x86: Add a workaround for quirky AVX-512 hardware behavior

On Intel CPUs certain AVX-512 shuffle instructions incorrectly
flag the upper halves of YMM registers as in use when writing
to XMM registers, which may cause AVX/SSE state transitions.

This behavior is not documented and only occurs on physical
hardware, not when using the Intel SDE, so as far as I can tell
it appears to be a hardware bug.

Work around the issue by using EVEX-only registers. This avoids
the problem at the cost of a slightly larger code size.

4 files changed, 234 insertions, 234 deletions

diff --git a/src/x86/ipred16_avx512.asm b/src/x86/ipred16_avx512.asm
index 4a1b060..1a307ad 100644
--- a/src/x86/ipred16_avx512.asm
+++ b/src/x86/ipred16_avx512.asm
@@ -114,20 +114,20 @@ cglobal ipred_paeth_16bpc, 3, 7, 10, dst, stride, tl, w, h
     vbroadcasti32x4     m2, [tlq]
     pshufb              m2, m7      ; left
     PAETH                4, 5, 6
-    vextracti32x4     xmm1, m0, 2
-    vextracti32x4     xmm2, ym0, 1
-    vextracti32x4     xmm3, m0, 3
+    vextracti32x4      xm1, m0, 2
+    vextracti32x4      xm8, ym0, 1
+    vextracti32x4      xm9, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+r6       ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm8
+    movq   [dstq+r6       ], xm9
     sub                 hd, 8
     jl .w4_end
     lea               dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+r6       ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm8
+    movhps [dstq+r6       ], xm9
     lea               dstq, [dstq+strideq*4]
     jg .w4_loop
 .w4_end:
@@ -220,19 +220,19 @@ cglobal ipred_smooth_v_16bpc, 3, 7, 7, dst, stride, tl, w, h, weights, stride3
     pshufb               m3, m4
     pmulhrsw             m3, m5
     paddw                m3, m6
-    vextracti32x4      xmm0, m3, 3
-    vextracti32x4      xmm1, ym3, 1
-    vextracti32x4      xmm2, m3, 2
-    movhps [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
+    vextracti32x4       xm0, m3, 3
+    vextracti32x4       xm1, ym3, 1
+    vextracti32x4       xm2, m3, 2
+    movhps [dstq+strideq*0], xm0
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
     movhps [dstq+stride3q ], xm3
     add                  hq, 8
     jg .end
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
+    movq   [dstq+strideq*0], xm0
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
     movq   [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     jl .w4_loop
@@ -337,20 +337,20 @@ cglobal ipred_smooth_h_16bpc, 3, 7, 7, dst, stride, tl, w, h, stride3
     psubw                m0, m6          ; left - right
     pmulhrsw             m0, m5
     paddw                m0, m6
-    vextracti32x4      xmm1, m0, 2
-    vextracti32x4      xmm2, ym0, 1
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, m0, 2
+    vextracti32x4       xm2, ym0, 1
+    vextracti32x4       xm3, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     sub                  hd, 8*2
     jl .end
     lea                dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     jg .w4_loop
 .end:
@@ -472,11 +472,11 @@ cglobal ipred_smooth_16bpc, 3, 7, 16, dst, stride, tl, w, h, v_weights, stride3
     vpdpwssd            m0, m1, m6
     vpermb              m0, m14, m0
     pavgw              ym0, ym15
-    vextracti32x4     xmm1, ym0, 1
+    vextracti32x4      xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     lea               dstq, [dstq+strideq*4]
     add         v_weightsq, 4*4
     sub                 hd, 4*2
@@ -624,11 +624,11 @@ cglobal pal_pred_16bpc, 4, 7, 4, dst, stride, pal, idx, w, h, stride3
     pmovzxbw            ym0, [idxq]
     add                idxq, 16
     vpermw              ym0, ym0, ym3
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
