Revert "Update ROCm doc (#35)"

This reverts commit 98ec60d.

index.md

@@ -28,7 +28,7 @@ All versions of Julia are supported, on Linux and Windows, and the functionality
 
 Similar, but much newer capabilities exist for [**Intel GPUs with oneAPI**](/oneapi/). Currently, full-featured kernel programming capabilities are available, but there is no support for vendor libraries such as oneMKL or oneDNN yet.
 
-Maturing support exists for [**AMD GPUs**](/rocm/) running on the ROCm stack. These GPUs can again be programmed in Julia at the kernel level or using high-level operations on arrays. Latest versions of Julia are supported, and the functionality is increasingly used by a variety of applications and libraries.
+Experimental support exists for [**AMD GPUs**](/rocm/) running on the ROCm stack. These GPUs can again be programmed in Julia at the kernel level or using array operations, but these capabilities are under heavy development and are not ready for general consumption yet.
 
 Experimental support also exists for [**Apple GPUs**](/metal/). Array programming and kernel programming are both supported.
 

rocm.md

@@ -18,36 +18,19 @@ title = "AMD ROCm"
 The Julia programming support for AMD GPUs based on the ROCm platform aims to
 provide similar capabilities as the [NVIDIA CUDA](/cuda/) stack, with support
 for both low-level kernel programming as well as an array-oriented interface.
-[AMDGPU.jl](https://github.com/JuliaGPU/AMDGPU.jl) offers comparable performance
-as HIP C++. The toolchain can easily be installed on latest version of Julia 
-using the integrated package manager.
 
-AMDGPU.jl makes it possible to program AMD GPUs at different abstraction levels:
+Julia support exists in the form of a single package:
 
-- by using the `ROCArray` type, providing a user-friendly yet powerful abstraction
-  that does not require any GPU programming experience;
-- by writing ROC kernels, with similar performance as kernels written in HIP C++;
-- by interfacing with HIP APIs and libraries directly, similar level of flexibility
-  you would expect from a C-based programming environment.
+- [AMDGPU.jl](https://github.com/JuliaGPU/AMDGPU.jl)
 
-The [documentation](https://amdgpu.juliagpu.org/stable/) of AMDGPU.jl demonstrates
-each of these approaches.
+This package contains everything needed to access the HSA runtime, program GPU
+kernels, and utilize a user-friendly array-based interface.
+The stack originally was divided into 3 separate packages, which still exist and may be of use for interested users and developers:
 
-# Performance
+- [HSARuntime.jl](https://github.com/JuliaGPU/HSARuntime.jl): interfacing with the HSA runtime
+- [AMDGPUnative.jl](https://github.com/JuliaGPU/AMDGPUnative.jl): GPU kernel programming support
+- [ROCArrays.jl](https://github.com/JuliaGPU/ROCArrays.jl): array programming interface
 
-Julia on the CPU is known for its good performance, approaching that of statically
-compiled languages like C. The same holds for programming AMD GPUs with kernels
-written using AMDGPU.jl, where we show preliminary performance to approach that of
-HIP C++ on a memcopy and 2D diffusion kernel:
-
-<div class="card mb-3">
-  <a href="/assets/img/amdgpu-performance.png">
-    <img src="/assets/img/amdgpu-performance.png" class=card-img-top alt>
-  </a>
-  <div class=card-body>
-    <p class=card-text>
-      Preliminary performance of a memcopy and 2D diffusion kernel implemented in
-      Julia with AMDGPU.jl and executed on a MI250x GPU.
-    </p>
-  </div>
-</div>
+At this point, the toolchain is a work in progress, although it is quite
+functional for simple usecases. We only officially support Julia 1.4 and Julia
+1.5.