diff --git a/test/avx10/amx.txt b/test/avx10/amx.txt
index 8752c7e..74cf06d 100644
--- a/test/avx10/amx.txt
+++ b/test/avx10/amx.txt
@@ -1,3 +1,19 @@
+ldtilecfg(ptr[rax + rcx * 4 + 64]);
+sttilecfg(ptr[rsp + rax * 8 + 128]);
+tileloadd(tmm3, ptr[rdi + rdx * 2 + 8]);
+tileloaddt1(tmm4, ptr[r8 + r9 + 32]);
+tilerelease();
+tilestored(ptr[r10 + r11 * 2 + 32], tmm2);
+tilezero(tmm7);
+tdpbssd(tmm1, tmm2, tmm3);
+tdpbsud(tmm2, tmm3, tmm4);
+tdpbusd(tmm3, tmm4, tmm5);
+tdpbuud(tmm4, tmm5, tmm6);
+tdpfp16ps(tmm5, tmm6, tmm7);
+tdpbf16ps(tmm5, tmm6, tmm7);
+tileloadd(tmm1, ptr[r8+r8]);
+tileloadd(tmm1, ptr[rax+rcx*4]);
+tileloadd(tmm1, ptr[r8+r9*1+0x40]);
 tileloaddrs(tmm3, ptr[rdi + rdx * 2 + 8]);
 tileloaddrst1(tmm4, ptr[r8 + r9 + 32]);
 
diff --git a/test/avx10/old.txt b/test/avx10/old.txt
index f5a143c..19a4995 100644
--- a/test/avx10/old.txt
+++ b/test/avx10/old.txt
@@ -191,21 +191,6 @@
 vdpbf16ps(xmm0 | k1, xmm1, ptr [rax + 64]);
 vdpbf16ps(ymm0 | k1, ymm1, ptr [rax + 64]);
 vdpbf16ps(zmm0 | k1, zmm1, ptr [rax + 64]);
-ldtilecfg(ptr[rax + rcx * 4 + 64]);
-sttilecfg(ptr[rsp + rax * 8 + 128]);
-tileloadd(tmm3, ptr[rdi + rdx * 2 + 8]);
-tileloaddt1(tmm4, ptr[r8 + r9 + 32]);
-tilerelease();
-tilestored(ptr[r10 + r11 * 2 + 32], tmm2);
-tilezero(tmm7);
-tdpbssd(tmm1, tmm2, tmm3);
-tdpbsud(tmm2, tmm3, tmm4);
-tdpbusd(tmm3, tmm4, tmm5);
-tdpbuud(tmm4, tmm5, tmm6);
-tdpbf16ps(tmm5, tmm6, tmm7);
-tileloadd(tmm1, ptr[r8+r8]);
-tileloadd(tmm1, ptr[rax+rcx*4]);
-tileloadd(tmm1, ptr[r8+r9*1+0x40]);
 vaddph(zmm0, zmm1, ptr[rax+64]);
 vaddph(ymm0, ymm1, ptr[rax+64]);
 vaddph(xmm0, xmm1, ptr[rax+64]);