1. It's not a special logic, it's adoption of OpenCL or compatible stuff, over time.
It's (effectively) the ability to use OpenCL for running more application-specific logic. Though this is already possible the idea is that the main system memory will be shared between the processor and GPU and therefore OpenCL will have access to the same memory space as applications. This does raise some important considerations about how this will operate within a Virtualized Memory environment and/or whether this is something that can be implemented without low-level Kernel support. OpenCL code written for the purpose of running on HSA won't run on other systems- that means that development of a program for HSA support would need to either not support systems that don't support HSA capabilities or they will need to write it twice. The thing about the "over time" qualifier is that has been said about a lot of new things. For example, VL Bus was going to penetrate the market "over time". The fact about technology is that if nobody supports it or uses it, it dies, regardless of any possible technical benefits. MCA was objectively superior to any of the competing bus designs at the time, but it failed for largely political reasons.
Using APIs like OpenCL has its immediate benefits as well as future benefits.
Generally- right now, that is, OpenCL is for shaders (and scientific processing with GPUs). In the latter case you need to copy the information the OpenCL will use to VRAM. This copying is of course slower than not copying- thing is, though, VRAM is many times faster than System Memory, which is a consideration largely unaddressed by the literature AMD provides. Otherwise, HSA, despite their insistence to the contrary, appears to effectively change how integrated GPUs share system memory, allowing them (in some manner) to use the same memory without 'partitioning' it (in some sense, making them more integrated)
I also think it should be close in syntax to C, so it's not like moving from C to .. I don't know, perl.
OpenCL typically uses a sort standardized C dialect, yes. But outside of low-level software, C tends not to be used for much. C++ tends to be as low as you will find the typical desktop application going. And OpenCL is a different platform and language not only from C but particularly from C++. In order to implement HSA support in a C++ project, existing logic will need to be ported to OpenCL, or new implementations will need to be written against both OpenCL as well as C++. It is a barrier to entry that comes with significant "negative points" in terms of costing, as it would require writing the same thing twice, or dropping support for any system that cannot run their OpenCL implementations (via HSA). Presumably, of course, if it gains enough traction, most developers won't have to worry about it- most JIT's would eventually implement support for it anyway.
2. Mantle achieved at least two goals for AMD, as far as I see - marketing and mantling DX12. Marketing, because now people know AMD has this thing and DX12, because they worked with MS and "contributed" their knowledge and way of thinking to Windows 10, which means better adoption for them
AMD doesn't appear to have worked any more closely with Microsoft than is typical for hardware manufacturers. Not aware of any specific contributions AMD would have made towards Windows 10, either- I mean, beyond the standard fare where Microsoft works with hardware manufacturers in general.
HSA is already in use. LibreOffice Calc uses it and benchmarks show it results in a huge improvements over just OpenCL with no HSA
Alright, fair enough. I was wrong. I was going to make a quip about LibreOffice here, but can't think of anything clever. So pretend I made a clever yet biting comment about LibreOffice.
From what I can gather- (And annoyingly, it's just laymans articles) the performance improvement was after significant refactoring. It seems to suggest the improvements were measured based on performance before the refactoring and performance after. I also downloaded the source for LibreOffice but can only find clew.c that seems to have any reference to OpenCL, which nothing seems to reference, which seems odd. maybe it was a test build?
If they measured improvements, that doesn't necessarily indicate the benefits were due to the changes themselves; their original logic may have been subpar or slower, for example; I presume it may also be comparing via the use of an AMD APU, by enabling or disabling, within the software, new OpenCL/HSA stuff. However, AMD's "Normal" CPU cores have also tended to have rather poor floating point performance (but better Integer performance), so using the GPU core may merely be covering up an existing deficiency in the main CPU core. Realistically, this is very likely what HSA is designed to do- their CPU technologies sort of stagnated (except top-notch virtualization instruct set support), but their GPU's moved forward; so it makes sense to try to take advantage of their GPU talent by having a GPU core try to help out the main CPU core and 'accelerate' it (thus, presumably, the term APU for the whole package).
if I seem somewhat hostile towards HSA, it's because all I can find from AMD is marketing garbage. When you (well not you, obviously, but AMD) write a claim in an infographic that something "will enable the creation of programs that cost less and take less time to develop, and provide more performance while using less power on HSA-enabled hardware." My bs detector goes off, because marketing wizards say that about everything. They said it about waterfall, they said it about agile and scrum. And as it turns out, there is no silver bullet. And HSA is being marketed as a silver bullet. Boom. Instant skeptical blacklist. The same infographic also says that HSA will "move humanity forward"... Talk about losing perspective.
So far, contradicting my "nobody will take advantage of it" claim, (according to the same doc) there are 6 applications that have implemented HSA; (this may be out of date- looks to be about a half-year old, so not sure if there are more) the aforementioned LibreOffice, AIDA64, Corel Aftershot Pro, PCMark, SiSoft Sandra, and AMD Catalyst. 6 Applications is not very many, especially when only 2 of those programs are actual applications (LibreOffice and Aftershot), and the others are benchmark tools.
