brintos

brintos / llvm-project-archived public Read only

0
0
Text · 7.4 KiB · 059c8da Raw
134 lines · plain
1# RUN: llvm-mc -triple=amdgcn -mcpu=gfx950 -disassemble -show-encoding < %s | FileCheck -check-prefix=GFX950 %s2 3# GFX950: v_dot2c_f32_bf16_e32 v5, v1, v2         ; encoding: [0x01,0x05,0x0a,0x2c]40x01,0x05,0x0a,0x2c5 6# GFX950: v_dot2c_f32_bf16_e32 v255, v1, v2       ; encoding: [0x01,0x05,0xfe,0x2d]70x01,0x05,0xfe,0x2d8 9# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x00]100xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0011 12# GFX950: v_dot2c_f32_bf16_dpp v255, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0xfe,0x2d,0x01,0xe4,0x00,0x00]130xfa,0x04,0xfe,0x2d,0x01,0xe4,0x00,0x0014 15# GFX950: v_dot2c_f32_bf16_dpp v5, v255, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0xff,0xe4,0x00,0x00]160xfa,0x04,0x0a,0x2c,0xff,0xe4,0x00,0x0017 18# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v255 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0xfe,0x0b,0x2c,0x01,0xe4,0x00,0x00]190xfa,0xfe,0x0b,0x2c,0x01,0xe4,0x00,0x0020 21# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[3,2,1,0] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x1b,0x00,0x00]220xfa,0x04,0x0a,0x2c,0x01,0x1b,0x00,0x0023 24# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_mirror row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x40,0x01,0x00]250xfa,0x04,0x0a,0x2c,0x01,0x40,0x01,0x0026 27# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_half_mirror row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x41,0x01,0x00]280xfa,0x04,0x0a,0x2c,0x01,0x41,0x01,0x0029 30# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_bcast:15 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x42,0x01,0x00]310xfa,0x04,0x0a,0x2c,0x01,0x42,0x01,0x0032 33# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_bcast:31 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x43,0x01,0x00]340xfa,0x04,0x0a,0x2c,0x01,0x43,0x01,0x0035 36# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 wave_shl:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x30,0x01,0x00]370xfa,0x04,0x0a,0x2c,0x01,0x30,0x01,0x0038 39# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 wave_rol:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x34,0x01,0x00]400xfa,0x04,0x0a,0x2c,0x01,0x34,0x01,0x0041 42# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 wave_shr:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x38,0x01,0x00]430xfa,0x04,0x0a,0x2c,0x01,0x38,0x01,0x0044 45# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 wave_ror:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x3c,0x01,0x00]460xfa,0x04,0x0a,0x2c,0x01,0x3c,0x01,0x0047 48# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_shl:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x01,0x01,0x00]490xfa,0x04,0x0a,0x2c,0x01,0x01,0x01,0x0050 51# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_shl:15 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x0f,0x01,0x00]520xfa,0x04,0x0a,0x2c,0x01,0x0f,0x01,0x0053 54# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_shr:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x11,0x01,0x00]550xfa,0x04,0x0a,0x2c,0x01,0x11,0x01,0x0056 57# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_shr:15 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x1f,0x01,0x00]580xfa,0x04,0x0a,0x2c,0x01,0x1f,0x01,0x0059 60# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_ror:1 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x21,0x01,0x00]610xfa,0x04,0x0a,0x2c,0x01,0x21,0x01,0x0062 63# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 row_ror:15 row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0x2f,0x01,0x00]640xfa,0x04,0x0a,0x2c,0x01,0x2f,0x01,0x0065 66# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x1 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x10]670xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x1068 69# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x3 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x30]700xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x3071 72# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0xf0]730xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0xf074 75# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0xf0]760xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0xf077 78# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x1 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x01]790xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0180 81# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x3 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x03]820xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0383 84# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0xf ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0f]850xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0f86 87# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0xf ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0f]880xfa,0x04,0x0a,0x2c,0x01,0xe4,0x00,0x0f89 90# GFX950: v_dot2c_f32_bf16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 bound_ctrl:1 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x08,0x00]910xfa,0x04,0x0a,0x2c,0x01,0xe4,0x08,0x0092 93# GFX950: v_dot2c_f32_bf16_dpp v5, -v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x10,0x00]940xfa,0x04,0x0a,0x2c,0x01,0xe4,0x10,0x0095 96# GFX950: v_dot2c_f32_bf16_dpp v5, |v1|, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x20,0x00]970xfa,0x04,0x0a,0x2c,0x01,0xe4,0x20,0x0098 99# GFX950: v_dot2c_f32_bf16_dpp v5, v1, -v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x40,0x00]1000xfa,0x04,0x0a,0x2c,0x01,0xe4,0x40,0x00101 102# GFX950: v_dot2c_f32_bf16_dpp v5, v1, |v2| quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0 ; encoding: [0xfa,0x04,0x0a,0x2c,0x01,0xe4,0x80,0x00]1030xfa,0x04,0x0a,0x2c,0x01,0xe4,0x80,0x00104 105# GFX950: v_dot2c_f32_bf16_e64 v5, v1, src_scc    ; encoding: [0x05,0x00,0x16,0xd1,0x01,0xfb,0x01,0x00]1060x05,0x00,0x16,0xd1,0x01,0xfb,0x01,0x00107 108# GFX950: v_dot2c_f32_bf16_e64 v5, v255, src_execz ; encoding: [0x05,0x00,0x16,0xd1,0xff,0xf9,0x01,0x00]1090x05,0x00,0x16,0xd1,0xff,0xf9,0x01,0x00110 111# GFX950: v_dot2c_f32_bf16_e64 v5, s101, s101     ; encoding: [0x05,0x00,0x16,0xd1,0x65,0xca,0x00,0x00]1120x05,0x00,0x16,0xd1,0x65,0xca,0x00,0x00113 114# GFX950: v_dot2c_f32_bf16_e64 v5, -1, flat_scratch_lo ; encoding: [0x05,0x00,0x16,0xd1,0xc1,0xcc,0x00,0x00]1150x05,0x00,0x16,0xd1,0xc1,0xcc,0x00,0x00116 117# GFX950: v_dot2c_f32_bf16_e64 v5, 0.5, -|flat_scratch_hi| ; encoding: [0x05,0x02,0x16,0xd1,0xf0,0xce,0x00,0x40]1180x05,0x02,0x16,0xd1,0xf0,0xce,0x00,0x40119 120# GFX950: v_dot2c_f32_bf16_e64 v5, src_execz, 0.5 mul:4 ; encoding: [0x05,0x00,0x16,0xd1,0xfc,0xe0,0x01,0x10]1210x05,0x00,0x16,0xd1,0xfc,0xe0,0x01,0x10122 123# GFX950: v_dot2c_f32_bf16_e64 v255, -|src_scc|, -1 clamp div:2 ; encoding: [0xff,0x81,0x16,0xd1,0xfd,0x82,0x01,0x38]1240xff,0x81,0x16,0xd1,0xfd,0x82,0x01,0x38125 126# GFX950: v_dot2c_f32_bf16_e32 v5, 10, v2         ; encoding: [0x8a,0x04,0x0a,0x2c]1270x8a,0x04,0x0a,0x2c128 129# GFX950: v_dot2c_f32_bf16_e32 v5, 0x64, v2       ; encoding: [0xff,0x04,0x0a,0x2c,0x64,0x00,0x00,0x00]1300xff,0x04,0x0a,0x2c,0x64,0x00,0x00,0x00131 132# GFX950: v_dot2c_f32_bf16_e32 v5, 0x4122, v2     ; encoding: [0xff,0x04,0x0a,0x2c,0x22,0x41,0x00,0x00]1330xff,0x04,0x0a,0x2c,0x22,0x41,0x00,0x00134