     movhps [dstq+strideq*1], xm0
-    movq   [dstq+strideq*2], xmm1
-    movhps [dstq+stride3q ], xmm1
+    movq   [dstq+strideq*2], xm1
+    movhps [dstq+stride3q ], xm1
     lea                dstq, [dstq+strideq*4]
     sub                  hd, 4
     jg .w4
diff --git a/src/x86/ipred_avx512.asm b/src/x86/ipred_avx512.asm
index 050ec9b..38c86b5 100644
--- a/src/x86/ipred_avx512.asm
+++ b/src/x86/ipred_avx512.asm
@@ -242,9 +242,9 @@ cglobal ipred_dc_8bpc, 3, 7, 5, dst, stride, tl, w, h, stride3
     jmp                  wq
 .w8:
     movq               xmm1, [tlq+1]
-    vextracti32x4      xmm2, ym0, 1
+    vextracti32x4       xm2, ym0, 1
     vpdpbusd            xm0, xmm1, xm3
-    paddd              xmm2, xm0
+    paddd              xmm2, xm2, xm0
     punpckhqdq         xmm0, xmm2, xmm2
     paddd              xmm0, xmm2
     psrlq              xmm1, xmm0, 32
@@ -275,9 +275,9 @@ cglobal ipred_dc_8bpc, 3, 7, 5, dst, stride, tl, w, h, stride3
     jmp                  wq
 .w16:
     movu               xmm1, [tlq+1]
-    vextracti32x4      xmm2, ym0, 1
+    vextracti32x4       xm2, ym0, 1
     vpdpbusd            xm0, xmm1, xm3
-    paddd              xmm2, xm0
+    paddd              xmm2, xm2, xm0
     punpckhqdq         xmm0, xmm2, xmm2
     paddd              xmm0, xmm2
     psrlq              xmm1, xmm0, 32
@@ -309,8 +309,8 @@ cglobal ipred_dc_8bpc, 3, 7, 5, dst, stride, tl, w, h, stride3
 .w32:
     movu                ym1, [tlq+1]
     vpdpbusd            ym0, ym1, ym3
-    vextracti32x4      xmm1, ym0, 1
-    paddd              xmm1, xm0
+    vextracti32x4       xm1, ym0, 1
+    paddd              xmm1, xm1, xm0
     punpckhqdq         xmm0, xmm1, xmm1
     paddd              xmm0, xmm1
     psrlq              xmm1, xmm0, 32
@@ -345,8 +345,8 @@ cglobal ipred_dc_8bpc, 3, 7, 5, dst, stride, tl, w, h, stride3
     movu                ym2, [tlq+33]
     vpdpbusd            ym0, ym1, ym3
     vpdpbusd            ym0, ym2, ym3
-    vextracti32x4      xmm1, ym0, 1
-    paddd              xmm1, xm0
+    vextracti32x4       xm1, ym0, 1
+    paddd              xmm1, xm1, xm0
     punpckhqdq         xmm0, xmm1, xmm1
     paddd              xmm0, xmm1
     psrlq              xmm1, xmm0, 32
@@ -524,12 +524,12 @@ INIT_YMM avx512icl
     pextrd [dstq+stride3q ], xm0, 3
     sub                  hd, 8
     jl .w4_ret
-    vextracti32x4      xmm0, m0, 1
+    vextracti32x4       xm0, m0, 1
     lea                dstq, [dstq+strideq*4]
-    movd   [dstq+strideq*0], xmm0
-    pextrd [dstq+strideq*1], xmm0, 1
-    pextrd [dstq+strideq*2], xmm0, 2
-    pextrd [dstq+stride3q ], xmm0, 3
+    movd   [dstq+strideq*0], xm0
+    pextrd [dstq+strideq*1], xm0, 1
+    pextrd [dstq+strideq*2], xm0, 2
+    pextrd [dstq+stride3q ], xm0, 3
     lea                dstq, [dstq+strideq*4]
     jg .w4_loop
 .w4_ret:
@@ -545,20 +545,20 @@ INIT_ZMM avx512icl
     vpbroadcastq         m4, [tlq+hq-8]
     pshufb               m4, m9
     PAETH
-    vextracti32x4      xmm1, m0, 2
-    vextracti32x4      xmm2, ym0, 1
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, m0, 2
+    vextracti32x4       xm2, ym0, 1
+    vextracti32x4       xm3, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     sub                  hd, 8
     jl .w8_ret
     lea                dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     jg .w8_loop
 .w8_ret:
@@ -639,18 +639,18 @@ cglobal ipred_smooth_v_8bpc, 3, 7, 7, dst, stride, tl, w, h, weights, stride3
     pmaddubsw            m0, m2, m0
     paddw                m0, m3
     vpermb               m0, m6, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movd   [dstq+strideq*0], xm0
-    movd   [dstq+strideq*1], xmm1
+    movd   [dstq+strideq*1], xm1
     pextrd [dstq+strideq*2], xm0, 2
-    pextrd [dstq+stride3q ], xmm1, 2
