whispercpp 1.3.1 → 1.3.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/.gitignore +4 -3
- data/README.md +92 -31
- data/Rakefile +26 -7
- data/ext/.gitignore +5 -7
- data/ext/dependencies.rb +61 -0
- data/ext/extconf.rb +21 -198
- data/ext/options.rb +221 -0
- data/ext/ruby_whisper.c +159 -0
- data/ext/ruby_whisper.h +17 -2
- data/ext/ruby_whisper_context.c +641 -0
- data/ext/ruby_whisper_error.c +52 -0
- data/ext/ruby_whisper_model.c +232 -0
- data/ext/ruby_whisper_params.c +1301 -0
- data/ext/ruby_whisper_segment.c +143 -0
- data/ext/ruby_whisper_transcribe.cpp +87 -0
- data/ext/ruby_whisper_vad_params.c +288 -0
- data/ext/sources/.dockerignore +3 -0
- data/ext/sources/.github/workflows/bindings-ruby.yml +21 -0
- data/ext/sources/CMakeGraphVizOptions.cmake +8 -0
- data/ext/sources/CMakeLists.txt +251 -0
- data/ext/sources/bindings/javascript/CMakeLists.txt +41 -0
- data/ext/sources/bindings/javascript/emscripten.cpp +93 -0
- data/ext/sources/bindings/javascript/libwhisper.worker.js +1 -0
- data/ext/sources/bindings/javascript/package-tmpl.json +26 -0
- data/ext/sources/bindings/javascript/package.json +26 -0
- data/ext/sources/bindings/javascript/whisper.js +19 -0
- data/ext/sources/build-xcframework.sh +547 -0
- data/ext/sources/ci/run.sh +336 -0
- data/ext/sources/close-issue.yml +28 -0
- data/ext/sources/cmake/DefaultTargetOptions.cmake +16 -0
- data/ext/sources/cmake/FindFFmpeg.cmake +163 -0
- data/ext/sources/cmake/build-info.cmake +60 -0
- data/ext/sources/cmake/git-vars.cmake +22 -0
- data/ext/sources/cmake/whisper-config.cmake.in +65 -0
- data/ext/sources/cmake/whisper.pc.in +10 -0
- data/ext/sources/examples/CMakeLists.txt +124 -0
- data/ext/sources/examples/addon.node/CMakeLists.txt +31 -0
- data/ext/sources/examples/addon.node/__test__/whisper.spec.js +37 -0
- data/ext/sources/examples/addon.node/addon.cpp +438 -0
- data/ext/sources/examples/addon.node/index.js +54 -0
- data/ext/sources/examples/addon.node/package.json +16 -0
- data/ext/sources/examples/bench/CMakeLists.txt +8 -0
- data/ext/sources/examples/bench/bench.cpp +175 -0
- data/ext/sources/examples/bench.wasm/CMakeLists.txt +49 -0
- data/ext/sources/examples/bench.wasm/emscripten.cpp +87 -0
- data/ext/sources/examples/bench.wasm/index-tmpl.html +284 -0
- data/ext/sources/examples/cli/CMakeLists.txt +8 -0
- data/ext/sources/examples/cli/cli.cpp +1294 -0
- data/ext/sources/examples/coi-serviceworker.js +146 -0
- data/ext/sources/examples/command/CMakeLists.txt +10 -0
- data/ext/sources/examples/command/command.cpp +776 -0
- data/ext/sources/examples/command/commands.txt +9 -0
- data/ext/sources/examples/command.wasm/CMakeLists.txt +50 -0
- data/ext/sources/examples/command.wasm/emscripten.cpp +327 -0
- data/ext/sources/examples/command.wasm/index-tmpl.html +414 -0
- data/ext/sources/examples/common-ggml.cpp +238 -0
- data/ext/sources/examples/common-ggml.h +18 -0
- data/ext/sources/examples/common-sdl.cpp +227 -0
- data/ext/sources/examples/common-sdl.h +49 -0
- data/ext/sources/examples/common-whisper.cpp +168 -0
- data/ext/sources/examples/common-whisper.h +24 -0
- data/ext/sources/examples/common.cpp +675 -0
- data/ext/sources/examples/common.h +322 -0
- data/ext/sources/examples/deprecation-warning/CMakeLists.txt +6 -0
- data/ext/sources/examples/deprecation-warning/deprecation-warning.cpp +38 -0
- data/ext/sources/examples/ffmpeg-transcode.cpp +368 -0
- data/ext/sources/examples/generate-karaoke.sh +57 -0
- data/ext/sources/examples/grammar-parser.cpp +423 -0
- data/ext/sources/examples/grammar-parser.h +29 -0
- data/ext/sources/examples/helpers.js +191 -0
- data/ext/sources/examples/json.hpp +24596 -0
- data/ext/sources/examples/livestream.sh +112 -0
- data/ext/sources/examples/lsp/CMakeLists.txt +9 -0
- data/ext/sources/examples/lsp/lsp.cpp +467 -0
- data/ext/sources/examples/lsp/whisper.vim +362 -0
- data/ext/sources/examples/miniaudio.h +93468 -0
- data/ext/sources/examples/python/test_whisper_processor.py +7 -0
- data/ext/sources/examples/python/whisper_processor.py +54 -0
- data/ext/sources/examples/quantize/CMakeLists.txt +6 -0
- data/ext/sources/examples/quantize/quantize.cpp +223 -0
- data/ext/sources/examples/server/CMakeLists.txt +12 -0
- data/ext/sources/examples/server/bench.js +29 -0
- data/ext/sources/examples/server/httplib.h +10497 -0
- data/ext/sources/examples/server/server.cpp +1091 -0
- data/ext/sources/examples/server.py +115 -0
- data/ext/sources/examples/stb_vorbis.c +5584 -0
- data/ext/sources/examples/stream/CMakeLists.txt +10 -0
- data/ext/sources/examples/stream/stream.cpp +429 -0
- data/ext/sources/examples/stream.wasm/CMakeLists.txt +49 -0
- data/ext/sources/examples/stream.wasm/emscripten.cpp +216 -0
- data/ext/sources/examples/stream.wasm/index-tmpl.html +414 -0
- data/ext/sources/examples/sycl/CMakeLists.txt +9 -0
- data/ext/sources/examples/sycl/build.sh +22 -0
- data/ext/sources/examples/sycl/ls-sycl-device.cpp +11 -0
- data/ext/sources/examples/sycl/run-whisper.sh +17 -0
- data/ext/sources/examples/talk-llama/CMakeLists.txt +40 -0
- data/ext/sources/examples/talk-llama/eleven-labs.py +80 -0
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +388 -0
- data/ext/sources/examples/talk-llama/llama-adapter.h +76 -0
- data/ext/sources/examples/talk-llama/llama-arch.cpp +1746 -0
- data/ext/sources/examples/talk-llama/llama-arch.h +437 -0
- data/ext/sources/examples/talk-llama/llama-batch.cpp +374 -0
- data/ext/sources/examples/talk-llama/llama-batch.h +89 -0
- data/ext/sources/examples/talk-llama/llama-chat.cpp +663 -0
- data/ext/sources/examples/talk-llama/llama-chat.h +58 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +2676 -0
- data/ext/sources/examples/talk-llama/llama-context.h +276 -0
- data/ext/sources/examples/talk-llama/llama-cparams.cpp +5 -0
- data/ext/sources/examples/talk-llama/llama-cparams.h +41 -0
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +1229 -0
- data/ext/sources/examples/talk-llama/llama-grammar.h +173 -0
- data/ext/sources/examples/talk-llama/llama-graph.cpp +1618 -0
- data/ext/sources/examples/talk-llama/llama-graph.h +640 -0
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +95 -0
- data/ext/sources/examples/talk-llama/llama-hparams.h +190 -0
- data/ext/sources/examples/talk-llama/llama-impl.cpp +167 -0
- data/ext/sources/examples/talk-llama/llama-impl.h +61 -0
- data/ext/sources/examples/talk-llama/llama-io.cpp +15 -0
- data/ext/sources/examples/talk-llama/llama-io.h +35 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2739 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +502 -0
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +379 -0
- data/ext/sources/examples/talk-llama/llama-memory.cpp +1 -0
- data/ext/sources/examples/talk-llama/llama-memory.h +32 -0
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +600 -0
- data/ext/sources/examples/talk-llama/llama-mmap.h +68 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +1138 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.h +169 -0
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +281 -0
