v_pk_fma_f16 GPU Native ISA AMD Vector

Vector Packed Arithmetic

v_pk_fma_f16

Multiply two packed half-precision float inputs component-wise and add a third input component-wise using fused multiply add, and store the result…

Encoding

Binary Layout (AMD CDNA 5)
NEG
63:61
OPSEL_HI[1:0]
60:59
SRC2
58:50
SRC1
49:41
SRC0
40:32
11001100
31:24
00001110
23:16
CLAMP
15
OPSEL_HI[2]
14
OPSEL
13:11
NEG_HI
10:8
VDST
7:0
 

The ENC_VOP3P layout from AMD's machine-readable ISA specification (AMD CDNA 5). Opcode 14 in OP rebuilds 0x00000000CC0E0000, an identifier AMD lists for this encoding. Not the same in every generation: AMD CDNA 4 (field layout), AMD CDNA 3 (field layout), AMD CDNA 2 (field layout), AMD CDNA 1 (field layout), AMD RDNA 4 (field layout), AMD RDNA 3.5 (field layout), AMD RDNA 3 (field layout), AMD RDNA 2 (field layout), AMD RDNA 1 (field layout).

Format VOP3P
Width 64 bits
Opcode 14
Identifier 0x00000000CC0E0000

Operands

  • VDST
    Written. Operand must be a vector GPR. Uses an 8-bit operand field. Data format: 16-bit half-precision floating point value. Packed 2-component value. (OPR_VGPR, FMT_NUM_PK2_F16)
  • SRC0
    Read. All operands. Covers all operands that are allowed as scalar or vector sources. Data format: 16-bit half-precision floating point value. Packed 2-component value. (OPR_SRC, FMT_NUM_PK2_F16)
  • SRC1
    Read. All operands. Covers all operands that are allowed as scalar or vector sources. Data format: 16-bit half-precision floating point value. Packed 2-component value. (OPR_SRC, FMT_NUM_PK2_F16)
  • SRC2
    Read. All operands. Covers all operands that are allowed as scalar or vector sources. Data format: 16-bit half-precision floating point value. Packed 2-component value. (OPR_SRC, FMT_NUM_PK2_F16)

In AMD's order, as its machine-readable ISA specification lists them for the ENC_VOP3P encoding (AMD CDNA 5).

GFX Target Compatibility

TargetSupport
gfx1100✅ Supported

Related

More in Vector Packed Arithmetic

Reference

AMDGPU / GFX ISA

Description

Multiply two packed half-precision float inputs component-wise and add a third input component-wise using fused multiply add, and store the result into a vector register.

Semantics

declare tmp : 32'B; tmp[15 : 0].f16 = fma(S0[15 : 0].f16, S1[15 : 0].f16, S2[15 : 0].f16); tmp[31 : 16].f16 = fma(S0[31 : 16].f16, S1[31 : 16].f16, S2[31 : 16].f16); D0.b32 = tmp

Example

v_pk_fma_f16 v5, v1, v2, s3

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

Sources