- Thread:
- Thread =
the execution unit in
CUDA
Each thread is executed by 1 CUDA core (= processor)
- Thread =
the execution unit in
CUDA
- Kernel:
- Kernel =
the program (function)
that is executed by
a CUDA thread
Example:
__global__ void hello( ) { printf("Hello World\n"); // CUDA C code // uses printf( ) in}CUDA C library
- Kernel =
the program (function)
that is executed by
a CUDA thread
- Properties of threads:
- Multiple threads can
simultaneously execute
(the same)
kernel program.
- The maximum number of simultaneously running threads depends on the GPU type !!!
- Threads are organized (= grouped together) in a certain structure to make management (scheduling) easier to do...
- Multiple threads can
simultaneously execute
(the same)
kernel program.
- Thread block:
- Threads are
organized (= grouped together) into
thread blocks:
- The number of threads in a thread block is limited to 512 or 1024 threads per block (depending on the type of GPU).
More later on these subjects:
- Thread blocks are conceptually
organized
into 1-dim, 2-dim or 3-dim arrays
of threads
(for ease of management).
- Each thread in a thread block has a unique thread ID.
- Threads are
organized (= grouped together) into
thread blocks:
- Special properties of
threads that are in the
same thread block:
- Threads
within the same
thread block
can communicate with each other
very efficiently
using the "shared" memory
that are accessible
only by threads
in the same thread block:
One thread can write a value in a shared variable (stored in the shared memory) and another thread can read the written value from the shared variable
- Also:
- Using this shared memory, (only) threads in the same thread block can synchronize (= "wait on") with each other
- Threads
within the same
thread block
can communicate with each other
very efficiently
using the "shared" memory
that are accessible
only by threads
in the same thread block:
- This also mean:
- Threads that are in
different thread blocks
cannot communicate via
the shared memory
- Therefore:
- Threads that are in
different thread blocks
cannot synchronize (= "wait")
with each other !!!
- Threads that are in
different thread blocks
cannot synchronize (= "wait")
with each other !!!
- Threads that are in
different thread blocks
cannot communicate via
the shared memory
- Grid:
- Grid =
a collection
of thread blocks of the
same dimension
that execute the
the same kernel (GPU) program
Grids will help you organize the threads that execute the same kernel program
- Grid =
a collection
of thread blocks of the
same dimension
that execute the
the same kernel (GPU) program
- Grid dimension:
- Dimension of a grid =
- # blocks in the grid and
- # threads in each (thread) block
Comment:
- This is a simplified definition
for "grid dimension"
of a 1-dimensional grid
- The grid dimension of higher dimensional grids are more complex to express...
- Dimension of a grid =
- Recall the
architecture of
a GPU processor:
- How the
threads and
thread blocks are
executed by
the multi-processors in
a CPU:
In other words:
So threads in the same thread block are running inside the same multi-processor !!!
The processors inside the same multi-processor can access the shared memory inside the multi-processor !!!
That's why the threads in the same thread block can communicate with each others via their shared memory !!!