Kernel source adaptation for execution on a graphics processing unit
Examples described herein relate to computer-readable medium comprising instructions stored thereon, that if executed by one or more processors, cause the one or more processors to: access a kernel source written in a shading language; select a compiler from a plurality of compilers based on the shading language; generate a kernel instruction set architecture and an associated runtime binding based on the kernel source and the selected compiler; and adapt state information to configure a graphics processing unit (GPU) based on the runtime binding to target any API runtime of choice.
1 . At least one non-transitory computer-readable medium comprising instructions stored thereon, that if executed by one or more processors, cause the one or more processors to:
execute a user mode driver that is to:
access a kernel source written in a shading language;
select a compiler from a plurality of compilers based on the shading language;
generate a kernel instruction set architecture and an associated runtime binding based on the kernel source and the selected compiler; and
adapt state information to configure a graphics processing unit (GPU) based on the runtime binding, wherein the runtime binding indicates: registers for passing values for arguments to particular GPU assembly language level instructions, a number of threads to dispatch, and state programming to execute the GPU assembly language level instructions.
2 . The computer-readable medium of claim 1 , wherein the shading language comprises one or more of: C for Metal (CM), High-level shader language (HLSL), OpenCL, Standard Portable Intermediate Representation V (SPIR-V), Metal Shading Language (Metal SL), or CUDA.
3 . The computer-readable medium of claim 1 , wherein the plurality of compilers comprise one or more of: DirectX Bytecode (DXBC) compiler, where DXBC can be an intermediate representation of DirectX-based High-Level Shading Language (HLSL); vector compiler; OpenCL compiler, SpirV compiler, Metal Shading Language (Metal SL) compiler, or CUDA compiler.
4 . The computer-readable medium of claim 1 , wherein the compiler is selected based on the shading language.
5 . The computer-readable medium of claim 1 , wherein the kernel instruction set architecture generated by the compiler is unchanged by adaptation of state information to configure the GPU.
6 . The computer-readable medium of claim 1 , wherein the state information comprises one or more of: resource binding mode, resource type information, single instruction multiple data (SIMD) mode, number of threads to dispatch and how to break a thread size, and/or variables.
7 . The computer-readable medium of claim 1 , wherein the generate a kernel instruction set architecture and an associated runtime binding is based on a just in time (JIT) operation using the selected compiler.
8 . The computer-readable medium of claim 1 , wherein the adapt state information to configure a GPU based on the runtime binding occurs using a user mode driver (UMD) at workload execution.
9 . The computer-readable medium of claim 1 , comprising instructions stored thereon, that if executed by one or more processors, cause the one or more processors to:
deploy the generated kernel instruction set architecture for execution on the GPU configured using the adapted state information.
10 . An apparatus comprising:
a processor and
a machine-readable storage storing instructions, the instructions executable by the processor to cause the processor to:
access a kernel source written in a shading language;
select a compiler from a plurality of compilers based on the shading language;
generate a kernel instruction set architecture and an associated runtime binding based on the kernel source and the selected compiler; and
adapt state information to configure a graphics processing unit (GPU) based on the runtime binding, wherein the runtime binding indicates: registers for passing values for arguments to particular GPU assembly language level instructions, a number of threads to dispatch, and state programming to execute the GPU assembly language level instructions.
11 . The apparatus of claim 10 , wherein the shading language comprises one or more of: C for Metal (CM), High-level shader language (HLSL), OpenCL, Standard Portable Intermediate Representation V (SPIR-V), Metal Shading Language (Metal SL), or CUDA.
12 . The apparatus of claim 10 , wherein the plurality of compilers comprise one or more of: DirectX Bytecode (DXBC) compiler, where DXBC can be an intermediate representation of DirectX-based High-Level Shading Language (HLSL); vector compiler; OpenCL compiler, SpirV compiler, Metal Shading Language (Metal SL) compiler, or CUDA compiler.
13 . The apparatus of claim 10 , wherein the compiler is selected based on the shading language.
14 . The apparatus of claim 10 , wherein the kernel instruction set architecture generated by the compiler is unchanged by adaptation of state information to configure the GPU.
15 . The apparatus of claim 10 , wherein the state information comprises resource binding mode, resource type information, single instruction multiple data (SIMD) mode, number of threads to dispatch and how to break a thread size, and/or variables.
16 . The apparatus of claim 10 , wherein the generate a kernel instruction set architecture and an associated runtime binding is based on a just in time (JIT) operation using the selected compiler.
17 . The apparatus of claim 10 , wherein the processor is to execute a user mode driver (UMD) to adapt state information to configure a GPU based on the runtime binding.
18 . The apparatus of claim 10 , further comprising the GPU to perform the generated kernel instruction set architecture using the adapted state information.
19 . A method comprising:
accessing a kernel source written in a shading language;
selecting a compiler from a plurality of compilers based on the shading language;
generating a kernel instruction set architecture and an associated runtime binding based on the kernel source and the selected compiler; and
adapting state information to configure a graphics processing unit (GPU) based on the runtime binding, wherein the runtime binding indicates: registers for passing values for arguments to particular GPU assembly language level instructions, a number of threads to dispatch, and state programming to execute the GPU assembly language level instructions.
20 . The method of claim 19 , wherein
the shading language comprises one or more of: C for Metal (CM), High-level shader language (HLSL), OpenCL, Standard Portable Intermediate Representation V (SPIR-V), Metal Shading Language (Metal SL), or CUDA and
the plurality of compilers comprise one or more of: DirectX Bytecode (DXBC) compiler, where DXBC can be an intermediate representation of DirectX-based High-Level Shading Language (HLSL); vector compiler; OpenCL compiler, SpirV compiler, Metal Shading Language (Metal SL) compiler, or CUDA compiler.
21 . The method of claim 19 , wherein the compiler is selected based on the shading language.
22 . The method of claim 19 , wherein the kernel instruction set architecture generated by the compiler is unchanged by adaptation of state information to configure the GPU.
23 . The method of claim 19 , wherein the state information comprises: resource binding mode, resource type information, single instruction multiple data (SIMD) mode, number of threads to dispatch and how to break a thread size, and/or variables.
24 . The method of claim 19 , wherein the generating a kernel instruction set architecture and an associated runtime binding is based on a just in time (JIT) operation using the selected compiler.
25 . The method of claim 19 , comprising:
deploying the generated kernel instruction set architecture for execution on the GPU configured using the adapted state information.