Question: I recall from lecture that we were supposed to 'clear the slate' in our minds as to avoid terminology confusion between SIMD and CUDA since they use similar terms such as threads. However, I feel like though SIMD threads and CUDA threads are different in terms of location where they run (CPU vs GPU), they are still fundamentally quite the same. They execute in parallel with one another and return when all are finished. In thinking this, am I missing something?
This comment was marked helpful 0 times.
kayvonf
Clarification: I advised that to avoid confusion, it would be wise to recognize that the pthread API, the ISPC programming language, and the CUDA programming language all use very different, and somewhat inconsistent terminology. (Not SIMD and CUDA... since SIMD is not a language).
Depending on context, the term "thread" can sometimes be used to generically refer to a "thread of control in a program", to a hardware execution context, to a pthread (a logical thread of control mapped to an execution context on a CPU core), or a CUDA thread (a logical thread of control that is mapped to a single "lane" of a SIMD set of ALUs in a GPU).
But yes, both a pthread, an ISPC program instance, and a CUDA thread present the program the abstraction of a logical "thread" of control. They just have very different mappings to hardware.
This comment was marked helpful 0 times.
GG
Although CUDA is initially designed to handle computational tasks, it can also be used for Graphics. We can map some resources from OpenGL and Direct3D into the address space of CUDA, to enable them to interoperate with each other. In this way, we can reduce data transfer between the CPU and GPU, which causes a significant bottleneck in performance and we can also improve the performance of many graphical algorithms. More details can be found here.
Question: I recall from lecture that we were supposed to 'clear the slate' in our minds as to avoid terminology confusion between SIMD and CUDA since they use similar terms such as threads. However, I feel like though SIMD threads and CUDA threads are different in terms of location where they run (CPU vs GPU), they are still fundamentally quite the same. They execute in parallel with one another and return when all are finished. In thinking this, am I missing something?
This comment was marked helpful 0 times.
Clarification: I advised that to avoid confusion, it would be wise to recognize that the pthread API, the ISPC programming language, and the CUDA programming language all use very different, and somewhat inconsistent terminology. (Not SIMD and CUDA... since SIMD is not a language).
Depending on context, the term "thread" can sometimes be used to generically refer to a "thread of control in a program", to a hardware execution context, to a pthread (a logical thread of control mapped to an execution context on a CPU core), or a CUDA thread (a logical thread of control that is mapped to a single "lane" of a SIMD set of ALUs in a GPU).
But yes, both a pthread, an ISPC program instance, and a CUDA thread present the program the abstraction of a logical "thread" of control. They just have very different mappings to hardware.
This comment was marked helpful 0 times.
Although CUDA is initially designed to handle computational tasks, it can also be used for Graphics. We can map some resources from OpenGL and Direct3D into the address space of CUDA, to enable them to interoperate with each other. In this way, we can reduce data transfer between the CPU and GPU, which causes a significant bottleneck in performance and we can also improve the performance of many graphical algorithms. More details can be found here.
This comment was marked helpful 0 times.