+    pextrd [dstq+stride3q ], xm1, 2
     add                  hq, 8
     jg .ret
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 1
-    pextrd [dstq+strideq*1], xmm1, 1
+    pextrd [dstq+strideq*1], xm1, 1
     pextrd [dstq+strideq*2], xm0, 3
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+stride3q ], xm1, 3
     lea                dstq, [dstq+strideq*4]
     jl .w4_loop
 .ret:
@@ -669,11 +669,11 @@ cglobal ipred_smooth_v_8bpc, 3, 7, 7, dst, stride, tl, w, h, weights, stride3
     pmaddubsw            m0, m2, m0
     paddw                m0, m3
     vpermb               m0, m6, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     lea                dstq, [dstq+strideq*4]
     add                  hq, 4
     jl .w8_loop
@@ -785,18 +785,18 @@ cglobal ipred_smooth_h_8bpc, 4, 7, 11, dst, stride, tl, w, h, stride3
     paddw                m0, m2
     paddw                m0, m1
     vpermb               m0, m8, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movd   [dstq+strideq*0], xm0
-    movd   [dstq+strideq*1], xmm1
+    movd   [dstq+strideq*1], xm1
     pextrd [dstq+strideq*2], xm0, 2
-    pextrd [dstq+stride3q ], xmm1, 2
+    pextrd [dstq+stride3q ], xm1, 2
     sub                  hd, 8
     jl .ret
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 1
-    pextrd [dstq+strideq*1], xmm1, 1
+    pextrd [dstq+strideq*1], xm1, 1
     pextrd [dstq+strideq*2], xm0, 3
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+stride3q ], xm1, 3
     lea                dstq, [dstq+strideq*4]
     jg .w4_loop
 .ret:
@@ -815,11 +815,11 @@ cglobal ipred_smooth_h_8bpc, 4, 7, 11, dst, stride, tl, w, h, stride3
     paddw                m0, m2
     paddw                m0, m1
     vpermb               m0, m8, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     lea                dstq, [dstq+strideq*4]
     sub                  hd, 4
     jg .w8_loop
@@ -937,18 +937,18 @@ cglobal ipred_smooth_8bpc, 4, 7, 16, dst, stride, tl, w, h, v_weights, stride3
     paddw                m1, m2
     pavgw                m0, m1
     vpermb               m0, m11, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movd   [dstq+strideq*0], xm0
-    movd   [dstq+strideq*1], xmm1
+    movd   [dstq+strideq*1], xm1
     pextrd [dstq+strideq*2], xm0, 2
-    pextrd [dstq+stride3q ], xmm1, 2
+    pextrd [dstq+stride3q ], xm1, 2
     sub                  hd, 8
     jl .ret
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 1
-    pextrd [dstq+strideq*1], xmm1, 1
+    pextrd [dstq+strideq*1], xm1, 1
     pextrd [dstq+strideq*2], xm0, 3
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+stride3q ], xm1, 3
     lea                dstq, [dstq+strideq*4]
     jg .w4_loop
 .ret:
@@ -978,11 +978,11 @@ cglobal ipred_smooth_8bpc, 4, 7, 16, dst, stride, tl, w, h, v_weights, stride3
     paddw                m1, m2
     pavgw                m0, m1
     vpermb               m0, m11, m0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     lea                dstq, [dstq+strideq*4]
     sub                  hd, 4
     jg .w8_loop
diff --git a/src/x86/mc16_avx512.asm b/src/x86/mc16_avx512.asm
index e83b18a..f8d1aaa 100644
--- a/src/x86/mc16_avx512.asm
+++ b/src/x86/mc16_avx512.asm
@@ -3615,32 +3615,32 @@ ALIGN function_align
 .w4:
     movq   [dstq          ], xm0
     movhps [dstq+strideq*1], xm0
-    vextracti32x4      xmm0, ym0, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    vextracti32x4       xm2, ym0, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
     cmp                  hd, 8
     jl .w4_end
-    vextracti32x4      xmm0, m0, 2
+    vextracti32x4       xm2, m0, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq          ], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m0, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq          ], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm0, m0, 3