- data/ext/sources/examples/talk-llama/llama-model-saver.h +37 -0
- data/ext/sources/examples/talk-llama/llama-model.cpp +13814 -0
- data/ext/sources/examples/talk-llama/llama-model.h +425 -0
- data/ext/sources/examples/talk-llama/llama-quant.cpp +966 -0
- data/ext/sources/examples/talk-llama/llama-quant.h +1 -0
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +2575 -0
- data/ext/sources/examples/talk-llama/llama-sampling.h +32 -0
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +3340 -0
- data/ext/sources/examples/talk-llama/llama-vocab.h +131 -0
- data/ext/sources/examples/talk-llama/llama.cpp +354 -0
- data/ext/sources/examples/talk-llama/llama.h +1377 -0
- data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +23 -0
- data/ext/sources/examples/talk-llama/speak +40 -0
- data/ext/sources/examples/talk-llama/speak.bat +1 -0
- data/ext/sources/examples/talk-llama/speak.ps1 +14 -0
- data/ext/sources/examples/talk-llama/talk-llama.cpp +808 -0
- data/ext/sources/examples/talk-llama/unicode-data.cpp +7034 -0
- data/ext/sources/examples/talk-llama/unicode-data.h +20 -0
- data/ext/sources/examples/talk-llama/unicode.cpp +849 -0
- data/ext/sources/examples/talk-llama/unicode.h +66 -0
- data/ext/sources/examples/vad-speech-segments/CMakeLists.txt +8 -0
- data/ext/sources/examples/vad-speech-segments/speech.cpp +143 -0
- data/ext/sources/examples/wchess/CMakeLists.txt +10 -0
- data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +19 -0
- data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +803 -0
- data/ext/sources/examples/wchess/libwchess/Chessboard.h +33 -0
- data/ext/sources/examples/wchess/libwchess/WChess.cpp +193 -0
- data/ext/sources/examples/wchess/libwchess/WChess.h +63 -0
- data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +117 -0
- data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +8 -0
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +249 -0
- data/ext/sources/examples/whisper.wasm/CMakeLists.txt +50 -0
- data/ext/sources/examples/whisper.wasm/emscripten.cpp +118 -0
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +658 -0
- data/ext/sources/ggml/CMakeLists.txt +390 -0
- data/ext/sources/ggml/cmake/BuildTypes.cmake +54 -0
- data/ext/sources/ggml/cmake/GitVars.cmake +22 -0
- data/ext/sources/ggml/cmake/common.cmake +26 -0
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +152 -0
- data/ext/{ggml → sources/ggml}/include/ggml-alloc.h +1 -1
- data/ext/{ggml → sources/ggml}/include/ggml-backend.h +9 -7
- data/ext/{ggml → sources/ggml}/include/ggml-cpp.h +2 -1
- data/ext/{ggml → sources/ggml}/include/ggml-cpu.h +9 -1
- data/ext/{ggml → sources/ggml}/include/ggml-metal.h +1 -1
- data/ext/{ggml → sources/ggml}/include/ggml-opt.h +49 -28
- data/ext/{ggml → sources/ggml}/include/ggml-rpc.h +6 -1
- data/ext/{ggml → sources/ggml}/include/ggml-vulkan.h +0 -2
- data/ext/{ggml → sources/ggml}/include/ggml.h +182 -265
- data/ext/sources/ggml/include/gguf.h +202 -0
- data/ext/sources/ggml/src/CMakeLists.txt +346 -0
- data/ext/{ggml → sources/ggml}/src/ggml-alloc.c +34 -29
- data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +107 -0
- data/ext/{ggml → sources/ggml}/src/ggml-backend-impl.h +1 -2
- data/ext/{ggml → sources/ggml}/src/ggml-backend-reg.cpp +87 -53
- data/ext/{ggml → sources/ggml}/src/ggml-backend.cpp +26 -14
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +87 -0
- data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +74 -0
- data/ext/sources/ggml/src/ggml-cann/Doxyfile +2579 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.cpp +10 -4
- data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.h +5 -5
- data/ext/{ggml → sources/ggml}/src/ggml-cann/aclnn_ops.cpp +1272 -1506
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +1125 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/common.h +135 -1
- data/ext/{ggml → sources/ggml}/src/ggml-cann/ggml-cann.cpp +564 -146
- data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +30 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/dup.cpp +3 -5
- data/ext/{ggml → sources/ggml}/src/ggml-common.h +12 -8
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +504 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.cpp +2 -1
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +158 -0
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.h +16 -0
- data/ext/sources/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
- data/ext/sources/ggml/src/ggml-cpu/common.h +72 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/cpu-feats-x86.cpp +5 -1
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +6431 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-impl.h +163 -41
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-quants.c +4029 -1117
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +3510 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu.cpp +67 -18
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +337 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +95 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +482 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +3544 -0
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +8903 -0
- data/ext/sources/ggml/src/ggml-cpu/ops.h +110 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +892 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +186 -0
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.h +28 -0
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +252 -0
- data/ext/sources/ggml/src/ggml-cpu/vec.h +818 -0
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +184 -0
- data/ext/sources/ggml/src/ggml-cuda/acc.cu +61 -0
- data/ext/sources/ggml/src/ggml-cuda/acc.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/arange.cu +34 -0
- data/ext/sources/ggml/src/ggml-cuda/arange.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/argmax.cu +91 -0
- data/ext/sources/ggml/src/ggml-cuda/argmax.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +104 -0
- data/ext/sources/ggml/src/ggml-cuda/argsort.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +363 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +9 -0
- data/ext/sources/ggml/src/ggml-cuda/clamp.cu +45 -0
- data/ext/sources/ggml/src/ggml-cuda/clamp.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +828 -0
- data/ext/sources/ggml/src/ggml-cuda/concat.cu +221 -0
- data/ext/sources/ggml/src/ggml-cuda/concat.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +89 -0
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +730 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +26 -0
- data/ext/sources/ggml/src/ggml-cuda/count-equal.cu +64 -0
- data/ext/sources/ggml/src/ggml-cuda/count-equal.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/cp-async.cuh +57 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +705 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cuh +11 -0
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +189 -0
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +103 -0
- data/ext/sources/ggml/src/ggml-cuda/diagmask.cu +40 -0
