Abstract: Techniques are disclosed relating to dedicated power function circuitry for a floating-point power instruction. In some embodiments, execution circuitry is configured to execute a floating-point power instruction to evaluate the power function xy as 2y log2x. In some embodiments, base-2 logarithm circuitry is configured to evaluate a base-2 logarithm for a first input (e.g., log2 x) by determining coefficients for a polynomial function and evaluating the polynomial function using the determined coefficients and the first input. In some embodiments, multiplication circuitry multiplies the base-2 logarithm result by a second input to generate a multiplication result. In some embodiments, base-2 power function circuitry is configured to evaluate a base-2 power function for the multiplication result. Disclosed techniques may advantageously increase performance and reduce power consumption of floating-point power function operations with reasonable area and accuracy, relative to traditional techniques.

Abstract: A method and an apparatus that schedule a plurality of executables in a schedule queue for execution in one or more physical compute devices such as CPUs or GPUs concurrently are described. One or more executables are compiled online from a source having an existing executable for a type of physical compute devices different from the one or more physical compute devices. Dependency relations among elements corresponding to scheduled executables are determined to select an executable to be executed by a plurality of threads concurrently in more than one of the physical compute devices. A thread initialized for executing an executable in a GPU of the physical compute devices are initialized for execution in another CPU of the physical compute devices if the GPU is busy with graphics processing threads.

Abstract: A method and an apparatus that schedule a plurality of executables in a schedule queue for execution in one or more physical compute devices such as CPUs or GPUs concurrently are described. One or more executables are compiled online from a source having an existing executable for a type of physical compute devices different from the one or more physical compute devices. Dependency relations among elements corresponding to scheduled executables are determined to select an executable to be executed by a plurality of threads concurrently in more than one of the physical compute devices. A thread initialized for executing an executable in a GPU of the physical compute devices are initialized for execution in another CPU of the physical compute devices if the GPU is busy with graphics processing threads.

Abstract: A method and an apparatus that schedule a plurality of executables in a schedule queue for execution in one or more physical compute devices such as CPUs or GPUs concurrently are described. One or more executables are compiled online from a source having an existing executable for a type of physical compute devices different from the one or more physical compute devices. Dependency relations among elements corresponding to scheduled executables are determined to select an executable to be executed by a plurality of threads concurrently in more than one of the physical compute devices. A thread initialized for executing an executable in a GPU of the physical compute devices are initialized for execution in another CPU of the physical compute devices if the GPU is busy with graphics processing threads.

Abstract: A method and an apparatus that allocate one or more physical compute devices such as Central Processing Units (CPUs) or Graphical Processing Units (GPUs) attached to a host processing unit running an application for executing one or more threads of the application are described. The allocation may be based on data representing a processing capability requirement from the application for executing an executable in the one or more threads. A compute device identifier may be associated with the allocated physical compute devices to schedule and execute the executable in the one or more threads concurrently in one or more of the allocated physical compute devices concurrently.

Abstract: The subject technology provides a framework for multi-processor training of neural networks. Multi-processor training of neural networks can include performing a forward pass of a training iteration using a neural processor, and performing a backward pass of the training iteration using a CPU or a GPU. Additional operations for facilitating the multi-processor training are disclosed.

Abstract: A method and an apparatus that execute a parallel computing program in a programming language for a parallel computing architecture are described. The parallel computing program is stored in memory in a system with parallel processors. The parallel computing program is stored in a memory to allocate threads between a host processor and a GPU. The programming language includes an API to allow an application to make calls using the API to allocate execution of the threads between the host processor and the GPU. The programming language includes host function data tokens for host functions performed in the host processor and kernel function data tokens for compute kernel functions performed in one or more compute processors, e.g., GPUs or CPUs, separate from the host processor.

Abstract: Graphics processing units (GPUs) and other compute units are allowed to enqueue tasks for themselves by themselves, without needing a host processor to queue the work for the GPU. Built-in functions enable kernels to enqueue kernels for execution on a device. In some embodiments, ndrange kernels execute over an N-dimensional range to provide data-parallel operations. Task kernels provide task-parallel operations. In some embodiments, kernels may be defined using clang block syntax. The order of execution of commands on a compute unit may be constrained or allow execution of commands out-of-order. Compute units may control when kernels enqueued by the compute unit begins execution.

Abstract: A method and an apparatus that execute a parallel computing program in a programming language for a parallel computing architecture are described. The parallel computing program is stored in memory in a system with parallel processors. The parallel computing program is stored in a memory to allocate threads between a host processor and a GPU. The programming language includes an API to allow an application to make calls using the API to allocate execution of the threads between the host processor and the GPU. The programming language includes host function data tokens for host functions performed in the host processor and kernel function data tokens for compute kernel functions performed in one or more compute processors, e.g., GPUs or CPUs, separate from the host processor.