+    movq   [dstq+strideq*2], xm0
+    movhps [dstq+stride3q ], xm0
     je .w4_end
     lea                dstq, [dstq+strideq*4]
     movq   [dstq          ], xm1
     movhps [dstq+strideq*1], xm1
-    vextracti32x4      xmm0, ym1, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
-    vextracti32x4      xmm0, m1, 2
+    vextracti32x4       xm0, ym1, 1
+    movq   [dstq+strideq*2], xm0
+    movhps [dstq+stride3q ], xm0
+    vextracti32x4       xm0, m1, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq          ], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m1, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq          ], xm0
+    movhps [dstq+strideq*1], xm0
+    vextracti32x4       xm1, m1, 3
+    movq   [dstq+strideq*2], xm1
+    movhps [dstq+stride3q ], xm1
 .w4_end:
     RET
 .w8_loop:
@@ -3860,33 +3860,33 @@ cglobal w_mask_420_16bpc, 4, 8, 16, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vpermb               m3, m15, m3
     movq   [dstq+strideq*0], xm0
     movhps [dstq+strideq*1], xm0
-    vextracti32x4      xmm0, ym0, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    vextracti32x4       xm2, ym0, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
     mova            [maskq], xm3
     cmp                  hd, 8
     jl .w4_end
-    vextracti32x4      xmm0, m0, 2
+    vextracti32x4       xm2, m0, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m0, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm0, m0, 3
+    movq   [dstq+strideq*2], xm0
+    movhps [dstq+stride3q ], xm0
     je .w4_end
     lea                dstq, [dstq+strideq*4]
     movq   [dstq+strideq*0], xm1
     movhps [dstq+strideq*1], xm1
-    vextracti32x4      xmm0, ym1, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
-    vextracti32x4      xmm0, m1, 2
+    vextracti32x4       xm2, ym1, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
+    vextracti32x4       xm2, m1, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m1, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm1, m1, 3
+    movq   [dstq+strideq*2], xm1
+    movhps [dstq+stride3q ], xm1
 .w4_end:
     RET
 .w8:
@@ -4090,32 +4090,32 @@ cglobal w_mask_422_16bpc, 4, 8, 15, dst, stride, tmp1, tmp2, w, h, mask, stride3
 .w4:
     movq   [dstq+strideq*0], xm0
     movhps [dstq+strideq*1], xm0
-    vextracti32x4      xmm0, ym0, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    vextracti32x4       xm2, ym0, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
     cmp                  hd, 8
     jl .w4_end
-    vextracti32x4      xmm0, m0, 2
+    vextracti32x4       xm2, m0, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m0, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm0, m0, 3
+    movq   [dstq+strideq*2], xm0
+    movhps [dstq+stride3q ], xm0
     je .w4_end
     lea                dstq, [dstq+strideq*4]
     movq   [dstq+strideq*0], xm1
     movhps [dstq+strideq*1], xm1
-    vextracti32x4      xmm0, ym1, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
-    vextracti32x4      xmm0, m1, 2
+    vextracti32x4       xm2, ym1, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
+    vextracti32x4       xm2, m1, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m1, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm1, m1, 3
+    movq   [dstq+strideq*2], xm1
+    movhps [dstq+stride3q ], xm1
 .w4_end:
     RET
 .w8_loop:
@@ -4249,32 +4249,32 @@ cglobal w_mask_444_16bpc, 4, 8, 13, dst, stride, tmp1, tmp2, w, h, mask, stride3
 .w4:
     movq   [dstq+strideq*0], xm0
     movhps [dstq+strideq*1], xm0
-    vextracti32x4      xmm0, ym0, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    vextracti32x4       xm2, ym0, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