- data/ext/sources/ggml/src/ggml-cuda/diagmask.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +881 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +1471 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +357 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +365 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +482 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +472 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +634 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +346 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +275 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cuh +15 -0
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +3505 -0
- data/ext/sources/ggml/src/ggml-cuda/gla.cu +93 -0
- data/ext/sources/ggml/src/ggml-cuda/gla.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +103 -0
- data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +396 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +324 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +3217 -0
- data/ext/sources/ggml/src/ggml-cuda/mmv.cu +336 -0
- data/ext/sources/ggml/src/ggml-cuda/mmv.cuh +12 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +595 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +12 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +458 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cuh +11 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +68 -0
- data/ext/sources/ggml/src/ggml-cuda/out-prod.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +49 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/pool2d.cu +94 -0
- data/ext/sources/ggml/src/ggml-cuda/pool2d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +190 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +27 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +456 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +31 -0
- data/ext/sources/ggml/src/ggml-cuda/scale.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +283 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +148 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +153 -0
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/sum.cu +45 -0
- data/ext/sources/ggml/src/ggml-cuda/sum.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +39 -0
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +10 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +78 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +47 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +289 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +59 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cu +51 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +1135 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/cuda.h +1 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/hip.h +57 -0
- data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/musa.h +7 -1
- data/ext/sources/ggml/src/ggml-cuda/wkv.cu +199 -0
- data/ext/sources/ggml/src/ggml-cuda/wkv.cuh +7 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +131 -0
- data/ext/{ggml → sources/ggml}/src/ggml-impl.h +64 -19
- data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +166 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +112 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +58 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +25 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +30 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +22 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +17 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +31 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +31 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +38 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +39 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +44 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +69 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +51 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +33 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +35 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +140 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +106 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +73 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +28 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +84 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +21 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +53 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +52 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +19 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +23 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +22 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +72 -0
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +71 -0
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +120 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +622 -0
- data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.m +2178 -1064
- data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.metal +1575 -1218
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +113 -0
- data/ext/sources/ggml/src/ggml-musa/mudnn.cu +112 -0
- data/ext/sources/ggml/src/ggml-musa/mudnn.cuh +12 -0
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +96 -0
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +5124 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +83 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +62 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl +268 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl +274 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +79 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl +190 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +81 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +96 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rope.cl +721 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +16 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +87 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +87 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +86 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +86 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +84 -0
- data/ext/{ggml → sources/ggml}/src/ggml-opt.cpp +373 -190
- data/ext/{ggml → sources/ggml}/src/ggml-quants.c +114 -120
- data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
- data/ext/{ggml → sources/ggml}/src/ggml-rpc/ggml-rpc.cpp +480 -73
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +189 -0
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +37 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +345 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +39 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/common.cpp +20 -32
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +589 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/concat.cpp +32 -33
- data/ext/sources/ggml/src/ggml-sycl/concat.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/conv.cpp +4 -2
