1 //===-- CaymanInstructions.td - CM Instruction defs -------*- tablegen -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // TableGen definitions for instructions which are available only on Cayman 11 // family GPUs. 12 // 13 //===----------------------------------------------------------------------===// 14 15 def isCayman : Predicate<"Subtarget->hasCaymanISA()">; 16 17 //===----------------------------------------------------------------------===// 18 // Cayman Instructions 19 //===----------------------------------------------------------------------===// 20 21 let SubtargetPredicate = isCayman in { 22 23 def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24", 24 [(set i32:$dst, (AMDGPUmad_i24 i32:$src0, i32:$src1, i32:$src2))], VecALU 25 >; 26 def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24", 27 [(set i32:$dst, (AMDGPUmul_i24 i32:$src0, i32:$src1))], VecALU 28 >; 29 30 def : IMad24Pat<MULADD_INT24_cm>; 31 32 let isVector = 1 in { 33 34 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>; 35 36 def MULLO_INT_cm : MULLO_INT_Common<0x8F>; 37 def MULHI_INT_cm : MULHI_INT_Common<0x90>; 38 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>; 39 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>; 40 def MULHI_INT_cm24 : MULHI_INT24_Common<0x5c>; 41 def MULHI_UINT_cm24 : MULHI_UINT24_Common<0xb2>; 42 43 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>; 44 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>; 45 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>; 46 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>; 47 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>; 48 def SIN_cm : SIN_Common<0x8D>; 49 def COS_cm : COS_Common<0x8E>; 50 } // End isVector = 1 51 52 def : RsqPat<RECIPSQRT_IEEE_cm, f32>; 53 54 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>; 55 56 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>; 57 58 // RECIP_UINT emulation for Cayman 59 // The multiplication scales from [0,1] to the unsigned integer range 60 def : R600Pat < 61 (AMDGPUurecip i32:$src0), 62 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)), 63 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1))) 64 >; 65 66 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> { 67 let ADDR = 0; 68 let POP_COUNT = 0; 69 let COUNT = 0; 70 } 71 72 73 74 def : R600Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>; 75 76 class RAT_STORE_DWORD <RegisterClass rc, ValueType vt, bits<4> mask> : 77 CF_MEM_RAT_CACHELESS <0x14, 0, mask, 78 (ins rc:$rw_gpr, R600_TReg32_X:$index_gpr), 79 "STORE_DWORD $rw_gpr, $index_gpr", 80 [(store_global vt:$rw_gpr, i32:$index_gpr)]> { 81 let eop = 0; // This bit is not used on Cayman. 82 } 83 84 def RAT_STORE_DWORD32 : RAT_STORE_DWORD <R600_TReg32_X, i32, 0x1>; 85 def RAT_STORE_DWORD64 : RAT_STORE_DWORD <R600_Reg64, v2i32, 0x3>; 86 def RAT_STORE_DWORD128 : RAT_STORE_DWORD <R600_Reg128, v4i32, 0xf>; 87 88 def RAT_STORE_TYPED_cm: CF_MEM_RAT_STORE_TYPED<0> { 89 let eop = 0; // This bit is not used on Cayman. 90 } 91 92 class VTX_READ_cm <string name, dag outs> 93 : VTX_WORD0_cm, VTX_READ<name, outs, []> { 94 95 // Static fields 96 let VC_INST = 0; 97 let FETCH_TYPE = 2; 98 let FETCH_WHOLE_QUAD = 0; 99 let SRC_REL = 0; 100 // XXX: We can infer this field based on the SRC_GPR. This would allow us 101 // to store vertex addresses in any channel, not just X. 102 let SRC_SEL_X = 0; 103 let SRC_SEL_Y = 0; 104 let STRUCTURED_READ = 0; 105 let LDS_REQ = 0; 106 let COALESCED_READ = 0; 107 108 let Inst{31-0} = Word0; 109 } 110 111 def VTX_READ_8_cm 112 : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", 113 (outs R600_TReg32_X:$dst_gpr)> { 114 115 let DST_SEL_X = 0; 116 let DST_SEL_Y = 7; // Masked 117 let DST_SEL_Z = 7; // Masked 118 let DST_SEL_W = 7; // Masked 119 let DATA_FORMAT = 1; // FMT_8 120 } 121 122 def VTX_READ_16_cm 123 : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", 124 (outs R600_TReg32_X:$dst_gpr)> { 125 let DST_SEL_X = 0; 126 let DST_SEL_Y = 7; // Masked 127 let DST_SEL_Z = 7; // Masked 128 let DST_SEL_W = 