blob: 8da3e73db7f5d488e64157acd6f63a9d62bf6017 [file] [log] [blame]
// clang-format off
byte bf16_00[] = { 0x62, 0x02, 0x17, 0x40, 0x72, 0xf4,}; //
byte bf16_01[] = { 0x62, 0x22, 0x17, 0x47, 0x72, 0xb4, 0xf5, 0x00, 0x00, 0x00, 0x10,}; //
byte bf16_02[] = { 0x62, 0x42, 0x17, 0x50, 0x72, 0x31,}; //
byte bf16_03[] = { 0x62, 0x62, 0x17, 0x40, 0x72, 0x71, 0x7f,}; //
byte bf16_04[] = { 0x62, 0x62, 0x17, 0xd7, 0x72, 0xb2, 0x00, 0xe0, 0xff, 0xff,}; //
byte bf16_05[] = { 0x62, 0x02, 0x7e, 0x48, 0x72, 0xf5,}; //
byte bf16_06[] = { 0x62, 0x22, 0x7e, 0x4f, 0x72, 0xb4, 0xf5, 0x00, 0x00, 0x00, 0x10,}; //
byte bf16_07[] = { 0x62, 0x42, 0x7e, 0x58, 0x72, 0x31,}; //
byte bf16_08[] = { 0x62, 0x62, 0x7e, 0x48, 0x72, 0x71, 0x7f,}; //
byte bf16_09[] = { 0x62, 0x62, 0x7e, 0xdf, 0x72, 0xb2, 0x00, 0xe0, 0xff, 0xff,}; //
byte bf16_10[] = { 0x62, 0x02, 0x16, 0x40, 0x52, 0xf4,}; //
byte bf16_11[] = { 0x62, 0x22, 0x16, 0x47, 0x52, 0xb4, 0xf5, 0x00, 0x00, 0x00, 0x10,}; //
byte bf16_12[] = { 0x62, 0x42, 0x16, 0x50, 0x52, 0x31,}; //
byte bf16_13[] = { 0x62, 0x62, 0x16, 0x40, 0x52, 0x71, 0x7f,}; //
byte bf16_14[] = { 0x62, 0x62, 0x16, 0xd7, 0x52, 0xb2, 0x00, 0xe0, 0xff, 0xff,}; //
ENC4(bf16_00, vcvtne2ps2bf16_mask, 0, R(ZMM30), R(K0), R(ZMM29), R(ZMM28));
ENC4(bf16_01, vcvtne2ps2bf16_mask, 0, R(ZMM30), R(K7), R(ZMM29), M(RBP, R14, 8, 0x10000000, OPSZ_64));
ENC4(bf16_02, vcvtne2ps2bf16_mask, 0, R(ZMM30), R(K0), R(ZMM29), M(R9, NULL, 0, 0, OPSZ_4));
ENC4(bf16_03, vcvtne2ps2bf16_mask, 0, R(ZMM30), R(K0), R(ZMM29), M(RCX, NULL, 0, 0x1fc0, OPSZ_64));
ENC4(bf16_04, vcvtne2ps2bf16_mask, Z, R(ZMM30), R(K7), R(ZMM29), M(RDX, NULL, 0, 0xffffe000, OPSZ_4));
ENC3(bf16_05, vcvtneps2bf16_mask, 0, Zh(ZMM30), R(K0), R(ZMM29));
ENC3(bf16_06, vcvtneps2bf16_mask, 0, Zh(ZMM30), R(K7), M(RBP, R14, 8, 0x10000000, OPSZ_64));
ENC3(bf16_07, vcvtneps2bf16_mask, 0, Zh(ZMM30), R(K0), M(R9, NULL, 0, 0, OPSZ_4));
ENC3(bf16_08, vcvtneps2bf16_mask, 0, Zh(ZMM30), R(K0), M(RCX, NULL, 0, 0x1fc0, OPSZ_64));
ENC3(bf16_09, vcvtneps2bf16_mask, Z, Zh(ZMM30), R(K7), M(RDX, NULL, 0, 0xffffe000, OPSZ_4));
ENC4(bf16_10, vdpbf16ps_mask, 0, R(ZMM30), R(K0), R(ZMM29), R(ZMM28));
ENC4(bf16_11, vdpbf16ps_mask, 0, R(ZMM30), R(K7), R(ZMM29), M(RBP, R14, 8, 0x10000000, OPSZ_64));
ENC4(bf16_12, vdpbf16ps_mask, 0, R(ZMM30), R(K0), R(ZMM29), M(R9, NULL, 0, 0, OPSZ_4));
ENC4(bf16_13, vdpbf16ps_mask, 0, R(ZMM30), R(K0), R(ZMM29), M(RCX, NULL, 0, 0x1fc0, OPSZ_64));
ENC4(bf16_14, vdpbf16ps_mask, Z, R(ZMM30), R(K7), R(ZMM29), M(RDX, NULL, 0, 0xffffe000, OPSZ_4));