- data/ext/sources/ggml/src/ggml-sycl/conv.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/convert.cpp +104 -28
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +34 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +700 -0
- data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +11 -0
- data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +791 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/dmmv.cpp +156 -17
- data/ext/sources/ggml/src/ggml-sycl/dmmv.hpp +27 -0
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +2957 -0
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1511 -0
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +75 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +99 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +309 -0
- data/ext/sources/ggml/src/ggml-sycl/getrows.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/ggml-sycl.cpp +1004 -1240
- data/ext/sources/ggml/src/ggml-sycl/gla.cpp +106 -0
- data/ext/sources/ggml/src/ggml-sycl/gla.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +136 -0
- data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +21 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/mmq.cpp +0 -1
- data/ext/sources/ggml/src/ggml-sycl/mmq.hpp +33 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/mmvq.cpp +261 -166
- data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +27 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/norm.cpp +204 -81
- data/ext/sources/ggml/src/ggml-sycl/norm.hpp +26 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/outprod.cpp +8 -17
- data/ext/sources/ggml/src/ggml-sycl/outprod.hpp +10 -0
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +74 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +83 -0
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +361 -0
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +20 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/softmax.cpp +35 -25
- data/ext/sources/ggml/src/ggml-sycl/softmax.hpp +20 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
- data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
- data/ext/{ggml → sources/ggml}/src/ggml-sycl/tsembd.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/tsembd.hpp +20 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +1215 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +293 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.hpp +10 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +196 -0
- data/ext/sources/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
- data/ext/{ggml → sources/ggml}/src/ggml-vulkan/ggml-vulkan.cpp +3130 -1087
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +51 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +69 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +23 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +242 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +462 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +699 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp +13 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +48 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +49 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +31 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +337 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +360 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +267 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +59 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +25 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +23 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +64 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp +9 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp +76 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +33 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +100 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +169 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +118 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +82 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +79 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +90 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +118 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +154 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +130 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +136 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +167 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +868 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +441 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +442 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +99 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +28 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +77 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +52 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +58 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +60 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +43 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +47 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +173 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +50 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +37 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +20 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp +7 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +41 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +1373 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +36 -0
- data/ext/{ggml → sources/ggml}/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -35
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
- data/ext/{ggml → sources/ggml}/src/ggml.c +676 -1820
- data/ext/sources/ggml/src/gguf.cpp +1330 -0
- data/ext/{include → sources/include}/whisper.h +68 -2
- data/ext/sources/src/CMakeLists.txt +143 -0
- data/ext/{src → sources/src}/coreml/whisper-decoder-impl.h +27 -15
- data/ext/{src → sources/src}/coreml/whisper-decoder-impl.m +35 -10
- data/ext/{src → sources/src}/coreml/whisper-encoder-impl.h +21 -9
- data/ext/{src → sources/src}/coreml/whisper-encoder-impl.m +28 -3
- data/ext/sources/src/coreml/whisper-encoder.mm +73 -0
- data/ext/sources/src/whisper-arch.h +197 -0
- data/ext/{src → sources/src}/whisper.cpp +1905 -374
- data/ext/sources/tests/CMakeLists.txt +105 -0
- data/ext/sources/tests/earnings21/eval.mk +58 -0
- data/ext/sources/tests/earnings21/eval.py +68 -0
- data/ext/sources/tests/earnings21/normalizers/__init__.py +2 -0
- data/ext/sources/tests/earnings21/normalizers/basic.py +80 -0
- data/ext/sources/tests/earnings21/normalizers/english.json +1741 -0
- data/ext/sources/tests/earnings21/normalizers/english.py +550 -0
- data/ext/sources/tests/earnings21/requirements.txt +6 -0
- data/ext/sources/tests/en-0-ref.txt +1 -0
- data/ext/sources/tests/en-1-ref.txt +1 -0
- data/ext/sources/tests/en-2-ref.txt +1 -0
- data/ext/sources/tests/es-0-ref.txt +1 -0
- data/ext/sources/tests/librispeech/eval.mk +39 -0
- data/ext/sources/tests/librispeech/eval.py +47 -0
