Well, supposedly RISC-V implementations will have none of this malarkey while still rivaling x64/ARM64 in processing speed at comparable technology/clock rates/prices, just with plain old loads-and-xors-and-stores?
Well, supposedly RISC-V implementations will have none of this malarkey while still rivaling x64/ARM64 in processing speed at comparable technology/clock rates/prices, just with plain old loads-and-xors-and-stores?
The core of modern RISC thought is basically: "The laws of physics mean that no matter how much hardware you throw at it, only some kinds of instructions can be implemented in a performant way. We should only include these kinds of instructions in the instruction set." Then you build more complex operations out of these simple building blocks, but the fact that every instruction provided can be reasonably implemented to run really fast, the CPU itself can be fast.
Masked vector adds belong in the set of instructions that can be implemented to be fast, and that's why they are included in the RVV RISC-V extension. An example of an instruction that cannot be implemented to be fast would be the humble x86 load+add, where you first look up a value in memory, and then add it to a register. The only reasonable way to implement this to be fast is to just split it into two separate operations which are also dispatched separately, and that is precisely what modern x86 does.
Though, granted, RVV is significantly more uniform than AVX-512 (albeit at the cost of not having some useful goodies).
The proposal originates with Andes, and one of their own ISAs. They have several RISC-V cores with an early draft of it.
All while being scalable, i.e. minimum vector register width (VLEN) is 128-bit for the 'v' extension, but hardware can implement up to 65536-bit vectors (and software can choose to either pretend they're 128-bit, or can be written such that it portably scales automatically); and if you want more than 128 bits portably there's LMUL, allowing grouping registers up to 8 in a group, giving up to at-least-1024-bit registers.
For shuffles it has vrgather, which supports all element width lookups and can move any element to any other element (yes, including at LMUL=8, though as you can imagine it can be expected to slow down quadratically with LMUL; and could even become a problem at LMUL=1 for hardware with large VLEN, whenever that becomes a thing).