src/amd/compiler/README.md

   1 # Unofficial GCN/RDNA ISA reference errata
   2
   3 ## v_sad_u32
   4
   5 The Vega ISA reference writes it's behaviour as:
   6 ```
   7 D.u = abs(S0.i - S1.i) + S2.u.
   8 ```
   9 This is incorrect. The actual behaviour is what is written in the GCN3 reference
  10 guide:
  11 ```
  12 ABS_DIFF (A,B) = (A>B) ? (A-B) : (B-A)
  13 D.u = ABS_DIFF (S0.u,S1.u) + S2.u
  14 ```
  15 The instruction doesn't subtract the S0 and S1 and use the absolute value (the
  16 _signed_ distance), it uses the _unsigned_ distance between the operands. So
  17 `v_sad_u32(-5, 0, 0)` would return `4294967291` (`-5` interpreted as unsigned),
  18 not `5`.
  19
  20 ## s_bfe_*
  21
  22 Both the Vega and GCN3 ISA references write that these instructions don't write
  23 SCC. They do.
  24
  25 ## v_bcnt_u32_b32
  26
  27 The Vega ISA reference writes it's behaviour as:
  28 ```
  29 D.u = 0;
  30 for i in 0 ... 31 do
  31 D.u += (S0.u[i] == 1 ? 1 : 0);
  32 endfor.
  33 ```
  34 This is incorrect. The actual behaviour (and number of operands) is what
  35 is written in the GCN3 reference guide:
  36 ```
  37 D.u = CountOneBits(S0.u) + S1.u.
  38 ```
  39
  40 ## SMEM stores
  41
  42 The Vega ISA references doesn't say this (or doesn't make it clear), but
  43 the offset for SMEM stores must be in m0 if IMM == 0.
  44
  45 The RDNA ISA doesn't mention SMEM stores at all, but they seem to be supported
  46 by the chip and are present in LLVM. AMD devs however highly recommend avoiding
  47 these instructions.
  48
  49 ## SMEM atomics
  50
  51 RDNA ISA: same as the SMEM stores, the ISA pretends they don't exist, but they
  52 are there in LLVM.
  53
  54 ## VMEM stores
  55
  56 All reference guides say (under "Vector Memory Instruction Data Dependencies"):
  57 > When a VM instruction is issued, the address is immediately read out of VGPRs
  58 > and sent to the texture cache. Any texture or buffer resources and samplers
  59 > are also sent immediately. However, write-data is not immediately sent to the
  60 > texture cache.
  61 Reading that, one might think that waitcnts need to be added when writing to
  62 the registers used for a VMEM store's data. Experimentation has shown that this
  63 does not seem to be the case on GFX8 and GFX9 (GFX6 and GFX7 are untested). It
  64 also seems unlikely, since NOPs are apparently needed in a subset of these
  65 situations.
  66
  67 ## MIMG opcodes on GFX8/GCN3
  68
  69 The `image_atomic_{swap,cmpswap,add,sub}` opcodes in the GCN3 ISA reference
  70 guide are incorrect. The Vega ISA reference guide has the correct ones.
  71
  72 ## VINTRP encoding
  73
  74 VEGA ISA doc says the encoding should be `110010` but `110101` works.
  75
  76 ## VOP1 instructions encoded as VOP3
  77
  78 RDNA ISA doc says that `0x140` should be added to the opcode, but that doesn't
  79 work. What works is adding `0x180`, which LLVM also does.
  80
  81 ## FLAT, Scratch, Global instructions
  82
  83 The NV bit was removed in RDNA, but some parts of the doc still mention it.
  84
  85 RDNA ISA doc 13.8.1 says that SADDR should be set to 0x7f when ADDR is used, but
  86 9.3.1 says it should be set to NULL. We assume 9.3.1 is correct and set it to
  87 SGPR_NULL.
  88
  89 ## Legacy instructions
  90
  91 Some instructions have a `_LEGACY` variant which implements "DX9 rules", in which
  92 the zero "wins" in multiplications, ie. `0.0*x` is always `0.0`. The VEGA ISA
  93 mentions `V_MAC_LEGACY_F32` but this instruction is not really there on VEGA.
  94
  95 ## RDNA L0, L1 cache and DLC, GLC bits
  96
  97 The old L1 cache was renamed to L0, and a new L1 cache was added to RDNA. The
  98 L1 cache is 1 cache per shader array. Some instruction encodings have DLC and
  99 GLC bits that interact with the cache.
 100
 101 * DLC ("device level coherent") bit: controls the L1 cache
 102 * GLC ("globally coherent") bit: controls the L0 cache
 103
 104 The recommendation from AMD devs is to always set these two bits at the same time,
 105 as it doesn't make too much sense to set them independently, aside from some
 106 circumstances (eg. we needn't set DLC when only one shader array is used).
