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
@@ -0,0 +1,360 @@
|
|
1
|
+
#version 450
|
2
|
+
|
3
|
+
#extension GL_EXT_control_flow_attributes : enable
|
4
|
+
#extension GL_EXT_shader_16bit_storage : require
|
5
|
+
|
6
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
|
7
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
|
8
|
+
|
9
|
+
#extension GL_KHR_shader_subgroup_basic : enable
|
10
|
+
#extension GL_KHR_memory_scope_semantics : enable
|
11
|
+
#extension GL_KHR_cooperative_matrix : enable
|
12
|
+
|
13
|
+
#include "types.comp"
|
14
|
+
#include "flash_attn_base.comp"
|
15
|
+
|
16
|
+
const uint32_t D_per_thread = D / D_split;
|
17
|
+
const uint32_t row_split = 4;
|
18
|
+
const uint32_t rows_per_thread = Br / row_split;
|
19
|
+
const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
|
20
|
+
const uint32_t cols_per_thread = Bc / cols_per_iter;
|
21
|
+
|
22
|
+
|
23
|
+
layout (binding = 0) readonly buffer Q {float data_q[];};
|
24
|
+
layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
|
25
|
+
layout (binding = 1) readonly buffer K {float16_t data_k[];};
|
26
|
+
layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
|
27
|
+
layout (binding = 2) readonly buffer V {float16_t data_v[];};
|
28
|
+
layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
|
29
|
+
layout (binding = 3) readonly buffer M {float16_t data_m[];};
|
30
|
+
|
31
|
+
// Store the output when doing grouped query attention.
|
32
|
+
// Rows index by Q's dimension 2, and the first N rows are valid.
|
33
|
+
D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
|
34
|
+
{
|
35
|
+
uint32_t offset = (iq2 + r) * D + c;
|
36
|
+
data_o[o_offset + offset] = D_TYPE(elem);
|
37
|
+
return elem;
|
38
|
+
}
|
39
|
+
|
40
|
+
// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd
|
41
|
+
const uint32_t MatBr = 16;
|
42
|
+
const uint32_t MatBc = 16;
|
43
|
+
|
44
|
+
shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
|
45
|
+
shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
|
46
|
+
|
47
|
+
const uint32_t qstride = D / 4 + 2; // in units of f16vec4
|
48
|
+
shared f16vec4 Qf[Br * qstride];
|
49
|
+
|
50
|
+
// Avoid padding for D==256 to make it fit in 48KB shmem.
|
51
|
+
const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
|
52
|
+
shared ACC_TYPE sfsh[Bc * sfshstride];
|
53
|
+
|
54
|
+
const uint32_t kshstride = D / 4 + 2; // in units of f16vec4
|
55
|
+
shared f16vec4 ksh[Bc * kshstride];
|
56
|
+
|
57
|
+
shared float slope[Br];
|
58
|
+
|
59
|
+
void main() {
|
60
|
+
#ifdef NEEDS_INIT_IQ_SHMEM
|
61
|
+
init_iq_shmem(gl_WorkGroupSize);
|
62
|
+
#endif
|
63
|
+
|
64
|
+
init_indices();
|
65
|
+
|
66
|
+
const uint32_t tid = gl_LocalInvocationIndex;
|
67
|
+
|
68
|
+
const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
|
69
|
+
const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
|
70
|
+
const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
|
71
|
+
const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
|
72
|
+
|
73
|
+
#define tile_row(r) (row_tid * rows_per_thread + (r))
|
74
|
+
|
75
|
+
uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
|
76
|
+
|
77
|
+
[[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
|
78
|
+
uint32_t d = (idx + tid) % (D / 4);
|
79
|
+
uint32_t r = (idx + tid) / (D / 4);
|
80
|
+
if (r < Br && d < D / 4 &&
|
81
|
+
i * Br + r < N) {
|
82
|
+
Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
|
83
|
+
}
|
84
|
+
}
|
85
|
+
barrier();
|
86
|
+
|
87
|
+
ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
|
88
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
89
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
90
|
+
Of[r][d] = ACC_TYPEV4(0.0);
|
91
|
+
}
|
92
|
+
}
|
93
|
+
|
94
|
+
float Lf[rows_per_thread], Mf[rows_per_thread];
|
95
|
+
|
96
|
+
// Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
|
97
|
+
const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
|
98
|
+
|
99
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
100
|
+
Lf[r] = 0;
|
101
|
+
Mf[r] = NEG_FLT_MAX_OVER_2;
|
102
|
+
}
|
103
|
+
|
104
|
+
// ALiBi
|
105
|
+
if (p.max_bias > 0.0f) {
|
106
|
+
if (tid < Br) {
|
107
|
+
uint r = tid;
|
108
|
+
slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