     cmp                  hd, 8
     jl .w4_end
-    vextracti32x4      xmm0, m0, 2
+    vextracti32x4       xm2, m0, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m0, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm0, m0, 3
+    movq   [dstq+strideq*2], xm0
+    movhps [dstq+stride3q ], xm0
     je .w4_end
     lea                dstq, [dstq+strideq*4]
     movq   [dstq+strideq*0], xm1
     movhps [dstq+strideq*1], xm1
-    vextracti32x4      xmm0, ym1, 1
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
-    vextracti32x4      xmm0, m1, 2
+    vextracti32x4       xm2, ym1, 1
+    movq   [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm2
+    vextracti32x4       xm2, m1, 2
     lea                dstq, [dstq+strideq*4]
-    movq   [dstq+strideq*0], xmm0
-    movhps [dstq+strideq*1], xmm0
-    vextracti32x4      xmm0, m1, 3
-    movq   [dstq+strideq*2], xmm0
-    movhps [dstq+stride3q ], xmm0
+    movq   [dstq+strideq*0], xm2
+    movhps [dstq+strideq*1], xm2
+    vextracti32x4       xm1, m1, 3
+    movq   [dstq+strideq*2], xm1
+    movhps [dstq+stride3q ], xm1
 .w4_end:
     RET
 .w8_loop:
diff --git a/src/x86/mc_avx512.asm b/src/x86/mc_avx512.asm
index eb3ca1c..7897f1d 100644
--- a/src/x86/mc_avx512.asm
+++ b/src/x86/mc_avx512.asm
@@ -449,9 +449,9 @@ cglobal put_bilin_8bpc, 4, 8, 0, dst, ds, src, ss, w, h, mxy
     pshufb              ym0, ym4
     pmaddubsw           ym0, ym5
     pmulhrsw            ym0, ym3
-    vpmovuswb          xmm0, ym0
-    movq       [dstq+dsq*0], xmm0
-    movhps     [dstq+dsq*1], xmm0
+    vpmovuswb           xm0, ym0
+    movq       [dstq+dsq*0], xm0
+    movhps     [dstq+dsq*1], xm0
     lea                dstq, [dstq+dsq*2]
     sub                  hd, 2
     jg .h_w8
@@ -755,9 +755,9 @@ cglobal put_bilin_8bpc, 4, 8, 0, dst, ds, src, ss, w, h, mxy
     pmulhw              ym1, ym6
     paddw               ym1, ym2
     pmulhrsw            ym1, ym7
-    vpmovuswb          xmm1, ym1
-    movq       [dstq+dsq*0], xmm1
-    movhps     [dstq+dsq*1], xmm1
+    vpmovuswb           xm1, ym1
+    movq       [dstq+dsq*0], xm1
+    movhps     [dstq+dsq*1], xm1
     lea                dstq, [dstq+dsq*2]
     sub                  hd, 2
     jg .hv_w8_loop
@@ -1588,13 +1588,13 @@ cglobal put_8tap_8bpc, 4, 9, 0, dst, ds, src, ss, w, h, mx, my, ss3
     jg .h_w4_loop
     RET
 .h_w8:
-    movu               xmm0, [srcq+ssq*0]
-    vinserti32x4        ym0, ymm0, [srcq+ssq*1], 1
+    movu                xm0, [srcq+ssq*0]
+    vinserti32x4        ym0, [srcq+ssq*1], 1
     lea                srcq, [srcq+ssq*2]
     WRAP_YMM PUT_8TAP_H   0, 1, 2, 3
-    vpmovuswb          xmm0, ym0
-    movq       [dstq+dsq*0], xmm0
-    movhps     [dstq+dsq*1], xmm0
+    vpmovuswb           xm0, ym0
+    movq       [dstq+dsq*0], xm0
+    movhps     [dstq+dsq*1], xm0
     lea                dstq, [dstq+dsq*2]
     sub                  hd, 2
     jg .h_w8
@@ -3308,17 +3308,17 @@ ALIGN function_align
     cmp                  hd, 8
     jg .w4_h16
     WRAP_YMM %1           0
-    vextracti32x4      xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movd   [dstq          ], xm0
     pextrd [dstq+strideq*1], xm0, 1
-    movd   [dstq+strideq*2], xmm1
-    pextrd [dstq+stride3q ], xmm1, 1
+    movd   [dstq+strideq*2], xm1
+    pextrd [dstq+stride3q ], xm1, 1
     jl .w4_ret
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq          ], xm0, 2
     pextrd [dstq+strideq*1], xm0, 3
-    pextrd [dstq+strideq*2], xmm1, 2
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+strideq*2], xm1, 2
+    pextrd [dstq+stride3q ], xm1, 3
 .w4_ret:
     RET
 .w4_h16:
@@ -3332,29 +3332,29 @@ ALIGN function_align
     cmp                  hd, 4