- data/ext/sources/tests/librispeech/normalizers/__init__.py +2 -0
- data/ext/sources/tests/librispeech/normalizers/basic.py +80 -0
- data/ext/sources/tests/librispeech/normalizers/english.json +1741 -0
- data/ext/sources/tests/librispeech/normalizers/english.py +550 -0
- data/ext/sources/tests/librispeech/requirements.txt +6 -0
- data/ext/sources/tests/run-tests.sh +130 -0
- data/ext/sources/tests/test-c.c +3 -0
- data/ext/sources/tests/test-vad-full.cpp +54 -0
- data/ext/sources/tests/test-vad.cpp +83 -0
- data/ext/sources/tests/test-whisper.js +58 -0
- data/extsources.rb +33 -5
- data/lib/whisper/model/uri.rb +149 -128
- data/sig/whisper.rbs +480 -0
- data/tests/helper.rb +28 -0
- data/tests/test_callback.rb +45 -3
- data/tests/test_error.rb +2 -2
- data/tests/test_model.rb +38 -0
- data/tests/test_package.rb +18 -3
- data/tests/test_params.rb +145 -8
- data/tests/test_segment.rb +10 -19
- data/tests/test_vad.rb +19 -0
- data/tests/test_vad_params.rb +103 -0
- data/tests/test_whisper.rb +37 -37
- data/whispercpp.gemspec +5 -4
- metadata +766 -111
- data/ext/cpu.mk +0 -9
- data/ext/examples/dr_wav.h +0 -8815
- data/ext/ggml/src/ggml-cann/aclnn_ops.h +0 -592
- data/ext/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +0 -4262
- data/ext/ggml/src/ggml-cpu/ggml-cpu.c +0 -14123
- data/ext/ggml/src/ggml-cpu/llamafile/sgemm.cpp +0 -1884
- data/ext/ggml/src/ggml-cpu/llamafile/sgemm.h +0 -14
- data/ext/ggml/src/ggml-metal/ggml-metal-impl.h +0 -288
- data/ext/ggml/src/ggml-sycl/element_wise.cpp +0 -1030
- data/ext/ggml/src/ggml-sycl/im2col.cpp +0 -126
- data/ext/ggml/src/ggml-sycl/rope.cpp +0 -276
- data/ext/ggml/src/ggml-sycl/wkv6.cpp +0 -141
- data/ext/metal-embed.mk +0 -17
- data/ext/metal.mk +0 -6
- data/ext/ruby_whisper.cpp +0 -1909
- data/ext/scripts/get-flags.mk +0 -38
- data/lib/whisper.rb +0 -2
- /data/ext/{ggml → sources/ggml}/include/ggml-blas.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-cann.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-cuda.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-kompute.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-opencl.h +0 -0
- /data/ext/{ggml → sources/ggml}/include/ggml-sycl.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/common.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/ggml-amx.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-blas/ggml-blas.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/ascendc_kernels.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f16.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f32.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/common.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-aarch64.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-hbm.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-hbm.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-quants.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-traits.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-traits.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-kompute/ggml-kompute.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-quants.h +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-threading.cpp +0 -0
- /data/ext/{ggml → sources/ggml}/src/ggml-threading.h +0 -0
- /data/ext/{src → sources/src}/coreml/whisper-encoder.h +0 -0
- /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.cpp +0 -0
- /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.h +0 -0
@@ -11,20 +11,26 @@
|
|
11
11
|
#include <vector>
|
12
12
|
|
13
13
|
#ifdef GGML_USE_CPU_HBM
|
14
|
-
#include "ggml-cpu-hbm.h"
|
14
|
+
# include "ggml-cpu-hbm.h"
|
15
15
|
#endif
|
16
16
|
|
17
|
-
#
|
18
|
-
#include
|
19
|
-
#include <sys/sysctl.h>
|
17
|
+
#ifdef GGML_USE_CPU_KLEIDIAI
|
18
|
+
# include "kleidiai/kleidiai.h"
|
20
19
|
#endif
|
21
20
|
|
22
21
|
#if defined(_WIN32)
|
23
|
-
#define WIN32_LEAN_AND_MEAN
|
24
|
-
#ifndef NOMINMAX
|
25
|
-
|
22
|
+
# define WIN32_LEAN_AND_MEAN
|
23
|
+
# ifndef NOMINMAX
|
24
|
+
# define NOMINMAX
|
25
|
+
# endif
|
26
|
+
# include <windows.h>
|
27
|
+
#else
|
28
|
+
# include <unistd.h>
|
26
29
|
#endif
|
27
|
-
|
30
|
+
|
31
|
+
#if defined(__APPLE__)
|
32
|
+
# include <sys/sysctl.h>
|
33
|
+
# include <sys/types.h>
|
28
34
|
#endif
|
29
35
|
|
30
36
|
// ggml-backend interface
|
@@ -39,6 +45,12 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
|
|
39
45
|
}
|
40
46
|
#endif
|
41
47
|
|
48
|
+
#ifdef GGML_USE_CPU_KLEIDIAI
|
49
|
+
if (ggml_backend_cpu_kleidiai_buffer_type()) {
|
50
|
+
bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
|
51
|
+
}
|
52
|
+
#endif
|
53
|
+
|
42
54
|
#ifdef GGML_USE_CPU_AARCH64
|
43
55
|
if (ggml_backend_cpu_aarch64_buffer_type()) {
|
44
56
|
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
|
@@ -60,8 +72,10 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_ty
|
|
60
72
|
}
|
61
73
|
|
62
74
|
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
|
63
|
-
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
|
64
|
-
if (extra && extra == buft)
|
75
|
+
for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
|
76
|
+
if (extra && extra == buft) {
|
77
|
+
return true;
|
78
|
+
}
|
65
79
|
}
|
66
80
|
return false;
|
67
81
|
}
|
@@ -284,14 +298,14 @@ struct ggml_backend_cpu_device_context {
|
|
284
298
|
&hKey) == ERROR_SUCCESS) {
|
285
299
|
DWORD cpu_brand_size = 0;
|
286
300
|
if (RegQueryValueExA(hKey,
|
287
|
-
|
301
|
+
"ProcessorNameString",
|
288
302
|
NULL,
|
289
303
|
NULL,
|
290
304
|
NULL,
|
291
305
|
&cpu_brand_size) == ERROR_SUCCESS) {
|
292
306
|
description.resize(cpu_brand_size);
|
293
307
|
if (RegQueryValueExA(hKey,
|
294
|
-
|
308
|
+
"ProcessorNameString",
|
295
309
|
NULL,
|
296
310
|
NULL,
|
297
311
|
(LPBYTE)&description[0], // NOLINT
|
@@ -320,9 +334,18 @@ static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t d
|
|
320
334
|
}
|
321
335
|
|
322
336
|
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
|
323
|
-
|
324
|
-
|
325
|
-
|
337
|
+
#ifdef _WIN32
|
338
|
+
MEMORYSTATUSEX status;
|
339
|
+
status.dwLength = sizeof(status);
|
340
|
+
GlobalMemoryStatusEx(&status);
|
341
|
+
*total = status.ullTotalPhys;
|
342
|
+
*free = status.ullAvailPhys;
|
343
|
+
#else
|
344
|
+
long pages = sysconf(_SC_PHYS_PAGES);
|
345
|
+
long page_size = sysconf(_SC_PAGE_SIZE);
|
346
|
+
*total = pages * page_size;
|
347
|
+
*free = *total;
|
348
|
+
#endif
|
326
349
|
|
327
350
|
GGML_UNUSED(dev);
|
328
351
|
}
|
@@ -403,12 +426,23 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
|
|
403
426
|
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
|
404
427
|
case GGML_OP_MUL_MAT:
|
405
428
|
return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
|
406
|
-
case
|
407
|
-
|
429
|
+
case GGML_OP_SOFT_MAX_BACK: {
|
430
|
+
if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
|
431
|
+
return false;
|
432
|
+
}
|
433
|
+
float max_bias = 0.0f;
|
434
|
+
|
435
|
+
memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
|
436
|
+
|
437
|
+
return max_bias == 0.0f;
|
438
|
+
}
|
408
439
|