|
109
|
+
}
|
110
|
+
barrier();
|
111
|
+
} else {
|
112
|
+
if (tid < Br) {
|
113
|
+
uint r = tid;
|
114
|
+
slope[r] = 1.0;
|
115
|
+
}
|
116
|
+
barrier();
|
117
|
+
}
|
118
|
+
|
119
|
+
#if BLOCK_SIZE > 1
|
120
|
+
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
|
121
|
+
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
|
122
|
+
#else
|
123
|
+
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
|
124
|
+
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
|
125
|
+
#endif
|
126
|
+
|
127
|
+
[[dont_unroll]]
|
128
|
+
for (uint32_t j = start_j; j < end_j; ++j) {
|
129
|
+
|
130
|
+
[[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
|
131
|
+
uint32_t d = (idx + tid) % (D / 4);
|
132
|
+
uint32_t c = (idx + tid) / (D / 4);
|
133
|
+
if (c < Bc && d < D / 4) {
|
134
|
+
#if BLOCK_SIZE > 1
|
135
|
+
uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
|
136
|
+
uint ib = coord / BLOCK_SIZE;
|
137
|
+
uint iqs = (coord % BLOCK_SIZE);
|
138
|
+
f16vec4 K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
|
139
|
+
#else
|
140
|
+
f16vec4 K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
|
141
|
+
#endif
|
142
|
+
|
143
|
+
ksh[c * kshstride + d] = K_Tf;
|
144
|
+
}
|
145
|
+
}
|
146
|
+
barrier();
|
147
|
+
|
148
|
+
// K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
|
149
|
+
// Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
|
150
|
+
// This is written transposed in order to allow for N being 8 if implementations need it
|
151
|
+
coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
|
152
|
+
coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
|
153
|
+
coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
|
154
|
+
|
155
|
+
for (uint32_t d = 0; d < D / 16; ++d) {
|
156
|
+
coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
|
157
|
+
|
158
|
+
uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
|
159
|
+
coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
|
160
|
+
|
161
|
+
SfMat = coopMatMulAdd(KMat, QMat, SfMat);
|
162
|
+
}
|
163
|
+
|
164
|
+
uint coord = gl_SubgroupID * MatBc * sfshstride;
|
165
|
+
coopMatStore(SfMat, sfsh, coord, sfshstride, gl_CooperativeMatrixLayoutRowMajor);
|
166
|
+
barrier();
|
167
|
+
|
168
|
+
if (p.logit_softcap != 0.0f) {
|
169
|
+
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
|
170
|
+
uint32_t c = (idx + tid) / Br;
|
171
|
+
uint32_t r = (idx + tid) % Br;
|
172
|
+
if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
|
173
|
+
sfsh[c * sfshstride + r] = ACC_TYPE(p.logit_softcap * tanh(sfsh[c * sfshstride + r]));
|
174
|
+
}
|
175
|
+
}
|
176
|
+
barrier();
|
177
|
+
}
|
178
|
+
|
179
|
+
if (p.mask != 0) {
|
180
|
+
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
|
181
|
+
uint32_t c = (idx + tid) % Bc;
|
182
|
+
uint32_t r = (idx + tid) / Bc;
|
183
|
+
if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
|
184
|
+
sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
|
185
|
+
}
|
186
|
+
}
|
187
|
+
barrier();
|
188
|
+
}
|
189
|
+
|
190
|
+
float eMf[rows_per_thread];
|
191
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
192
|
+
float rowmaxf = sfsh[tile_row(r) + (0 * cols_per_iter + col_tid) * sfshstride];
|
193
|
+
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
|
194
|
+
rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride]));
|
195
|
+
}
|
196
|
+
float Moldf = Mf[r];
|
197
|
+
|
198
|
+
// M = max(rowmax, Mold)
|
199
|
+
// P = e^(S - M)
|
200
|
+
// eM = e^(Mold - M)
|
201
|
+
Mf[r] = max(rowmaxf, Moldf);
|
202
|
+
eMf[r] = exp(Moldf - Mf[r]);
|
203
|
+
}
|
204
|
+
|
205
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
206
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
207
|
+
Of[r][d] = float16_t(eMf[r]) * Of[r][d];
|
208
|
+
}
|
209
|
+
}
|
210
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
211
|
+
Lf[r] = eMf[r]*Lf[r];
|
212
|
+
}
|
213
|
+
|
214
|
+
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
|
215
|
+
float Pf[rows_per_thread];
|
216
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
217
|
+
Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
|
218
|
+
Lf[r] += Pf[r];
|
219
|
+
}
|
220
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
