[AMDGPU][MC][GFX908] dot opcodes operating 4-bit and 8-bit packed data should not support op_sel

[dpreobra]

Bugzilla Link	44912
Version	trunk
OS	Windows NT
CC	@atamazov,@arsenm,@rampitec

Extended Description

dot opcodes that have packed data smaller than 16 bits should not allow op_sel/op_sel_hi modifiers. This is confirmed by Justin Smith.

Examples of failed tests:

v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1]
v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[1,1]

[dpreobra]

assigned to @dpreobra

[jwanggit86]

According to ISA document, this is what v_dot4_u32_u8 does:

tmp = S2.i32;
tmp += i8_to_i32(S0[7 : 0].i8) * i8_to_i32(S1[7 : 0].i8);
tmp += i8_to_i32(S0[15 : 8].i8) * i8_to_i32(S1[15 : 8].i8);
tmp += i8_to_i32(S0[23 : 16].i8) * i8_to_i32(S1[23 : 16].i8);
tmp += i8_to_i32(S0[31 : 24].i8) * i8_to_i32(S1[31 : 24].i8);
D0.i32 = tmp

All the 4 bytes of both src0 and src1 are used. So it seems op_sel is indeed superfluous for this instruction. Same for v_dot8.

[arsenm]

According to ISA document, this is what v_dot4_u32_u8 does:

tmp = S2.i32;
tmp += i8_to_i32(S0[7 : 0].i8) * i8_to_i32(S1[7 : 0].i8);
tmp += i8_to_i32(S0[15 : 8].i8) * i8_to_i32(S1[15 : 8].i8);
tmp += i8_to_i32(S0[23 : 16].i8) * i8_to_i32(S1[23 : 16].i8);
tmp += i8_to_i32(S0[31 : 24].i8) * i8_to_i32(S1[31 : 24].i8);
D0.i32 = tmp

All the 4 bytes of both src0 and src1 are used. So it seems op_sel is indeed superfluous for this instruction. Same for v_dot8.

Should cross check that, the pseudocode documentation for these kinds of cases often misses mentioning the modifiers

[jwanggit86]

Closed by PR 100485.