I’m an optimization algorithm engineer, and I’m very interested on the hardware architectures of ARM Cortex-A? and Mali Midgard. As far as I know for the ARM Cortex-A?, the memory system of ARM consists of the following:
L1 Cache： L1D/L1P
L2 Cache
I searched Internet for Mail Midgard hardware architecture, and found this:
L1 Cache： Data/Texture Cache
L2 Cache (global memory and constant memory)
I wonder if Midgard has any private memory (Shader Core exclusive) and local memory (sharing in cluster), as ARM has the L1P cache code, Midgard is supposed to have the similar Cache/Memory.
Pls share with me the detailed information about the Midgard hardware architecture, many thanks.
There isn't a huge amount of public information about the hardware architecture available; it's not something that a typically user needs to know as it is mostly hidden by drivers. We have a public technical overview here:
The Mali GPU: An Abstract Machine, Part 3 - The Midgard Shader Core
I wonder if Midgard has any private memory (Shader Core exclusive) and local memory (sharing in cluster)
No - only the generic L1/L2 data caches caches for compute applications (the framebuffer tile memory is local to a shader core for fragment shading graphics workloads).
HTH,
Pete
Dear Pete
The figure above shows that one Shader Core has one Load/Store Pipe, one Texture Pipe, and two Arithmetic Pipes. My question is when multiple Arithmetic Pipes are operating in parallel (especially 4 Arithmetic Pipes), the performance of data read and write would limit the Arithmetic Pipe's operation capability?
And, is it possible for each Arithmetic Pipe in a Shader Core to run different applications at the same time?
The figure above tells that each Arithmetic Pipe has a register file, I wonder the organization and size of the file, please.
For example, ARMv7 Neon has 32 D(64bit) registers or 15 Q(128bit) registers, but ARMv8 has a different register organization.
It is going to depend on your algorithm and the ratio of memory access to arithmetic.
If you are running a program where that matters you are using it wrong. GPGPU workloads need lots (ideally a minimum of high tens of thousands) of parallel work items / threads to fill the GPU; so statistically you will be running one program at a time.
I wonder the organization and size of the register file
For Midgard GPUs this would be a useful starting point:
ARM Mali Compute Architecture Fundamentals
HTH, Pete