In today's increasingly parallel and heterogeneous compute landscape, emerging platforms incorporate a variety of processor and accelerator types to improve performance and cost efficiency compared to homogeneous CPU systems. Alongside this compute paradigm shift, there has been a rise in the popularity of cloud computing, which enables computation offloading to third-party data centres, reducing on-site costs. As a result, cloud platforms now support heterogeneous resources, including CPUs, GPUs, and FPGAs, to serve specialised workloads. However, adoption of these resources poses a variety of challenges. This thesis presents three major research contributions covering novel approaches to tackle a selection of these challenges, focusing on heterogeneous cloud resource management and diverse high-level design.

First, managed cloud Platform-as-a-Service (PaaS) and Function-as-a-Service (FaaS) systems are enhanced with mechanisms to support heterogeneity. Specifically, the extended ORIAN PaaS platform supports application deployment with heterogeneous elasticity and the SLATE FaaS framework enables runtime scheduling of serverless computations across diverse targets with seamless accelerator support. Abstracted management of heterogeneous resources improves accessibility to specialised hardware, enabling a wider pool of developers to tap into their performance, power, and cost-saving potential.

Second, a meta-programming approach for programmatic high-level design optimisation is introduced. Meta-programs, with programmatic access to source code, tools, and platforms, are used to develop mapping and optimisation strategies for specialised, heterogeneous targets separately from application descriptions. Meta-program strategies can be reused for multiple applications and hardware targets, significantly reducing the manual optimisation effort and expertise currently required.

Finally, a novel framework for automating diverse design-flows is presented, supporting generation of multiple designs from a single high-level source. Modular optimisation tasks codified using meta-programs are organised into branching design-flows with support for path selection automation. Automated design-flows targeting diverse hardware platforms improve developer productivity and design portability, facilitating experimentation across application domains, hardware targets, and optimisation strategies.