     jne .w8_h8
     WRAP_YMM %1           0
-    vextracti128       xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq          ], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     RET
 .w8_loop:
     %1_INC_PTR            2
     lea                dstq, [dstq+strideq*4]
 .w8_h8:
     %1                    0
-    vextracti32x4      xmm1, ym0, 1
-    vextracti32x4      xmm2, m0, 2
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, ym0, 1
+    vextracti32x4       xm2, m0, 2
+    vextracti32x4       xm3, m0, 3
     movq   [dstq          ], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     movhps [dstq          ], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     sub                  hd, 8
     jg .w8_loop
     RET
@@ -3415,8 +3415,8 @@ ALIGN function_align
     paddw                m0, [tmp2q+(%1+0)*mmsize]
     mova                 m1, [tmp1q+(%1+1)*mmsize]
     paddw                m1, [tmp2q+(%1+1)*mmsize]
-    pmulhrsw             m0, m2
-    pmulhrsw             m1, m2
+    pmulhrsw             m0, m4
+    pmulhrsw             m1, m4
     packuswb             m0, m1
 %endmacro
 
@@ -3425,13 +3425,13 @@ ALIGN function_align
     add               tmp2q, %1*mmsize
 %endmacro
 
-cglobal avg_8bpc, 4, 7, 3, dst, stride, tmp1, tmp2, w, h, stride3
+cglobal avg_8bpc, 4, 7, 5, dst, stride, tmp1, tmp2, w, h, stride3
 %define base r6-avg_avx512icl_table
     lea                  r6, [avg_avx512icl_table]
     tzcnt                wd, wm
     movifnidn            hd, hm
     movsxd               wq, dword [r6+wq*4]
-    vpbroadcastd         m2, [base+pw_1024]
+    vpbroadcastd         m4, [base+pw_1024]
     add                  wq, r6
     BIDIR_FN            AVG
 
@@ -3573,17 +3573,17 @@ cglobal w_mask_420_8bpc, 4, 8, 16, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vinserti128         ym5, [wm_420_perm4+32], 1
     vpermb              ym4, ym5, ym4
     vpdpbusd            ym8, ym4, ym9
-    vextracti128       xmm1, m0, 1
+    vextracti32x4       xm1, m0, 1
     movd   [dstq+strideq*0], xm0
     pextrd [dstq+strideq*1], xm0, 1
-    movd   [dstq+strideq*2], xmm1
-    pextrd [dstq+stride3q ], xmm1, 1
+    movd   [dstq+strideq*2], xm1
+    pextrd [dstq+stride3q ], xm1, 1
     jl .w4_end
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 2
     pextrd [dstq+strideq*1], xm0, 3
-    pextrd [dstq+strideq*2], xmm1, 2
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+strideq*2], xm1, 2
+    pextrd [dstq+stride3q ], xm1, 3
 .w4_end:
     vpermb              ym8, ym10, ym8
     movq            [maskq], xm8
@@ -3609,11 +3609,11 @@ cglobal w_mask_420_8bpc, 4, 8, 16, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vpdpbusd            ym8, ym4, ym9
     vpermb               m8, m10, m8
     mova            [maskq], xm8
-    vextracti128       xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     RET
 .w8_loop:
     add               tmp1q, 128
@@ -3627,18 +3627,18 @@ cglobal w_mask_420_8bpc, 4, 8, 16, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vpdpbusd             m1, m4, m9
     vpermb               m1, m10, m1
     mova            [maskq], xm1
-    vextracti32x4      xmm1, ym0, 1
-    vextracti32x4      xmm2, m0, 2
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, ym0, 1
+    vextracti32x4       xm2, m0, 2
+    vextracti32x4       xm3, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     sub                  hd, 8
     jg .w8_loop
     RET
@@ -3766,17 +3766,17 @@ cglobal w_mask_422_8bpc, 4, 8, 14, dst, stride, tmp1, tmp2, w, h, mask, stride3
     movhps             xm10, [wm_422_mask+16]
     vpdpwssd            ym8, ym4, ym9
     vpermb              ym8, ym10, ym8
-    vextracti128       xmm1, m0, 1
+    vextracti32x4       xm1, m0, 1
     movd   [dstq+strideq*0], xm0
