The Midgard Shader Core
ysh329 opened this issue · comments
Mali Midgard GPU架构优化细节
Midgard架构包括了T600、T700以及800系列，Arm官方给了对该架构的优化细节，下面将逐一展开，并结合我的理解。这部分内容主题来自其官网对Midgard GPU在OpenCL的文档。
kernel中所有线程结束的时间是相同的
- Midgard GPU的计算Branches are computationally cheap on Mali Midgard GPUs. This means you can use loops in kernels without any performance impact.
- Your kernels can include different code segments but try to ensure the kernels exit at the same time.
- A workaround to this is to use a bucket algorithm.
Make your kernel code as simple as possible
- This assists the auto-vectorization process.
- Using loops and branches might make auto-vectorization more difficult.
Use vector operations in kernel code
- Use vector operations in kernel code to help the compiler to map them to vector instructions.
Vectorize your code
- Mali Midgard GPUs perform computation with vectors. These enable you to perform multiple operations per instruction.
- Vectorizing your code makes the best use of the Mali Midgard GPU hardware so ensure that you vectorize your code for maximum performance.
- Mali Midgard GPUs contain 128-bit wide vector registers.
Note
The Midgard compiler can auto-vectorize some scalar code.
Vectorize incrementally
Vectorize in incremental steps. For example, start processing one pixel at a time, then two, then four.
Avoid processing single values
Avoid writing kernels that operate on single bytes or other small values. Write kernels that work on vectors.
Use 128-bit vectors
Vector sizes of 128-bits are optimal. Vector sizes greater than 128-bits are broken into 128-bit parts and operated on separately. For example, adding two 256-bit vectors takes twice as long as adding two 128-bit vectors. You can use vector sizes less than 128 bits without issue.
The disadvantage of using vectors greater than 128 bits is that they can increase code size. Increased code size uses more instruction cache space and this can reduce performance.