• State Accepted Answer
  • Locked Locked
  • Replies 3 replies
  • Subscribers 136 subscribers
  • Views 1433 views
  • Users 0 members are here
Options
Related
How was your experience today?
This discussion has been locked.
You can no longer post new replies to this discussion. If you have a question you can start a new discussion

Mali Memory Bandwidth and Access Pattern

opm

Hi there,

I am playing around with the Mali-G78 in a Pixel 6a and Streamline.

As a trial I want to copy the content of a 64MB buffer into another - here my shader code:

```
#version 450

layout (local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
layout(constant_id = 0) const int kBufferSizeElements = (64*1024*1024)/4;
layout(set = 0, binding = 0) buffer InputBuffer {uint input_buffer[kBufferSizeElements];};
layout(set = 0, binding = 1) buffer OutputBuffer {uint output_buffer[kBufferSizeElements];};

void main() {
   output_buffer[gl_GlobalInvocationID.x] = input_buffer[gl_GlobalInvocationID.x];
}

```

When I use Streamline to check the GPU counters, I observe the following:

  • the memory bandwidth (read and write together) are roughly 30 GB/s, even though the spec of the Pixel 6a talks of 50GB/s max. Is this max value realistic to achieve or am I already relatively high with the 30 GB/s?
  • the load/store counters indicate that I am doing partial read/writes only - there are no full reads/writes reported. This I don't understand at all. I expect the warp, so 16 threads, to access the L/S unit with all adjacent addresses. According to the "Mali-G78 Performance Counters Reference Guide" threads should access a 64 byte range for max performance, which I do I think. What is the reason for these partial read/writes ? 
  • when I change the buffers from type `uint` to `uvec4` (and dispatch 4x less), then I get rid of the partial reads/writes. However, performance drops from 30GB/s to less than 25GB/s, memory write stalls increase from 46% to 62% and the time for the copy increases from ~4.5ms to ~5.5ms. Why is it getting slower when I have full read/writes only ? Obviously, I am accessing the memory in a less optimal way, but why?

Generally, can you explain how the memory should be accessed for best performance ?

And finally a short question on the datasheet: the "FP32 operations/cycle" is this per Arithmetic unit or per core ?

Parents
  • +1
    opm said:
    Is this max value realistic to achieve or am I already relatively high with the 30 GB/s?

    I can't speak to any Pixel 6 specifics, but generally you'll never hit 100% DRAM utilization in real world usage scenarios. Hitting 75% of theoretical is probably as good as it gets before DRAM paging overheads start to dominate. 

    opm said:
    What is the reason for these partial read/writes ? 

    For most Mali cores accesses are coalesced in groups of 4 threads, not across the whole warp. This is why the uvec4 helps - you only need 4 threads to fill the width of the cache load path. The latest Immortalis-G715 can coalesce across the whole warp, so this should improve on newer microarchitectures.

    opm said:
    Why is it getting slower when I have full read/writes only ?

    Is the only change to the shader the change to a uvec4 access size? Is your workgroup size is still 16 threads?

    TBH it's hard to say without more data; memory performance close to peak is very sensitive to small changes in timing and thread scheduling.  The large increase in write stalls implies the issue may be outside of the GPU (i.e. that's the external bus saying it's busy). Perhaps the more efficient access pattern caused the bus or DRAM clock to drop?

Reply
  • +1
    opm said:
    Is this max value realistic to achieve or am I already relatively high with the 30 GB/s?

    I can't speak to any Pixel 6 specifics, but generally you'll never hit 100% DRAM utilization in real world usage scenarios. Hitting 75% of theoretical is probably as good as it gets before DRAM paging overheads start to dominate. 

    opm said:
    What is the reason for these partial read/writes ? 

    For most Mali cores accesses are coalesced in groups of 4 threads, not across the whole warp. This is why the uvec4 helps - you only need 4 threads to fill the width of the cache load path. The latest Immortalis-G715 can coalesce across the whole warp, so this should improve on newer microarchitectures.

    opm said:
    Why is it getting slower when I have full read/writes only ?

    Is the only change to the shader the change to a uvec4 access size? Is your workgroup size is still 16 threads?

    TBH it's hard to say without more data; memory performance close to peak is very sensitive to small changes in timing and thread scheduling.  The large increase in write stalls implies the issue may be outside of the GPU (i.e. that's the external bus saying it's busy). Perhaps the more efficient access pattern caused the bus or DRAM clock to drop?

Children