7; // Masked 129 let DATA_FORMAT = 5; // FMT_16 130 131 } 132 133 def VTX_READ_32_cm 134 : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", 135 (outs R600_TReg32_X:$dst_gpr)> { 136 137 let DST_SEL_X = 0; 138 let DST_SEL_Y = 7; // Masked 139 let DST_SEL_Z = 7; // Masked 140 let DST_SEL_W = 7; // Masked 141 let DATA_FORMAT = 0xD; // COLOR_32 142 143 // This is not really necessary, but there were some GPU hangs that appeared 144 // to be caused by ALU instructions in the next instruction group that wrote 145 // to the $src_gpr registers of the VTX_READ. 146 // e.g. 147 // %t3_x = VTX_READ_PARAM_32_eg killed %t2_x, 24 148 // %t2_x = MOV %zero 149 //Adding this constraint prevents this from happening. 150 let Constraints = "$src_gpr.ptr = $dst_gpr"; 151 } 152 153 def VTX_READ_64_cm 154 : VTX_READ_cm <"VTX_READ_64 $dst_gpr.XY, $src_gpr", 155 (outs R600_Reg64:$dst_gpr)> { 156 157 let DST_SEL_X = 0; 158 let DST_SEL_Y = 1; 159 let DST_SEL_Z = 7; 160 let DST_SEL_W = 7; 161 let DATA_FORMAT = 0x1D; // COLOR_32_32 162 } 163 164 def VTX_READ_128_cm 165 : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", 166 (outs R600_Reg128:$dst_gpr)> { 167 168 let DST_SEL_X = 0; 169 let DST_SEL_Y = 1; 170 let DST_SEL_Z = 2; 171 let DST_SEL_W = 3; 172 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32 173 174 // XXX: Need to force VTX_READ_128 instructions to write to the same register 175 // that holds its buffer address to avoid potential hangs. We can't use 176 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst 177 // registers are different sizes. 178 } 179 180 //===----------------------------------------------------------------------===// 181 // VTX Read from parameter memory space 182 //===----------------------------------------------------------------------===// 183 def : R600Pat<(i32:$dst_gpr (vtx_id3_az_extloadi8 ADDRVTX_READ:$src_gpr)), 184 (VTX_READ_8_cm MEMxi:$src_gpr, 3)>; 185 def : R600Pat<(i32:$dst_gpr (vtx_id3_az_extloadi16 ADDRVTX_READ:$src_gpr)), 186 (VTX_READ_16_cm MEMxi:$src_gpr, 3)>; 187 def : R600Pat<(i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)), 188 (VTX_READ_32_cm MEMxi:$src_gpr, 3)>; 189 def : R600Pat<(v2i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)), 190 (VTX_READ_64_cm MEMxi:$src_gpr, 3)>; 191 def : R600Pat<(v4i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)), 192 (VTX_READ_128_cm MEMxi:$src_gpr, 3)>; 193 194 //===----------------------------------------------------------------------===// 195 // VTX Read from constant memory space 196 //===----------------------------------------------------------------------===// 197 def : R600Pat<(i32:$dst_gpr (vtx_id2_az_extloadi8 ADDRVTX_READ:$src_gpr)), 198 (VTX_READ_8_cm MEMxi:$src_gpr, 2)>; 199 def : R600Pat<(i32:$dst_gpr (vtx_id2_az_extloadi16 ADDRVTX_READ:$src_gpr)), 200 (VTX_READ_16_cm MEMxi:$src_gpr, 2)>; 201 def : R600Pat<(i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)), 202 (VTX_READ_32_cm MEMxi:$src_gpr, 2)>; 203 def : R600Pat<(v2i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)), 204 (VTX_READ_64_cm MEMxi:$src_gpr, 2)>; 205 def : R600Pat<(v4i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)), 206 (VTX_READ_128_cm MEMxi:$src_gpr, 2)>; 207 208 //===----------------------------------------------------------------------===// 209 // VTX Read from global memory space 210 //===----------------------------------------------------------------------===// 211 def : R600Pat<(i32:$dst_gpr (vtx_id1_az_extloadi8 ADDRVTX_READ:$src_gpr)), 212 (VTX_READ_8_cm MEMxi:$src_gpr, 1)>; 213 def : R600Pat<(i32:$dst_gpr (vtx_id1_az_extloadi16 ADDRVTX_READ:$src_gpr)), 214 (VTX_READ_16_cm MEMxi:$src_gpr, 1)>; 215 def : R600Pat<(i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)), 216 (VTX_READ_32_cm MEMxi:$src_gpr, 1)>; 217 def : R600Pat<(v2i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)), 218 (VTX_READ_64_cm MEMxi:$src_gpr, 1)>; 219 def : R600Pat<(v4i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)), 220 (VTX_READ_128_cm MEMxi:$src_gpr, 1)>; 221 222 } // End let SubtargetPredicate = isCayman 223