On 27.03.2020 19:20, Andrew Cooper wrote:
> On 24/03/2020 12:30, Jan Beulich wrote:
>> --- a/tools/tests/x86_emulator/evex-disp8.c
>> +++ b/tools/tests/x86_emulator/evex-disp8.c
>> @@ -550,6 +550,12 @@ static const struct test avx512_4vnniw_5
>> INSN(p4dpwssds, f2, 0f38, 53, el_4, d, vl),
>> };
>>
>> +static const struct test avx512_bf16_all[] = {
>> + INSN(vcvtne2ps2bf16, f2, 0f38, 72, vl, d, vl),
>> + INSN(vcvtneps2bf16, f3, 0f38, 72, vl, d, vl),
>> + INSN(vdpbf16ps, f3, 0f38, 52, vl, d, vl),
>> +};
>> +
>> static const struct test avx512_bitalg_all[] = {
>> INSN(popcnt, 66, 0f38, 54, vl, bw, vl),
>> INSN(pshufbitqmb, 66, 0f38, 8f, vl, b, vl),
>> @@ -984,6 +990,7 @@ void evex_disp8_test(void *instr, struct
>> RUN(avx512pf, 512);
>> RUN(avx512_4fmaps, 512);
>> RUN(avx512_4vnniw, 512);
>> + RUN(avx512_bf16, all);
>> RUN(avx512_bitalg, all);
>> RUN(avx512_ifma, all);
>> RUN(avx512_vbmi, all);
>> --- a/tools/tests/x86_emulator/test_x86_emulator.c
>> +++ b/tools/tests/x86_emulator/test_x86_emulator.c
>> @@ -4516,6 +4516,80 @@ int main(int argc, char **argv)
>> else
>> printf("skipped\n");
>>
>> + if ( stack_exec && cpu_has_avx512_bf16 )
>> + {
>> + decl_insn(vcvtne2ps2bf16);
>> + decl_insn(vcvtneps2bf16);
>> + decl_insn(vdpbf16ps);
>> + static const struct {
>> + float f[16];
>> + } in1 = {{
>> + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
>> + }}, in2 = {{
>> + 1, -2, 3, -4, 5, -6, 7, -8, 9, -10, 11, -12, 13, -14, 15, -16
>> + }}, out = {{
>> + 1 * 1 + 2 * 2, 3 * 3 + 4 * 4,
>> + 5 * 5 + 6 * 6, 7 * 7 + 8 * 8,
>> + 9 * 9 + 10 * 10, 11 * 11 + 12 * 12,
>> + 13 * 13 + 14 * 14, 15 * 15 + 16 * 16,
>> + 1 * 1 - 2 * 2, 3 * 3 - 4 * 4,
>> + 5 * 5 - 6 * 6, 7 * 7 - 8 * 8,
>> + 9 * 9 - 10 * 10, 11 * 11 - 12 * 12,
>> + 13 * 13 - 14 * 14, 15 * 15 - 16 * 16
>> + }};
>> +
>> + printf("%-40s", "Testing vcvtne2ps2bf16 64(%ecx),%zmm1,%zmm2...");
>> + asm volatile ( "vmovups %1, %%zmm1\n"
>> + put_insn(vcvtne2ps2bf16,
>> + /* vcvtne2ps2bf16 64(%0), %%zmm1, %%zmm2 */
>> + ".byte 0x62, 0xf2, 0x77, 0x48, 0x72, 0x51,
>> 0x01")
>> + :: "c" (NULL), "m" (in2) );
>> + set_insn(vcvtne2ps2bf16);
>> + regs.ecx = (unsigned long)&in1 - 64;
>> + rc = x86_emulate(&ctxt, &emulops);
>> + if ( rc != X86EMUL_OKAY || !check_eip(vcvtne2ps2bf16) )
>> + goto fail;
>> + printf("pending\n");
>> +
>> + printf("%-40s", "Testing vcvtneps2bf16 64(%ecx),%ymm3...");
>> + asm volatile ( put_insn(vcvtneps2bf16,
>> + /* vcvtneps2bf16 64(%0), %%ymm3 */
>> + ".byte 0x62, 0xf2, 0x7e, 0x48, 0x72, 0x59,
>> 0x01")
>> + :: "c" (NULL) );
>> + set_insn(vcvtneps2bf16);
>> + rc = x86_emulate(&ctxt, &emulops);
>> + if ( rc != X86EMUL_OKAY || !check_eip(vcvtneps2bf16) )
>> + goto fail;
>> + asm ( "vmovdqa %%ymm2, %%ymm5\n\t"
>> + "vpcmpeqd %%zmm3, %%zmm5, %%k0\n\t"
>> + "kmovw %%k0, %0"
>> + : "=g" (rc) : "m" (out) );
>> + if ( rc != 0xffff )
>> + goto fail;
>> + printf("pending\n");
>> +
>> + printf("%-40s", "Testing vdpbf16ps 128(%ecx),%zmm2,%zmm4...");
>> + asm volatile ( "vmovdqa %%ymm3, %0\n\t"
>> + "vmovdqa %%ymm3, %1\n"
>> + put_insn(vdpbf16ps,
>> + /* vdpbf16ps 128(%2), %%zmm2, %%zmm4 */
>> + ".byte 0x62, 0xf2, 0x6e, 0x48, 0x52, 0x61,
>> 0x02")
>> + : "=&m" (res[0]), "=&m" (res[8])
>> + : "c" (NULL)
>> + : "memory" );
>> + set_insn(vdpbf16ps);
>> + regs.ecx = (unsigned long)res - 128;
>> + rc = x86_emulate(&ctxt, &emulops);
>> + if ( rc != X86EMUL_OKAY || !check_eip(vdpbf16ps) )
>> + goto fail;
>> + asm ( "vcmpeqps %1, %%zmm4, %%k0\n\t"
>> + "kmovw %%k0, %0"
>> + : "=g" (rc) : "m" (out) );
>> + if ( rc != 0xffff )
>> + goto fail;
>> + printf("okay\n");
>> + }
>
> I've just tried this out on an SDP.
>
> Testing vcvtne2ps2bf16 64(%ecx),%zmm1,%zmm2...pending
> Testing vcvtneps2bf16 64(%ecx),%ymm3... pending
> Testing vdpbf16ps 128(%ecx),%zmm2,%zmm4...okay
> ...
> Testing avx512_bf16/all disp8 handling...okay
>
> What is the "pending" supposed to signify? I can see that these three
> are linked, and that is fine, but at the point we've checked the
> intermediate results, it should be "okay", no?
I didn't think so, and hence I've used "pending". Whether the result of
the first two is indeed correct is only known after the 3rd.
Jan