case GGML_OP_IM2COL_BACK:
|
409
440
|
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
|
441
|
+
case GGML_OP_GET_ROWS_BACK:
|
442
|
+
return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
|
410
443
|
case GGML_OP_OUT_PROD:
|
411
|
-
return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)
|
444
|
+
return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
|
445
|
+
src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
|
412
446
|
default:
|
413
447
|
return true;
|
414
448
|
}
|
@@ -492,6 +526,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
|
|
492
526
|
if (ggml_cpu_has_fma()) {
|
493
527
|
features.push_back({ "FMA", "1" });
|
494
528
|
}
|
529
|
+
if (ggml_cpu_has_bmi2()) {
|
530
|
+
features.push_back({ "BMI2", "1" });
|
531
|
+
}
|
495
532
|
if (ggml_cpu_has_avx512()) {
|
496
533
|
features.push_back({ "AVX512", "1" });
|
497
534
|
}
|
@@ -522,16 +559,25 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
|
|
522
559
|
if (ggml_cpu_has_sve()) {
|
523
560
|
features.push_back({ "SVE", "1" });
|
524
561
|
}
|
562
|
+
if (ggml_cpu_has_dotprod()) {
|
563
|
+
features.push_back({ "DOTPROD", "1" });
|
564
|
+
}
|
525
565
|
if (ggml_cpu_get_sve_cnt() > 0) {
|
526
566
|
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
|
527
567
|
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
|
528
568
|
}
|
569
|
+
if (ggml_cpu_has_sme()) {
|
570
|
+
features.push_back({ "SME", "1" });
|
571
|
+
}
|
529
572
|
if (ggml_cpu_has_riscv_v()) {
|
530
573
|
features.push_back({ "RISCV_V", "1" });
|
531
574
|
}
|
532
575
|
if (ggml_cpu_has_vsx()) {
|
533
576
|
features.push_back({ "VSX", "1" });
|
534
577
|
}
|
578
|
+
if (ggml_cpu_has_vxe()) {
|
579
|
+
features.push_back({ "VXE", "1" });
|
580
|
+
}
|
535
581
|
if (ggml_cpu_has_wasm_simd()) {
|
536
582
|
features.push_back({ "WASM_SIMD", "1" });
|
537
583
|
}
|
@@ -547,6 +593,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
|
|
547
593
|
#ifdef GGML_USE_OPENMP
|
548
594
|
features.push_back({ "OPENMP", "1" });
|
549
595
|
#endif
|
596
|
+
#ifdef GGML_USE_CPU_KLEIDIAI
|
597
|
+
features.push_back({ "KLEIDIAI", "1" });
|
598
|
+
#endif
|
550
599
|
#ifdef GGML_USE_CPU_AARCH64
|
551
600
|
features.push_back({ "AARCH64_REPACK", "1" });
|
552
601
|
#endif
|
@@ -0,0 +1,337 @@
|
|
1
|
+
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
2
|
+
// SPDX-License-Identifier: MIT
|
3
|
+
//
|
4
|
+
|
5
|
+
// KleidiAI micro-kernels
|
6
|
+
#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
|
7
|
+
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
|
8
|
+
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
|
9
|
+
#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
|
10
|
+
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
|
11
|
+
#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
|
12
|
+
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
|
13
|
+
#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
|
14
|
+
|
15
|
+
#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
|
16
|
+
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
|
17
|
+
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
|
18
|
+
|
19
|
+
#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
|
20
|
+
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
|
21
|
+
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
|
22
|
+
|
23
|
+
#include "kai_common.h"
|
24
|
+
|
25
|
+
#include "kernels.h"
|
26
|
+
|
27
|
+
#define NELEMS(x) sizeof(x) / sizeof(*x)
|
28
|
+
static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
|
29
|
+
#if defined(__ARM_FEATURE_SME)
|
30
|
+
{
|
31
|
+
/* SME GEMM */
|
32
|
+
/* .kern_info = */ {
|
33
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
34
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
35
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
36
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
37
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
38
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
39
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
40
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
41
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
42
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
43
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
|
44
|
+
},
|
45
|
+
/* SME GEMV */
|
46
|
+
/* .kern_info = */ {
|
47
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
48
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
49
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
50
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
51
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
52
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
53
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
54
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
55
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
56
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
57
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
|
58
|
+
},
|
59
|
+
/* .lhs_info = */ {
|
60
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
|
61
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
|
62
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
|
63
|
+
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
|
64
|
+
},
|
65
|
+
/* .rhs_info = */ {
|
66
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
|
67
|
+
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
|
68
|
+
},
|
69
|
+
/* .required_cpu = */ CPU_FEATURE_SME,
|
70
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
71
|
+
/* .rhs_type = */ GGML_TYPE_Q4_0,
|
72
|
+
/* .op_type = */ GGML_TYPE_F32,
|
73
|
+
},
|
74
|
+
{
|
75
|
+
/* SME GEMM */
|
76
|
+
/* .kern_info = */ {
|
77
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
78
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
79
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
80
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
81
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
82
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
83
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
84
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
85
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
86
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
87
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
88
|
+
},
|
89
|
+
/* SME GEMV */
|
90
|
+
/* .kern_info = */ {
|
91
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
92
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
93
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
94
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
95
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
96
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
97
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
98
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
99
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
100
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
101
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
|
102
|
+
},
|
103
|
+
