diff cleanup

This commit is contained in:
qazal
2026-03-23 12:36:03 +00:00
parent e0d151560a
commit 003bd9534c
+83 -40
View File
@@ -102,46 +102,89 @@ class InstOp(Enum):
VALU_CMPX = 0x73 # v_cmpx_*
class InstOpRDNA4(Enum):
"""SQTT instruction operation types for RDNA4 (gfx1200). Complete EINST enum from rocprof-trace-decoder gfx12wave.cpp."""
SALU = 0; SMEM = 1; SMEM_WR = 2; JUMP = 3; JUMP_NO = 4; CALL = 5; TRAP = 6; SALU_NO_EXEC = 7; FATAL_HALT = 8; MESSAGE = 9 # noqa: E702
VALU_1 = 10; VALU_TRANS = 11; VALU_B1 = 12; VALU_B2 = 13; VALU_B4 = 14; VALU_B16 = 15 # noqa: E702
VINTERP = 18; BARRIER_WAIT = 19; EXPREQ_GDS = 20; EXPREQ_GFX = 21 # noqa: E702
FLAT_RD_2 = 28; FLAT_WR_3 = 29; FLAT_WR_4 = 30; FLAT_WR_5 = 31; FLAT_WR_6 = 32 # noqa: E702
VMEM_RD_1 = 33; VMEM_RD_2 = 34; VMEM_WR_1 = 35; VMEM_WR_2 = 36; VMEM_WR_3 = 37; VMEM_WR_4 = 38; VMEM_WR_5 = 39; VMEM_WR_6 = 40 # noqa: E702
LDS_RD = 41; LDS_WR_1 = 42; LDS_WR_2 = 43; LDS_WR_3 = 44; LDS_WR_4 = 45; LDS_WR_5 = 46 # noqa: E702
BUF_RD_1 = 47; BUF_RD_2 = 48; BUF_WR_1 = 49; BUF_WR_2 = 50; BUF_WR_3 = 51; BUF_WR_4 = 52; BUF_WR_5 = 53; BUF_WR_6 = 54 # noqa: E702
IMG_SAMPLE_1 = 55; IMG_SAMPLE_2 = 56; IMG_SAMPLE_3 = 57; IMG_SAMPLE_4 = 58; IMG_SAMPLE_5 = 59; IMG_SAMPLE_6 = 60 # noqa: E702
IMG_SAMPLE_7 = 61; IMG_SAMPLE_8 = 62; IMG_SAMPLE_9 = 63; IMG_SAMPLE_10 = 64; IMG_SAMPLE_11 = 65; IMG_SAMPLE_12 = 66 # noqa: E702
IMG_SAMPLE_RESERVED = 67; IMG_RD_1 = 68; IMG_RD_2 = 69; IMG_RD_3 = 70; IMG_RD_4 = 71 # noqa: E702
IMG_WR_2 = 72; IMG_WR_3 = 73; IMG_WR_4 = 74; IMG_WR_5 = 75; IMG_WR_6 = 76; IMG_WR_7 = 77; IMG_WR_8 = 78 # noqa: E702
OTHER_LDS_1 = 80; OTHER_LDS_2 = 81; OTHER_LDS_3 = 82; OTHER_LDS_4 = 83; OTHER_LDS_5 = 84 # noqa: E702
OTHER_FLAT_2 = 85; OTHER_FLAT_3 = 86; OTHER_FLAT_4 = 87; OTHER_FLAT_5 = 88; OTHER_FLAT_6 = 89 # noqa: E702
RAYTRACE_8 = 103; RAYTRACE_9 = 104; RAYTRACE_11 = 105; RAYTRACE_12 = 106 # noqa: E702
LDS_DIR_LOAD = 110; LDS_PARAM_LOAD = 111; SALU_WR_EXEC = 114; VALU1_WR_EXEC = 115; VALU_B2_WR_EXEC = 116 # noqa: E702
RFE = 117; LDS_BVH_6 = 118; OTHER_LDS_6 = 119; OTHER_LDS_10 = 120; LDS_BVH_10 = 121 # noqa: E702
BARRIER_SIGNAL = 122; DYN_VGPR = 135; TRY_LOCK = 136; UNLOCK = 137; BARRIER_JOIN = 138 # noqa: E702
WMMA_8 = 140; WMMA_16 = 141; WMMA_32 = 142; WMMA_64 = 143; VALU_DPFP = 146; VALU_DERATE = 147 # noqa: E702
ICPREF = 150; KCPREF = 151; SALU_FLOAT3 = 152; VALU_SCL_TRANS = 153; SALU_2 = 155; SALU_5 = 156 # noqa: E702
# 188-221: vmem_other_simd range (OTHER_VMEM_1..OTHER_VMEM_34)
OTHER_VMEM_1 = 188; OTHER_VMEM_2 = 189; OTHER_VMEM_3 = 190; OTHER_VMEM_4 = 191 # noqa: E702
OTHER_VMEM_5 = 192; OTHER_VMEM_6 = 193; OTHER_VMEM_7 = 194; OTHER_VMEM_8 = 195 # noqa: E702
OTHER_VMEM_9 = 196; OTHER_VMEM_10 = 197; OTHER_VMEM_11 = 198; OTHER_VMEM_12 = 199 # noqa: E702
OTHER_VMEM_13 = 200; OTHER_VMEM_14 = 201; OTHER_VMEM_15 = 202; OTHER_VMEM_16 = 203 # noqa: E702
