flat_atomic_pk_add_bf16 GPU Native ISA AMD Vector

Flat Memory

flat_atomic_pk_add_bf16

Add a packed 2-component BF16 float value in the data register to a location in the flat aperture.

Encoding

Binary Layout (AMD CDNA 5)
IOFFSET
95:72
VADDR
71:64
unassigned
63
VSRC
62:55
TH
54:52
SCOPE
51:50
SVE
49
SCALE_OFFSET
48
unassigned
47:40
VDST
39:32
11101100
31:24
unassigned
23:22
01011010
21:14
unassigned
13:8
NV
7
SADDR
6:0
 

The ENC_VFLAT layout from AMD's machine-readable ISA specification (AMD CDNA 5). Opcode 90 in OP rebuilds 0x0000000000000000EC168000, an identifier AMD lists for this encoding. Bits marked unassigned have no field in the specification. Not the same in every generation: AMD CDNA 4 (layout not verifiable), AMD CDNA 3 (layout not verifiable), AMD RDNA 4 (field layout).

Format FLAT
Width 96 bits
Opcode 90
Identifier 0x0000000000000000EC168000

Operands

  • VDST
    Written. Operand must be a vector GPR. Uses an 8-bit operand field. Data format: Value can be anything. (OPR_VGPR, FMT_ANY)
  • VADDR
    Read. Operand must be a vector GPR. Uses an 8-bit operand field. Data format: Value can be anything. (OPR_VGPR, FMT_ANY)
  • VSRC
    Read. Operand must be a vector GPR. Uses an 8-bit operand field. Data format: Value can be anything. (OPR_VGPR, FMT_ANY)
  • SADDR
    Read. Any scalar GPR operand including VCC and NULL. Data format: Value can be anything. (OPR_SREG, FMT_ANY)

In AMD's order, as its machine-readable ISA specification lists them for the ENC_VFLAT encoding (AMD CDNA 5).

GFX Target Compatibility

Per-target GFX compatibility has not yet been verified for this instruction.

Floating-point addition handles NAN/INF/denorm.

Related

More in Flat Memory

Reference

AMDGPU / GFX ISA

Description

Add a packed 2-component BF16 float value in the data register to a location in the flat aperture. Store the original value from flat aperture into a vector register iff the SC0 bit is set.

Semantics

tmp = MEM[ADDR]; src = DATA; dst[31 : 16].bf16 = tmp[31 : 16].bf16 + src[31 : 16].bf16; dst[15 : 0].bf16 = tmp[15 : 0].bf16 + src[15 : 0].bf16; MEM[ADDR] = dst.b32; RETURN_DATA = tmp

Example

flat_atomic_pk_add_bf16 v[0:1], v2

A real instruction accepted by the LLVM assembler, taken verbatim from LLVM's own AMDGPU MC test suite (gfx12_asm_vflat.s). Not from an AMD document, and not authored here.

Sources