Abstract: A compiler and library provide the ability to compile a programming language according to a defined language model into a programming language independent, machine independent intermediate representation, for conversion into an executable on a target programmable device. The language model allows writing programs that perform data-parallel graphics and non-graphics tasks.

Abstract: A method and an apparatus that allocate one or more physical compute devices such as CPUs or GPUs attached to a host processing unit running an application for executing one or more threads of the application are described. The allocation may be based on data representing a processing capability requirement from the application for executing an executable in the one or more threads. A compute device identifier may be associated with the allocated physical compute devices to schedule and execute the executable in the one or more threads concurrently in one or more of the allocated physical compute devices concurrently.

Abstract: Graphics processing units (GPUs) and other compute units are allowed to enqueue tasks for themselves by themselves, without needing a host processor to queue the work for the GPU. Built-in functions enable kernels to enqueue kernels for execution on a device. In some embodiments, ndrange kernels execute over an N-dimensional range to provide data-parallel operations. Task kernels provide task-parallel operations. In some embodiments, kernels may be defined using clang block syntax. The order of execution of commands on a compute unit may be constrained or allow execution of commands out-of-order. Compute units may control when kernels enqueued by the compute unit begins execution.

Abstract: A method and an apparatus that allocate one or more physical compute devices such as CPUs (Central Processing Unit) or GPUs (Graphics Processing Unit) attached to a host processing unit running an application for executing one or more threads of the application are described. The allocation may be based on data representing a processing capability requirement from the application for executing an executable in the one or more threads. A compute device identifier may be associated with the allocated physical compute devices to schedule and execute the executable in the one or more threads concurrently in one or more of the allocated physical compute devices concurrently.

Abstract: A method and an apparatus that execute a parallel computing program in a programming language for a parallel computing architecture are described. The parallel computing program is stored in memory in a system with parallel processors. The parallel computing program is stored in a memory to allocate threads between a host processor and a GPU. The programming language includes an API to allow an application to make calls using the API to allocate execution of the threads between the host processor and the GPU. The programming language includes host function data tokens for host functions performed in the host processor and kernel function data tokens for compute kernel functions performed in one or more compute processors, e.g., GPUs or CPUs, separate from the host processor.

Abstract: A compiler and library provide the ability to compile a programming language according to a defined language model into a programming language independent, machine independent intermediate representation, for conversion into an executable on a target programmable device. The language model allows writing programs that perform data-parallel graphics and non-graphics tasks.

Abstract: Graphics processing units (GPUs) and other compute units are allowed to enqueue tasks for themselves by themselves, without needing a host processor to queue the work for the GPU. Built-in functions enable kernels to enqueue kernels for execution on a device. In some embodiments, ndrange kernels execute over an N-dimensional range to provide data-parallel operations. Task kernels provide task-parallel operations. In some embodiments, kernels may be defined using clang block syntax. The order of execution of commands on a compute unit may be constrained or allow execution of commands out-of-order. Compute units may control when kernels enqueued by the compute unit begins execution.

Abstract: Techniques are disclosed relating to a hardware-supported flexible data structure for graphics processing. In some embodiments, dimensions of the data structure are configurable in an X direction, a Y direction, a number of samples per pixel, and an amount of data per sample. In some embodiments, these attributes are configurable using hardware registers. In some embodiments, the data structure is persistent across a tile being processed such that local memory context is accessible to both rendering threads of a render pass and mid-render compute threads.

Abstract: A compiler and library provide the ability to compile a programming language according to a defined language model into a programming language independent, machine independent intermediate representation, for conversion into an executable on a target programmable device. The language model allows writing programs that perform data-parallel graphics and non-graphics tasks.

Abstract: A system decouples the source code language from the eventual execution environment by compiling the source code language into a unified intermediate representation that conforms to a language model allowing both parallel graphical operations and parallel general-purpose computational operations. The intermediate representation may then be distributed to end-user computers, where an embedded compiler can compile the intermediate representation into an executable binary targeted for the CPUs and GPUs available in that end-user device. The intermediate representation is sufficient to define both graphics and non-graphics compute kernels and shaders. At install-time or later, the intermediate representation file may be compiled for the specific target hardware of the given end-user computing system.

Abstract: Techniques are disclosed relating to memory consistency in a memory hierarchy with relaxed ordering. In some embodiments, an apparatus includes a first level cache that is shared by a plurality of shader processing elements and a second level cache that is shared by the shader processing elements and at least a texture processing unit. In some embodiments, the apparatus is configured to execute operations specified by graphics instructions that include (1) an attribute of the operation that specifies a type of memory consistency to be imposed for the operation and (2) scope information for the attribute that specifies whether the memory consistency specified by the attribute should be enforced at the first level cache or the second level cache. In some embodiments, the apparatus is configured to determine whether to sequence memory accesses at the first level cache and the second level cache based on the attribute and the scope.