221
|
+
#if BLOCK_SIZE > 1
|
222
|
+
uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
|
223
|
+
uint ib = coord / BLOCK_SIZE;
|
224
|
+
uint iqs = (coord % BLOCK_SIZE);
|
225
|
+
vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
|
226
|
+
#else
|
227
|
+
vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
|
228
|
+
#endif
|
229
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
230
|
+
Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
|
231
|
+
}
|
232
|
+
}
|
233
|
+
}
|
234
|
+
|
235
|
+
barrier();
|
236
|
+
}
|
237
|
+
|
238
|
+
// reduce across threads
|
239
|
+
|
240
|
+
float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread];
|
241
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
242
|
+
FLOAT_TYPE M = Mf[r];
|
243
|
+
tmpsh[tid] = M;
|
244
|
+
// Compute max across the row
|
245
|
+
barrier();
|
246
|
+
[[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
|
247
|
+
M = max(M, tmpsh[tid ^ s]);
|
248
|
+
barrier();
|
249
|
+
tmpsh[tid] = M;
|
250
|
+
barrier();
|
251
|
+
}
|
252
|
+
rowmaxf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
|
253
|
+
barrier();
|
254
|
+
}
|
255
|
+
|
256
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
257
|
+
Moldf[r] = Mf[r];
|
258
|
+
|
259
|
+
// M = max(rowmax, Mold)
|
260
|
+
// eM = e^(Mold - M)
|
261
|
+
Mf[r] = max(rowmaxf[r], Moldf[r]);
|
262
|
+
eMf[r] = exp(Moldf[r] - Mf[r]);
|
263
|
+
|
264
|
+
Lf[r] = eMf[r]*Lf[r];
|
265
|
+
}
|
266
|
+
|
267
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
268
|
+
FLOAT_TYPE L = Lf[r];
|
269
|
+
tmpsh[tid] = L;
|
270
|
+
// Compute sum across the row
|
271
|
+
barrier();
|
272
|
+
[[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
|
273
|
+
L += tmpsh[tid ^ s];
|
274
|
+
barrier();
|
275
|
+
tmpsh[tid] = L;
|
276
|
+
barrier();
|
277
|
+
}
|
278
|
+
Lf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
|
279
|
+
barrier();
|
280
|
+
}
|
281
|
+
|
282
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
283
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
284
|
+
|
285
|
+
Of[r][d] = float16_t(eMf[r]) * Of[r][d];
|
286
|
+
tmpshv4[tid] = Of[r][d];
|
287
|
+
|
288
|
+
barrier();
|
289
|
+
[[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
|
290
|
+
Of[r][d] += tmpshv4[tid ^ s];
|
291
|
+
barrier();
|
292
|
+
tmpshv4[tid] = Of[r][d];
|
293
|
+
barrier();
|
294
|
+
}
|
295
|
+
Of[r][d] = tmpshv4[d_tid + row_tid * threads_per_rowgroup];
|
296
|
+
barrier();
|
297
|
+
}
|
298
|
+
}
|
299
|
+
|
300
|
+
// If there is split_k, then the split_k resolve shader does the final
|
301
|
+
// division by L. Store the intermediate O value and per-row m and L values.
|
302
|
+
if (p.k_num > 1) {
|
303
|
+
uint32_t o_offset = D * p.ne1 * split_k_index;
|
304
|
+
|
305
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
306
|
+
if (tile_row(r) < N) {
|
307
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
308
|
+
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
|
309
|
+
perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
|
310
|
+
}
|
311
|
+
}
|
312
|
+
}
|
313
|
+
}
|
314
|
+
|
315
|
+
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
|
316
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
317
|
+
if (tile_row(r) < N) {
|
318
|
+
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
|
319
|
+
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
|
320
|
+
}
|
321
|
+
}
|
322
|
+
|
323
|
+
return;
|
324
|
+
}
|
325
|
+
|
326
|
+
float Lfrcp[rows_per_thread];
|
327
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
328
|
+
Lfrcp[r] = 1.0 / Lf[r];
|
329
|
+
}
|
330
|
+
|
331
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
332
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
333
|
+
Of[r][d] *= float16_t(Lfrcp[r]);
|
334
|
+
}
|
335
|
+
}
|
336
|
+
|
337
|
+
uint32_t o_offset = iq3*p.ne2*p.ne1;
|
338
|
+
|
339
|
+
if (p.gqa_ratio > 1) {
|
340
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
341
|
+
if (tile_row(r) < N) {
|
342
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
343
|
+
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
|
344
|
+
perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
|
345
|
+
}
|
346
|
+
}
|
347
|
+
}
|
348
|
+
}
|
349
|
+
} else {