/* .lhs_info = */ {
|
104
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
|
105
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
|
106
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
|
107
|
+
/* .pack_func = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
|
108
|
+
},
|
109
|
+
/* .rhs_info = */ {
|
110
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
|
111
|
+
/* .pack_func = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
|
112
|
+
},
|
113
|
+
/* .required_cpu = */ CPU_FEATURE_SME,
|
114
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
115
|
+
/* .rhs_type = */ GGML_TYPE_F16,
|
116
|
+
/* .op_type = */ GGML_TYPE_F32,
|
117
|
+
},
|
118
|
+
#endif
|
119
|
+
#if defined(__APPLE__)
|
120
|
+
#if defined(__ARM_FEATURE_DOTPROD)
|
121
|
+
{
|
122
|
+
/* DOTPROD GEMM */
|
123
|
+
/* .kern_info = */ {
|
124
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
125
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
126
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
127
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
128
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
129
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
130
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
131
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
132
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
133
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
134
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
135
|
+
},
|
136
|
+
/* DOTPROD GEMV */
|
137
|
+
/* .kern_info = */ {
|
138
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
139
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
140
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
141
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
142
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
143
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
144
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
145
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
146
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
147
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
148
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
149
|
+
},
|
150
|
+
/* .lhs_info = */ {
|
151
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
|
152
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
|
153
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
|
154
|
+
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
|
155
|
+
},
|
156
|
+
/* .rhs_info = */ {
|
157
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
158
|
+
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
159
|
+
},
|
160
|
+
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
|
161
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
162
|
+
/* .rhs_type = */ GGML_TYPE_Q4_0,
|
163
|
+
/* .op_type = */ GGML_TYPE_F32,
|
164
|
+
},
|
165
|
+
#endif
|
166
|
+
#if defined(__ARM_FEATURE_MATMUL_INT8)
|
167
|
+
{
|
168
|
+
/* i8mm GEMM */
|
169
|
+
/* .kern_info = */ {
|
170
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
171
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
172
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
173
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
174
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
175
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
176
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
177
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
178
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
179
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
180
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
181
|
+
},
|
182
|
+
/* i8mm GEMV */
|
183
|
+
/* .kern_info = */ {
|
184
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
185
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
186
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
187
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
188
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
189
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
190
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
191
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
192
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
193
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
194
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
195
|
+
},
|
196
|
+
/* .lhs_info = */ {
|
197
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
|
198
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
|
199
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
|
200
|
+
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
|
201
|
+
},
|
202
|
+
/* .rhs_info = */ {
|
203
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
204
|
+
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
205
|
+
},
|
206
|
+
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
|
207
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
208
|
+
/* .rhs_type = */ GGML_TYPE_Q4_0,
|
209
|
+
/* .op_type = */ GGML_TYPE_F32,
|
210
|
+
},
|
211
|
+
#endif
|
212
|
+
#else
|
213
|
+
#if defined(__ARM_FEATURE_MATMUL_INT8)
|
214
|
+
{
|
215
|
+
/* i8mm GEMM */
|
216
|
+
/* .kern_info = */ {
|
217
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
218
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
219
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
220
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
221
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
222
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
223
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
224
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
225
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
226
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
227
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
|
228
|
+
},
|
229
|
+
/* i8mm GEMV */
|
230
|
+
/* .kern_info = */ {
|
231
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
232
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
233
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
234
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
235
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
236
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
237
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
238
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
239
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
240
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
241
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
|
242
|
+
},
|
243
|
+
/* .lhs_info = */ {
|
244
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
|
245
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
|
246
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
|
247
|
+
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
|
248
|
+
},
|
249
|
+
/* .rhs_info = */ {
|
250
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
251
|
+
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
252
|
+
},
|
253
|
+
