20 “P” Standard Extension for Packed-SIMD Instructions, Version 0.1

Discussions at the 5th RISC-V workshop indicated a desire to drop this packed-SIMD proposal for floating-point registers in favor of standardizing on the V extension for large floating-point SIMD operations. However, there was interest in packed-SIMD fixed-point operations for use in the integer registers of small RISC-V implementations.

In this chapter, we outline a standard packed-SIMD extension for RISC-V. We’ve reserved the instruction subset name “P” for a future standard set of packed-SIMD extensions. Many other extensions can build upon a packed-SIMD extension, taking advantage of the wide data registers and datapaths separate from the integer unit.

Packed-SIMD extensions, first introduced with the Lincoln Labs TX-2 [tx2], have become a popular way to provide higher throughput on data-parallel codes. Earlier commercial microprocessor implementations include the Intel i860, HP PA-RISC MAX [lee-max-ieeemicro1996], SPARC VIS [tremblay-vis-ieeemicro1996], MIPS MDMX [gwennap-mdmx-mpr1996], PowerPC AltiVec [diefendorff-altivec-ieeemicro2000], Intel x86 MMX/SSE [peleg-mmx-ieeemicro1996 raman-sse-ieeemicro2000], while recent designs include Intel x86 AVX [lomont-avx-irm2011] and ARM Neon [goodacre-armisa-computer2005]. We describe a standard framework for adding packed-SIMD in this chapter, but are not actively working on such a design. In our opinion, packed-SIMD designs represent a reasonable design point when reusing existing wide datapath resources, but if significant additional resources are to be devoted to data-parallel execution then designs based on traditional vector architectures are a better choice and should use the V extension.

A RISC-V packed-SIMD extension reuses the floating-point registers (f0-f31). These registers can be defined to have widths of FLEN=32 to FLEN=1024. The standard floating-point instruction subsets require registers of width 32 bits (“F”), 64 bits (“D”), or 128 bits (“Q”).

It is natural to use the floating-point registers for packed-SIMD values rather than the integer registers (PA-RISC and Alpha packed-SIMD extensions) as this frees the integer registers for control and address values, simplifies reuse of scalar floating-point units for SIMD floating-point execution, and leads naturally to a decoupled integer/floating-point hardware design. The floating-point load and store instruction encodings also have space to handle wider packed-SIMD registers. However, reusing the floating-point registers for packed-SIMD values does make it more difficult to use a recoded internal format for floating-point values.

The existing floating-point load and store instructions are used to load and store various-sized words from memory to the f registers. The base ISA supports 32-bit and 64-bit loads and stores, but the LOAD-FP and STORE-FP instruction encodings allows 8 different widths to be encoded as shown in Table [psimdwidth]. When used with packed-SIMD operations, it is desirable to support non-naturally aligned loads and stores in hardware.

width field	Code	Size in bits
000	B	8
001	H	16
010	W	32
011	D	64
100	Q	128
101	Q2	256
110	Q4	512
111	Q8	1024

Packed-SIMD computational instructions operate on packed values in f registers. Each value can be 8-bit, 16-bit, 32-bit, 64-bit, or 128-bit, and both integer and floating-point representations can be supported. For example, a 64-bit packed-SIMD extension can treat each register as 1×64-bit, 2×32-bit, 4×16-bit, or 8×8-bit packed values.

Simple packed-SIMD extensions might fit in unused 32-bit instruction opcodes, but more extensive packed-SIMD extensions will likely require a dedicated 30-bit instruction space.