OTHER_VMEM_17 = 204; OTHER_VMEM_18 = 205; OTHER_VMEM_19 = 206; OTHER_VMEM_20 = 207 # noqa: E702
OTHER_VMEM_21 = 208; OTHER_VMEM_22 = 209; OTHER_VMEM_23 = 210; OTHER_VMEM_24 = 211 # noqa: E702
OTHER_VMEM_25 = 212; OTHER_VMEM_26 = 213; OTHER_VMEM_27 = 214; OTHER_VMEM_28 = 215 # noqa: E702
OTHER_VMEM_29 = 216; OTHER_VMEM_30 = 217; OTHER_VMEM_31 = 218; OTHER_VMEM_32 = 219 # noqa: E702
OTHER_VMEM_33 = 220; OTHER_VMEM_34 = 221 # noqa: E702
# 222-255: block_store range (BLOCK_STORE_1..BLOCK_STORE_34)
BLOCK_STORE_1 = 222; BLOCK_STORE_2 = 223; BLOCK_STORE_3 = 224; BLOCK_STORE_4 = 225 # noqa: E702
BLOCK_STORE_5 = 226; BLOCK_STORE_6 = 227; BLOCK_STORE_7 = 228; BLOCK_STORE_8 = 229 # noqa: E702
BLOCK_STORE_9 = 230; BLOCK_STORE_10 = 231; BLOCK_STORE_11 = 232; BLOCK_STORE_12 = 233 # noqa: E702
BLOCK_STORE_13 = 234; BLOCK_STORE_14 = 235; BLOCK_STORE_15 = 236; BLOCK_STORE_16 = 237 # noqa: E702
BLOCK_STORE_17 = 238; BLOCK_STORE_18 = 239; BLOCK_STORE_19 = 240; BLOCK_STORE_20 = 241 # noqa: E702
BLOCK_STORE_21 = 242; BLOCK_STORE_22 = 243; BLOCK_STORE_23 = 244; BLOCK_STORE_24 = 245 # noqa: E702
BLOCK_STORE_25 = 246; BLOCK_STORE_26 = 247; BLOCK_STORE_27 = 248; BLOCK_STORE_28 = 249 # noqa: E702
BLOCK_STORE_29 = 250; BLOCK_STORE_30 = 251; BLOCK_STORE_31 = 252; BLOCK_STORE_32 = 253 # noqa: E702
BLOCK_STORE_33 = 254; BLOCK_STORE_34 = 255 # noqa: E702
"""SQTT instruction operation types for RDNA4 (gfx1200). Different encoding from RDNA3."""
SALU = 0x0
SMEM = 0x1
SMEM_WR = 0x2
JUMP = 0x3
JUMP_NO = 0x4
CALL = 0x5
TRAP = 0x6
SALU_NO_EXEC = 0x7
FATAL_HALT = 0x8
MESSAGE = 0x9
VALU_1 = 0xa
VALU_TRANS = 0xb
VALU_B1 = 0xc
VALU_B2 = 0xd
VALU_B4 = 0xe
VALU_B16 = 0xf
VINTERP = 0x12
BARRIER_WAIT = 0x13
FLAT_RD_2 = 0x1c
FLAT_WR_3 = 0x1d
FLAT_WR_4 = 0x1e
FLAT_WR_5 = 0x1f
FLAT_WR_6 = 0x20
VMEM_RD_1 = 0x21
VMEM_RD_2 = 0x22
VMEM_WR_1 = 0x23
VMEM_WR_2 = 0x24
VMEM_WR_3 = 0x25
VMEM_WR_4 = 0x26
VMEM_WR_5 = 0x27
VMEM_WR_6 = 0x28
LDS_RD = 0x29
LDS_WR_1 = 0x2a
LDS_WR_2 = 0x2b
LDS_WR_3 = 0x2c
LDS_WR_4 = 0x2d
LDS_WR_5 = 0x2e
BUF_RD_1 = 0x2f
BUF_RD_2 = 0x30
BUF_WR_1 = 0x31
BUF_WR_2 = 0x32
BUF_WR_3 = 0x33
BUF_WR_4 = 0x34
BUF_WR_5 = 0x35
BUF_WR_6 = 0x36
OTHER_LDS_1 = 0x50
OTHER_LDS_2 = 0x51
OTHER_LDS_3 = 0x52
OTHER_LDS_4 = 0x53
OTHER_LDS_5 = 0x54
OTHER_FLAT_2 = 0x55
OTHER_FLAT_3 = 0x56
OTHER_FLAT_4 = 0x57
OTHER_FLAT_5 = 0x58
OTHER_FLAT_6 = 0x59
LDS_DIR_LOAD = 0x6e
LDS_PARAM_LOAD = 0x6f
SALU_WR_EXEC = 0x72
VALU1_WR_EXEC = 0x73
VALU_B2_WR_EXEC = 0x74
RFE = 0x75
OTHER_LDS_6 = 0x77
OTHER_LDS_10 = 0x78
BARRIER_SIGNAL = 0x7a
DYN_VGPR = 0x87
TRY_LOCK = 0x88
UNLOCK = 0x89
BARRIER_JOIN = 0x8a
WMMA_8 = 0x8c
WMMA_16 = 0x8d
WMMA_32 = 0x8e
WMMA_64 = 0x8f
VALU_DPFP = 0x92
VALU_DERATE = 0x93
ICPREF = 0x96
KCPREF = 0x97
SALU_FLOAT3 = 0x98
VALU_SCL_TRANS = 0x99
SALU_2 = 0x9b
SALU_5 = 0x9c
OTHER_VMEM = 0xbc # 0xbc-0xdd: vmem_other_simd
for _i in range(34): InstOpRDNA4._value2member_map_[0xbc + _i] = InstOpRDNA4.OTHER_VMEM
class InstOpCDNA(Enum):
SMEM_RD = 0