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
|
254
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
255
|
+
/* .rhs_type = */ GGML_TYPE_Q4_0,
|
256
|
+
/* .op_type = */ GGML_TYPE_F32,
|
257
|
+
},
|
258
|
+
#endif
|
259
|
+
#if defined(__ARM_FEATURE_DOTPROD)
|
260
|
+
{
|
261
|
+
/* DOTPROD GEMM */
|
262
|
+
/* .kern_info = */ {
|
263
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
264
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
265
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
266
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
267
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
268
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
269
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
270
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
271
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
272
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
273
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
|
274
|
+
},
|
275
|
+
/* DOTPROD GEMV */
|
276
|
+
/* .kern_info = */ {
|
277
|
+
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
278
|
+
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
279
|
+
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
280
|
+
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
281
|
+
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
282
|
+
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
283
|
+
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
284
|
+
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
285
|
+
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
286
|
+
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
287
|
+
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
|
288
|
+
},
|
289
|
+
/* .lhs_info = */ {
|
290
|
+
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
|
291
|
+
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
|
292
|
+
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
|
293
|
+
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
|
294
|
+
},
|
295
|
+
/* .rhs_info = */ {
|
296
|
+
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
297
|
+
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
|
298
|
+
},
|
299
|
+
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
|
300
|
+
/* .lhs_type = */ GGML_TYPE_F32,
|
301
|
+
/* .rhs_type = */ GGML_TYPE_Q4_0,
|
302
|
+
/* .op_type = */ GGML_TYPE_F32,
|
303
|
+
},
|
304
|
+
#endif
|
305
|
+
#endif
|
306
|
+
};
|
307
|
+
|
308
|
+
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
|
309
|
+
ggml_kleidiai_kernels * kernel = nullptr;
|
310
|
+
|
311
|
+
if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
|
312
|
+
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
|
313
|
+
if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
|
314
|
+
gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
|
315
|
+
gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
|
316
|
+
gemm_gemv_kernels[i].op_type == tensor->type) {
|
317
|
+
kernel = &gemm_gemv_kernels[i];
|
318
|
+
break;
|
319
|
+
}
|
320
|
+
}
|
321
|
+
}
|
322
|
+
|
323
|
+
return kernel;
|
324
|
+
}
|
325
|
+
|
326
|
+
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
|
327
|
+
ggml_kleidiai_kernels * kernels = nullptr;
|
328
|
+
|
329
|
+
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
|
330
|
+
if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
|
331
|
+
kernels = &gemm_gemv_kernels[i];
|
332
|
+
break;
|
333
|
+
}
|
334
|
+
}
|
335
|
+
|
336
|
+
return kernels;
|
337
|
+
}
|
@@ -0,0 +1,95 @@
|
|
1
|
+
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
2
|
+
// SPDX-License-Identifier: MIT
|
3
|
+
//
|
4
|
+
|
5
|
+
#pragma once
|
6
|
+
|
7
|
+
#include <functional>
|
8
|
+
#include <variant>
|
9
|
+
#include "ggml.h"
|
10
|
+
|
11
|
+
enum cpu_feature {
|
12
|
+
CPU_FEATURE_NONE = 0,
|
13
|
+
CPU_FEATURE_DOTPROD = 1,
|
14
|
+
CPU_FEATURE_I8MM = 2,
|
15
|
+
CPU_FEATURE_SVE = 4,
|
16
|
+
CPU_FEATURE_SME = 8
|
17
|
+
};
|
18
|
+
inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
|
19
|
+
lhs = static_cast<cpu_feature>(lhs | rhs);
|
20
|
+
return lhs;
|
21
|
+
}
|
22
|
+
inline cpu_feature operator|(cpu_feature lhs, cpu_feature rhs) {
|
23
|
+
return static_cast<cpu_feature>(static_cast<int>(lhs) | static_cast<int>(rhs));
|
24
|
+
}
|
25
|
+
|
26
|
+
struct kernel_info {
|
27
|
+
size_t (*get_m_step)(void);
|
28
|
+
size_t (*get_n_step)(void);
|
29
|
+
size_t (*get_mr)(void);
|
30
|
+
size_t (*get_nr)(void);
|
31
|
+
size_t (*get_kr)(void);
|
32
|
+
size_t (*get_sr)(void);
|
33
|
+
std::variant<
|
34
|
+
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
|
35
|
+
std::function<size_t(size_t m_idx, size_t k)>
|
36
|
+
> get_lhs_offset;
|
37
|
+
std::variant<
|
38
|
+
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
|
39
|
+
std::function<size_t(size_t n_idx, size_t k)>
|
40
|
+
> get_rhs_packed_offset;
|
41
|
+
size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
|
42
|
+
size_t (*get_dst_size)(size_t m, size_t n);
|
43
|
+
std::variant<
|
44
|
+
std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
|
45
|
+
float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
|
46
|
+
std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
|
47
|
+
size_t dst_stride_col, float clamp_min, float clamp_max)>
|
48
|
+
> run_kernel;
|
49
|
+
};
|
50
|
+
|
51
|
+
struct lhs_packing_info {
|
52
|
+
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
|
53
|
+
std::variant<
|
54
|
+
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
|
55
|
+
std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
|
56
|
+
> get_packed_offset;
|
57
|
+
std::variant<
|
58
|
+
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
|
59
|
+
std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
|
60
|
+
> packed_size;
|
61
|
+
std::variant<
|
62
|
+
std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
|
63
|
+
size_t lhs_stride, void* lhs_packed)>,
|
64
|
+
std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
|
65
|
+
void* lhs_packed)>
|
66
|
+
> pack_func;
|
67
|
+
};
|
68
|
+
|
69
|
+
struct rhs_packing_info {
|
70
|
+
std::variant<
|
71
|
+
std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
|
72
|
+
std::function<size_t(size_t n, size_t k)>
|
73
|
+
> packed_size;
|
74
|
+
std::variant<
|
75
|
+
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
|
76
|
+
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
|
77
|
+
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
|
78
|
+
const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
|
79
|
+
> pack_func;
|
80
|
+
};
|
81
|
+
|
82
|
+
struct ggml_kleidiai_kernels {
|
83
|
+
kernel_info gemm;
|
84
|
+
kernel_info gemv;
|
85
|
+
lhs_packing_info lhs_info;
|
86
|
+
rhs_packing_info rhs_info;
|
87
|
+
|
88
|
+
cpu_feature required_cpu;
|
89
|
+
ggml_type lhs_type;
|
90
|
+
ggml_type rhs_type;
|
91
|
+
ggml_type op_type;
|
92
|
+
};
|
93
|
+
|
94
|
+
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
|
95
|
+
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);
|