The following instructions assembled with nasm version 2.11.06 use a displacement multiplier of 4 bytes in the memory operand, the correct multiplier is 1 for b and 2 for w, respectively. The d and q suffix instructions are coded correctly.
{evex} vpinsrb xmm1,xmm2,[rsi+40h], 2
{evex} vpinsrw xmm1,xmm2,[rsi+40h], 2
{evex} vpextrb [rsi+40h],xmm2, 2
{evex} vpextrw [rsi+40h],xmm2, 2