr0, xr4, 0 // out[9]
xvmulev_xvmaddod_lasx xr11, xr12, xr20, xr20, xr6, xr4
xvmulev_xvmaddod_lasx xr11, xr12, xr20, xr21, xr16, xr19
xvilvl.w xr11, xr4, xr6
xvilvl.w xr12, xr19, xr16
xvilvh.w xr4, xr4, xr6
xvilvh.w xr19, xr19, xr16
- xvssrarni.h.w xr4, xr11, 12 // out[5]
xvssrarni.h.w xr19, xr12, 12 // out[10]
- xvneg.h xr4, xr4
+ xvsrari.w xr11, xr11, 12
+ xvsrari.w xr4, xr4, 12
+ xvneg.w xr11, xr11
+ xvneg.w xr4, xr4
+ xvssrarni.h.w xr4, xr11, 0 // out[5]
.endm
function inv_txfm_add_adst_adst_16x16_8bpc_lasx
--
GitLab