CUDA x86

Not really a problem, unlike GPUs CPUs support load-execute instructions and the ALU:MEM ratio is much more favorable.

 

Sure but that makes it relatively *more* bad to waste instructions on predication. Instructions on CPUs are much more precious than on GPUs.

 

CPUs support Load/execute instructions, as in reg/mem operands?
A lot of CPUs don't support those.
It's not a particularly important distinction between the CPU and GPU realms.

 

True, so I mean, x86 CPUs, other may have more registers and a more sane SIMD, but anyway ALU:MEM ratio for CPUs still favorable.

 

IMO it's an important distinction; with load/execute instructions and in-order-execution each cache miss generated by one of those instructions means you lose a HW thread. On a in-order core without load/execute instruction the compiler can load data early and postpone execution of instructions that depend on that as much as necessary (and balance register pressure..) in order to increase your chance to hide latency without blocking a thread.

 

That's a distinction between different architectures, but it isn't a CPU vs GPU distinction.

In the case of x86, the memory operand is optional, though sorely needed because of the small number of registers in standard mode.
Nvidia claimed that with Fermi it revamped its internal ISA to remove reg/mem instructions and to go load/store, so there were GPUs that did have load/execute.

 

Every x86 supports reg, mem operations. Every single one. Read the ISA.

David

 

x86 isn't the only CPU out there.
I'm saying that reg/mem is a distinction between architectures, but it is not one that can distinguish between a CPU and GPU.
G80's internal ISA, per Nvidia, had memory operand instructions, and non-x86 CPUs can be load/store.

 

For CUDA on x86, it turns out that x86 is the only CPU out there...

DK

 

There are at least two architectures with reg/mem instructions in CUDA, unless we want to up the semantics ante by stating some arbitrary compute level as being CUDA-enough: G80 and x86. (edit: or 2.5, GT200 and possibly G80, then x86)

Granted, there are no GPUs in CUDA x86, so we can go one more step and nullify this side discussion.