 107
 108 Stores and atomics always bypass the L1 cache, so they don't support the DLC bit,
 109 and it shouldn't be set in these cases. Setting the DLC for these cases can result
 110 in graphical glitches.
 111
 112 ## RDNA S_DCACHE_WB
 113
 114 The S_DCACHE_WB is not mentioned in the RDNA ISA doc, but it is needed in order
 115 to achieve correct behavior in some SSBO CTS tests.
 116
 117 ## RDNA subvector mode
 118
 119 The documentation of S_SUBVECTOR_LOOP_BEGIN and S_SUBVECTOR_LOOP_END is not clear
 120 on what sort of addressing should be used, but it says that it
 121 "is equivalent to an S_CBRANCH with extra math", so the subvector loop handling
 122 in ACO is done according to the S_CBRANCH doc.
 123
 124 # Hardware Bugs
 125
 126 ## SMEM corrupts VCCZ on SI/CI
 127
 128 https://github.com/llvm/llvm-project/blob/acb089e12ae48b82c0b05c42326196a030df9b82/llvm/lib/Target/AMDGPU/SIInsertWaits.cpp#L580-L616
 129 After issuing a SMEM instructions, we need to wait for the SMEM instructions to
 130 finish and then write to vcc (for example, `s_mov_b64 vcc, vcc`) to correct vccz
 131
 132 Currently, we don't do this.
 133
 134 ## RDNA / GFX10 hazards
 135
 136 ### SMEM store followed by a load with the same address
 137
 138 We found that an `s_buffer_load` will produce incorrect results if it is preceded
 139 by an `s_buffer_store` with the same address. Inserting an `s_nop` between them
 140 does not mitigate the issue, so an `s_waitcnt lgkmcnt(0)` must be inserted.
 141 This is not mentioned by LLVM among the other GFX10 bugs, but LLVM doesn't use
 142 SMEM stores, so it's not surprising that they didn't notice it.
 143
 144 ### VMEMtoScalarWriteHazard
 145
 146 Triggered by:
 147 VMEM/FLAT/GLOBAL/SCRATCH/DS instruction reads an SGPR (or EXEC, or M0).
 148 Then, a SALU/SMEM instruction writes the same SGPR.
 149
 150 Mitigated by:
 151 A VALU instruction or an `s_waitcnt vmcnt(0)` between the two instructions.
 152
 153 ### SMEMtoVectorWriteHazard
 154
 155 Triggered by:
 156 An SMEM instruction reads an SGPR. Then, a VALU instruction writes that same SGPR.
 157 Despite LLVM
 158
 159 Mitigated by:
 160 Any non-SOPP SALU instruction (except `s_setvskip`, `s_version`, and any non-lgkmcnt `s_waitcnt`).
 161
 162 ### Offset3fBug
 163
 164 Any branch that is located at offset 0x3f will be buggy. Just insert some NOPs to make sure no branch
 165 is located at this offset.
 166
 167 ### InstFwdPrefetchBug
 168
 169 According to LLVM, the `s_inst_prefetch` instruction can cause a hang.
 170 There are no further details.
 171
 172 ### LdsMisalignedBug
 173
 174 When there is a misaligned multi-dword FLAT load/store instruction in WGP mode,
 175 it needs to be split into multiple single-dword FLAT instructions.
 176
 177 ACO doesn't use FLAT load/store on GFX10, so is unaffected.
 178
 179 ### FlatSegmentOffsetBug
 180
 181 The 12-bit immediate OFFSET field of FLAT instructions must always be 0.
 182 GLOBAL and SCRATCH are unaffected.
 183
 184 ACO doesn't use FLAT load/store on GFX10, so is unaffected.
 185
 186 ### VcmpxPermlaneHazard
 187
 188 Triggered by:
 189 Any permlane instruction that follows any VOPC instruction.
 190 Confirmed by AMD devs that despite the name, this doesn't only affect v_cmpx.
 191
 192 Mitigated by: any VALU instruction except `v_nop`.
 193
 194 ### VcmpxExecWARHazard
 195
 196 Triggered by:
 197 Any non-VALU instruction reads the EXEC mask. Then, any VALU instruction writes the EXEC mask.
 198
 199 Mitigated by:
 200 A VALU instruction that writes an SGPR (or has a valid SDST operand), or `s_waitcnt_depctr 0xfffe`.
 201 Note: `s_waitcnt_depctr` is an internal instruction, so there is no further information
 202 about what it does or what its operand means.
 203
 204 ### LdsBranchVmemWARHazard
 205
 206 Triggered by:
 207 VMEM/GLOBAL/SCRATCH instruction, then a branch, then a DS instruction,
 208 or vice versa: DS instruction, then a branch, then a VMEM/GLOBAL/SCRATCH instruction.
 209
 210 Mitigated by:
 211 Only `s_waitcnt_vscnt null, 0`. Needed even if the first instruction is a load.