BFloat16 sparse outer product, accumulating
This instruction generates BFloat16 outer product by multiplying the 1-in-2 selected elements from the dense sub-matrices in the two first source vectors with the corresponding elements of the compressed sparse sub-matrix in the second source vector and accumulates the results to the destination ZA tile elements.
The outer product is generated by multiplying the selected 1-in-2 BFloat16 value from each overlapping 16-bit containers of the two SVLH×1 sub-matrices in the first source vectors by the BFloat16 value from the corresponding 16-bit container of the compressed 1×SVLH sub-matrix in the second source vector.
The 1-in-2 BFloat16 value in the first source vectors is selected by 2-bit controls in the indexed segment of the control vector register. If the control bit corresponding to an element in the first source vectors is 0, the element is discarded and does not contribute to the result. If both bits of the 2-bit control corresponding to the elements of the first source vectors are 1, only the element corresponding to the least significant bit is selected.
The instruction multiplies the selected elements of sub-matrices of BFloat16 values held in the first source vectors by the corresponding elements of sub-matrix of BFloat16 values in the second source vector. The resulting BFloat16 SVLH×SVLH outer product is then destructively added to the BFloat16 destination tile. This is equivalent to performing a single multiply-accumulate to each of the destination tile elements.
This instruction follows SME2 ZA-targeting non-widening BFloat16 numerical behaviors.
This instruction is unpredicated.
| 31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
| 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | Zm | 0 | 0 | 0 | K | Zk | Zn | i2 | 1 | 0 | 0 | ZAda | |||||||||
if !IsFeatureImplemented(FEAT_SME_TMOP) || !IsFeatureImplemented(FEAT_SME_B16B16) then EndOfDecode(Decode_UNDEF); constant integer n = UInt(Zn:'0'); constant integer m = UInt(Zm); constant integer k = UInt('1':K:'1':Zk); constant integer index = UInt(i2); constant integer da = UInt(ZAda);
| <ZAda> |
Is the name of the ZA tile ZA0-ZA1, encoded in the "ZAda" field. |
| <Zn1> |
Is the name of the first scalable vector register of the first source multi-vector group, encoded as "Zn" times 2. |
| <Zn2> |
Is the name of the second scalable vector register of the first source multi-vector group, encoded as "Zn" times 2 plus 1. |
| <Zm> |
Is the name of the second source scalable vector register, encoded in the "Zm" field. |
| <Zk> |
Is the name of the control vector register Z20-Z23 or Z28-Z31, encoded in the "K:Zk" fields. |
| <index> |
Is the control segment index, in the range 0 to 3, encoded in the "i2" field. |
CheckStreamingSVEAndZAEnabled(); constant integer VL = CurrentVL; constant integer dim = VL DIV 16; constant integer csize = VL DIV 8; constant bits(VL) op2 = Z[m, VL]; constant bits(VL) op3 = Z[k, VL]; constant bits(csize) ctrl = Elem[op3, index, csize]; constant bits(dim*dim*16) op4 = ZAtile[da, 16, dim*dim*16]; bits(dim*dim*16) result; for row = 0 to dim-1 for col = 0 to dim-1 integer i = 0; bits(16) elem1 = BFZero('0', 16); for r = 0 to 1 constant bits(VL) op1 = Z[n+r, VL]; if i < 1 && Elem[ctrl, 2*col + r, 1] == '1' then elem1 = Elem[op1, row, 16]; i = i + 1; constant bits(16) elem2 = Elem[op2, col, 16]; constant bits(16) sum = Elem[op4, row*dim+col, 16]; Elem[result, row*dim+col, 16] = BFMulAdd_ZA(sum, elem1, elem2, FPCR); ZAtile[da, 16, dim*dim*16] = result;
Internal version only: aarchmrs v2024-12_rel, pseudocode v2024-12_rel ; Build timestamp: 2024-12-15T22:18
Copyright © 2010-2024 Arm Limited or its affiliates. All rights reserved. This document is Non-Confidential.