X-HEEP Matrix Spec

Version: 0.1

The X-HEEP Matrix Extension is closely derived from the T-HEAD RVM proposal. However, a number of features are absent to maintain hardware simplicity and efficiency. This document discusses these details.

Memory

Registers

There are 8 tile registers, m0-m7, each holding a 4x4 matrix of 32-bit values. Smaller datatypes like 8-bit and 16-bit can be packed into this space, yielding 4x16 and 4x8 configurations. However, the same load instructions are used; only the arithmetic changes (more below).

System memory model

The accelerator does not guarantee memory coherency with the CPU. That is, the effects of a matrix accelerator store instruction may not be visible in a load to the same address from the CPU, despite being in program order. The opposite is also true, a CPU store instruction may not be immediately visible to a matrix load.

Note that this is not modeled in the Spike simulator, all of the instructions will appear to have sequential consistency.

Instructions

Matrix multiplication

Let S represent the size of individual matrix elements in bytes. Let ms1 hold matrix A[4][16/S], ms2 hold B[4][16/S], and md hold C[4][4]. The pseudocode for the matrix multiplication instruction <mnemonic> md, ms1, ms2 is

for (int i = 0; i < 4; i++) {
    for (int j = 0; j < 4; j++) {
        for (int k = 0; k < (16/S); k++) {
            C[i][j] += A[i][k] * B[j][k];
        }
    }
}

That is, the instruction computes C += A * B^T. Below, the available mnemonics for different datatypes are given.

operand type	accumulator type	S	mnemonic
fp32	fp32	4	fmmacc.s
int32	int32	4	mmasa.w
int16	int32	2	mmada.h
int8	int32	1	mmaqa.b

Zero

Clear register md with 0s:

mzero md

Load/store

Load/store from address in rs1, row stride in bytes in rs2, for register md/ms1:

mld.w md, (rs1), rs2
mst.w ms1, (rs1), rs2

Element stride is assumed to be 4 in bytes; only 32-bit word memory instructions are available.

Instruction Encoding

All instructions share 7'b0101011 (CUSTOM 1) as the major opcode, and func3 is 3'b000.

Arithmetic

31 27	26 25	24	23 21	20 18	17 15	14 12	11 10	9 7	6 0	mnemonic
00001	00	0	ms2	ms1	md	func3	10	000	major opcode	fmmacc.s
11110	00	0	ms2	ms1	md	func3	10	000	major opcode	mmasa.w
11100	00	0	ms2	ms1	md	func3	01	000	major opcode	mmada.h
00010	00	0	ms2	ms1	md	func3	00	000	major opcode	mmaqa.b
11111	00	0	000	000	md	func3	00	000	major opcode	mzero

Load/store

31 27	26 25	24 20	19 15	14 12	11 10	9 7	6 0	mnemonic
00000	10	rs2	rs1	func3	10	md	major opcode	mld.w
00001	10	rs2	rs1	func3	10	ms1	major opcode	mst.w

Toolchain/Simulator

See BUILDING.md for instructions.