     pextrd [dstq+strideq*1], xm0, 1
-    movd   [dstq+strideq*2], xmm1
-    pextrd [dstq+stride3q ], xmm1, 1
+    movd   [dstq+strideq*2], xm1
+    pextrd [dstq+stride3q ], xm1, 1
     jl .w4_end
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 2
     pextrd [dstq+strideq*1], xm0, 3
-    pextrd [dstq+strideq*2], xmm1, 2
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+strideq*2], xm1, 2
+    pextrd [dstq+stride3q ], xm1, 3
 .w4_end:
     pand                xm8, xm11
     mova            [maskq], xm8
@@ -3801,11 +3801,11 @@ cglobal w_mask_422_8bpc, 4, 8, 14, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vpermb              ym8, ym10, ym8
     pand                xm8, xm11
     mova            [maskq], xm8
-    vextracti128       xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     RET
 .w8_loop:
     add               tmp1q, 128
@@ -3819,18 +3819,18 @@ cglobal w_mask_422_8bpc, 4, 8, 14, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vpermb               m1, m10, m1
     pand                ym1, ym11
     mova            [maskq], ym1
-    vextracti32x4      xmm1, ym0, 1
-    vextracti32x4      xmm2, m0, 2
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, ym0, 1
+    vextracti32x4       xm2, m0, 2
+    vextracti32x4       xm3, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     sub                  hd, 8
     jg .w8_loop
     RET
@@ -3936,17 +3936,17 @@ cglobal w_mask_444_8bpc, 4, 8, 12, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vinserti128         ym8, [wm_444_mask+32], 1
     vpermb              ym4, ym8, ym4
     mova            [maskq], ym4
-    vextracti128       xmm1, m0, 1
+    vextracti32x4      xm1, m0, 1
     movd   [dstq+strideq*0], xm0
     pextrd [dstq+strideq*1], xm0, 1
-    movd   [dstq+strideq*2], xmm1
-    pextrd [dstq+stride3q ], xmm1, 1
+    movd   [dstq+strideq*2], xm1
+    pextrd [dstq+stride3q ], xm1, 1
     jl .w4_end
     lea                dstq, [dstq+strideq*4]
     pextrd [dstq+strideq*0], xm0, 2
     pextrd [dstq+strideq*1], xm0, 3
-    pextrd [dstq+strideq*2], xmm1, 2
-    pextrd [dstq+stride3q ], xmm1, 3
+    pextrd [dstq+strideq*2], xm1, 2
+    pextrd [dstq+stride3q ], xm1, 3
 .w4_end:
     RET
 .w4_h16:
@@ -3965,11 +3965,11 @@ cglobal w_mask_444_8bpc, 4, 8, 12, dst, stride, tmp1, tmp2, w, h, mask, stride3
     vinserti128         ym8, [wm_444_mask+32], 1
     vpermb              ym4, ym8, ym4
     mova            [maskq], ym4
-    vextracti128       xmm1, ym0, 1
+    vextracti32x4       xm1, ym0, 1
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
+    movq   [dstq+strideq*1], xm1
     movhps [dstq+strideq*2], xm0
-    movhps [dstq+stride3q ], xmm1
+    movhps [dstq+stride3q ], xm1
     RET
 .w8_loop:
     add               tmp1q, 128
@@ -3980,18 +3980,18 @@ cglobal w_mask_444_8bpc, 4, 8, 12, dst, stride, tmp1, tmp2, w, h, mask, stride3
     W_MASK                0, 4, 0, 1, 1
     vpermb               m4, m8, m4
     mova            [maskq], m4
-    vextracti32x4      xmm1, ym0, 1
-    vextracti32x4      xmm2, m0, 2
-    vextracti32x4      xmm3, m0, 3
+    vextracti32x4       xm1, ym0, 1
+    vextracti32x4       xm2, m0, 2
+    vextracti32x4       xm3, m0, 3
     movq   [dstq+strideq*0], xm0
-    movq   [dstq+strideq*1], xmm1
-    movq   [dstq+strideq*2], xmm2
-    movq   [dstq+stride3q ], xmm3
+    movq   [dstq+strideq*1], xm1
+    movq   [dstq+strideq*2], xm2
+    movq   [dstq+stride3q ], xm3
     lea                dstq, [dstq+strideq*4]
     movhps [dstq+strideq*0], xm0
-    movhps [dstq+strideq*1], xmm1
-    movhps [dstq+strideq*2], xmm2
-    movhps [dstq+stride3q ], xmm3
+    movhps [dstq+strideq*1], xm1
+    movhps [dstq+strideq*2], xm2
+    movhps [dstq+stride3q ], xm3
     sub                  hd, 8
     jg .w8_loop
     RET