Index: src/x86/filmgrain16_avx2.asm
--- src/x86/filmgrain16_avx2.asm.orig
+++ src/x86/filmgrain16_avx2.asm
@@ -163,6 +163,7 @@ cglobal generate_grain_y_16bpc, 3, 9, 14, buf, fg_data
     jmp              r5
 
 .ar1:
+    _CET_ENDBR
     DEFINE_ARGS buf, fg_data, max, shift, val3, min, cf3, x, val0
     mov          shiftd, [fg_dataq+FGData.ar_coeff_shift]
     movsx          cf3d, byte [fg_dataq+FGData.ar_coeffs_y+3]
@@ -214,9 +215,11 @@ cglobal generate_grain_y_16bpc, 3, 9, 14, buf, fg_data
     dec              hd
     jg .y_loop_ar1
 .ar0:
+    _CET_ENDBR
     RET
 
 .ar2:
+    _CET_ENDBR
     DEFINE_ARGS buf, fg_data, bdmax, shift
     mov          shiftd, [fg_dataq+FGData.ar_coeff_shift]
     movq            xm0, [fg_dataq+FGData.ar_coeffs_y+5]    ; cf5-11
@@ -286,6 +289,7 @@ cglobal generate_grain_y_16bpc, 3, 9, 14, buf, fg_data
     RET
 
 .ar3:
+    _CET_ENDBR
     DEFINE_ARGS buf, fg_data, bdmax, shift
     mov          shiftd, [fg_dataq+FGData.ar_coeff_shift]
     sar          bdmaxd, 1
@@ -447,6 +451,7 @@ cglobal generate_grain_uv_%1_16bpc, 4, 11, 8, buf, buf
 
 INIT_YMM avx2
 .ar0:
+    _CET_ENDBR
     DEFINE_ARGS buf, bufy, fg_data, uv, bdmax, shift
     imul            uvd, 28
     mov          shiftd, [fg_dataq+FGData.ar_coeff_shift]
@@ -566,6 +571,7 @@ INIT_YMM avx2
 
 INIT_XMM avx2
 .ar1:
+    _CET_ENDBR
     DEFINE_ARGS buf, bufy, fg_data, uv, max, cf3, min, val3, x, shift
     imul            uvd, 28
     mov          shiftd, [fg_dataq+FGData.ar_coeff_shift]
@@ -645,6 +651,7 @@ INIT_XMM avx2
 
 INIT_YMM avx2
 .ar2:
+    _CET_ENDBR
 %if WIN64
     ; xmm6 and xmm7 already saved
     %assign xmm_regs_used 13 + %2
@@ -746,6 +753,7 @@ INIT_YMM avx2
     RET
 
 .ar3:
+    _CET_ENDBR
 %if WIN64
     ; xmm6 and xmm7 already saved
     %assign stack_offset 32