This paper presents a programming language interface, a complete scripting language, to describe composable compiler transformations. These transformation programs can be written, shared and reused by non-expert application and library developers. From a compiler writer’s perspective, a scripting language interface permits rapid prototyping of compiler algorithms that can mix levels and compose different sequences of transformations, producing readable code as output. From a library or application developer’s perspective, the use of transformation programs permits expression of clean high-level code, and a separate description of how to map that code to architectural features, easing maintenance and porting to new architectures. We illustrate this interface in the context of CUDA-CHiLL, a source-to-source compiler transformation and code generation framework that transforms sequential loop nests to high-performance GPU code. We show how this high-level transformation and code generation language can be used to express: (1) complex transformation sequences, exemplified by a single loop restructuring construct used to generate a series of tiling and permute commands; and, (2) complex code generation sequences to produce CUDA code from a high-level specification. We demonstrate that the automatically-generated code either performs closely or outperforms two hand-tuned GPU library kernels from Nvidia’s CUBLAS 2.2 and 3.2 libraries.<\/p>\n","protected":false},"excerpt":{"rendered":"
This paper presents a programming language interface, a complete scripting language, to describe composable compiler transformations. These transformation programs can be written, shared and reused by non-expert application and library developers. From a compiler writer’s perspective, a scripting language interface permits rapid prototyping of compiler algorithms that can mix levels and compose different sequences of […]<\/p>\n","protected":false},"author":351,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":false,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[36,11,89,3],"tags":[1787,215,1782,238,14,242,20,234,378],"class_list":["post-6453","post","type-post","status-publish","format-standard","hentry","category-algorithms","category-computer-science","category-nvidia-cuda","category-paper","tag-algorithms","tag-code-generation","tag-computer-science","tag-cublas","tag-cuda","tag-mpi","tag-nvidia","tag-nvidia-geforce-gtx-280","tag-tesla-c2050"],"views":2209,"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/hgpu.org\/index.php?rest_route=\/wp\/v2\/posts\/6453","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/hgpu.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/hgpu.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/hgpu.org\/index.php?rest_route=\/wp\/v2\/users\/351"}],"replies":[{"embeddable":true,"href":"https:\/\/hgpu.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6453"}],"version-history":[{"count":0,"href":"https:\/\/hgpu.org\/index.php?rest_route=\/wp\/v2\/posts\/6453\/revisions"}],"wp:attachment":[{"href":"https:\/\/hgpu.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6453"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/hgpu.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6453"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/hgpu.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6453"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}