User documentation for GPU integration with Pyccel

developer_docs/GPU_getting_started.md

@@ -0,0 +1,86 @@
+# Getting started GPU
+
+## Decorators
+
+* Kernel
+
+ In Pyccel, the `@kernel` decorator is used to indicate that a function should be treated as a CUDA kernel. A CUDA kernel is a function that is executed on a GPU and is typically used to perform a parallel computation over a large dataset.
+
+By applying the @kernel decorator to a function, Pyccel recognizes it as a CUDA kernel that can be launched on a GPU using the appropriate syntax.
+
+> kernels can not return a variable that is why we pass the returned variable in the kernel arguments.
+
+ ```Python
+ from pyccel import cuda
+ from pyccel.decorators import kernel
+
+ @kernel
+ def my_kernel(x: 'float64[:]', y: 'float64[:]', out: 'float64[:]'):
+ i = cuda.threadIdx(0) + cuda.blockIdx(0) * cuda.blockDim(0)
+ if i >= x.shape[0]:
+ return
+ out[i] = x[i] + y[i]
+ ```
+
+ We can launch this kernel on a GPU using the following syntax:
+
+ ```Python
+ ng = 128 # size of the grid
+ tn = 64 # size of the block
+ my_kernel[ng, tn](x, y, out)
+ ```
+
+* Device
+
+ The `@device` decorator is similar to the `@kernel` decorator, but indicates that the function should be compiled and executed on the GPU as a device function, rather than a kernel.
+
+ Device functions are similar to kernels, but are executed within the context of a kernel. They can be called only from kernels, and are typically used for operations that are too small to justify launching a separate kernel, or for operations that need to be performed repeatedly within the context of a kernel.
+
+ ```Python
+ from pyccel import cuda
+ from pyccel.decorators import device
+
+ @device
+ def my_device_function(x: 'float32'):
+ return x * x
+
+ # Call the function from a kernel
+ @kernel
+ def my_kernel(x: 'float32[:]', y: 'float32[:]'):
+ i = cuda.threadIdx(0) + cuda.blockIdx(0) * cuda.blockDim(0)
+ if i >= x.shape[0]:
+ return
+ y[i] = my_device_function(x[i])
+ ```
+
+## Built-in variables
+
+* cuda.threadIdx(dim_tpb): Returns the index of the current thread within the block in a given dimension dim_tpb.
+
+* cuda.blockDim(dim_tpb): Returns the size of the thread block in a given dimension dim_tpb.
+
+> dim_tpb: is an integer between 0 and the number of dimensions of the thread block minus one.
+
+* cuda.blockIdx(dim_grid): Returns the index of the current block within the grid in a given dimension dim.
+
+* cuda.gridDim(dim_grid): Returns the size of the grid in a given dimension dim.
+
+> dim_grid is an integer between 0 and the number of dimensions of the grid minus one.
+
+These built-in variables are provided by the CUDA runtime and are available to all CUDA kernels, regardless of the programming language being used. They are used to obtain information about the execution configuration of the kernel, such as the indices and dimensions of the threads and blocks. They are an important part of the CUDA programming model and are often used to compute global indices and other information about the execution configuration of the kernel.
+
+ ```Python
+ from pyccel import cuda
+ from pyccel.decorators import kernel
+
+ @kernel
+ def my_kernel(x: 'float64[:]', y: 'float64[:]', out: 'float64[:]'):
+ i = cuda.threadIdx(0) + cuda.blockIdx(0) * cuda.blockDim(0)
+ if i >= x.shape[0]:
+ return
+ out[i] = x[i] + y[i]
+ ```
+
+The kernel uses the `cuda.threadIdx(0)`, `cuda.blockIdx(0)`, and `cuda.blockDim(0)` built-in variables to compute the global index of the current thread within the 1D input arrays.
+
+The global index is computed as the sum of the thread index and the block index, multiplied by the block size, in the first dimension. This allows each thread to compute its own index within the input arrays, and to exit if its index falls outside the bounds of the arrays.