|
350
|
+
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
|
351
|
+
if (i * Br + tile_row(r) < N) {
|
352
|
+
[[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
|
353
|
+
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
|
354
|
+
data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
|
355
|
+
}
|
356
|
+
}
|
357
|
+
}
|
358
|
+
}
|
359
|
+
}
|
360
|
+
}
|
@@ -0,0 +1,267 @@
|
|
1
|
+
#version 450
|
2
|
+
|
3
|
+
#extension GL_EXT_control_flow_attributes : enable
|
4
|
+
#extension GL_EXT_shader_16bit_storage : require
|
5
|
+
|
6
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
|
7
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
|
8
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
|
9
|
+
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
|
10
|
+
|
11
|
+
#extension GL_KHR_memory_scope_semantics : enable
|
12
|
+
#extension GL_KHR_cooperative_matrix : enable
|
13
|
+
#extension GL_NV_cooperative_matrix2 : enable
|
14
|
+
#extension GL_EXT_buffer_reference : enable
|
15
|
+
#extension GL_KHR_shader_subgroup_ballot : enable
|
16
|
+
#extension GL_KHR_shader_subgroup_vote : enable
|
17
|
+
#extension GL_EXT_null_initializer : enable
|
18
|
+
|
19
|
+
#include "types.comp"
|
20
|
+
#include "dequant_funcs_cm2.comp"
|
21
|
+
#include "flash_attn_base.comp"
|
22
|
+
|
23
|
+
layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
|
24
|
+
layout (binding = 1) readonly buffer K {uint8_t data_k[];};
|
25
|
+
layout (binding = 2) readonly buffer V {uint8_t data_v[];};
|
26
|
+
layout (binding = 3) readonly buffer M {uint8_t data_m[];};
|
27
|
+
|
28
|
+
ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
|
29
|
+
return max(x, y);
|
30
|
+
}
|
31
|
+
|
32
|
+
ACC_TYPE smearReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
|
33
|
+
return x;
|
34
|
+
}
|
35
|
+
|
36
|
+
// Replace matrix elements >= numRows or numCols with 'replace'
|
37
|
+
ACC_TYPE replacePadding(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem, const in ACC_TYPE replace, const in uint32_t numRows, const in uint32_t numCols) {
|
38
|
+
if (row >= numRows || col >= numCols) {
|
39
|
+
return replace;
|
40
|
+
}
|
41
|
+
return elem;
|
42
|
+
}
|
43
|
+
|
44
|
+
ACC_TYPE Exp(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem)
|
45
|
+
{
|
46
|
+
return exp(elem);
|
47
|
+
}
|
48
|
+
|
49
|
+
ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem0, const in ACC_TYPE elem1)
|
50
|
+
{
|
51
|
+
return max(elem0, elem1);
|
52
|
+
}
|
53
|
+
|
54
|
+
#if defined(BLOCK_SIZE)
|
55
|
+
#define DECODEFUNC , DEQUANTFUNC
|
56
|
+
#else
|
57
|
+
#define DECODEFUNC
|
58
|
+
#endif
|
59
|
+
|
60
|
+
// Store the output when doing grouped query attention.
|
61
|
+
// Rows index by Q's dimension 2, and the first N rows are valid.
|
62
|
+
D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
|
63
|
+
{
|
64
|
+
if (r < N && c < D) {
|
65
|
+
uint32_t offset = (iq2 + r) * D + c;
|
66
|
+
data_o[o_offset + offset] = D_TYPE(elem);
|
67
|
+
}
|
68
|
+
return elem;
|
69
|
+
}
|
70
|
+
|
71
|
+
void main() {
|
72
|
+
#ifdef NEEDS_INIT_IQ_SHMEM
|
73
|
+
init_iq_shmem(gl_WorkGroupSize);
|
74
|
+
#endif
|
75
|
+
|
76
|
+
init_indices();
|
77
|
+
|
78
|
+
tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
|
79
|
+
tensorLayoutNV<2, Clamp> tensorLayoutK = createTensorLayoutNV(2, Clamp);
|
80
|
+
tensorLayoutNV<2, Clamp> tensorLayoutV = createTensorLayoutNV(2, Clamp);
|
81
|
+
|
82
|
+
tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0);
|
83
|
+
|
84
|
+
#if defined(BLOCK_SIZE)
|
85
|
+
tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE);
|
86
|
+
tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
|
87
|
+
#endif
|
88
|
+
|
89
|
+
tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
|
90
|
+
tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
|
91
|
+
tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
|
92
|
+
|
93
|
+
// hint to the compiler that strides are aligned for the aligned variant of the shader
|
94
|
+
if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
|
95
|
+
{
|
96
|
+
q_stride &= ~7;
|
97
|
+
#if !defined(BLOCK_SIZE)
|
98
|
+
k_stride &= ~7;
|
99
|
+
v_stride &= ~7;
|
100
|
+
#endif
|
101
|
+
m_stride &= ~7;
|
102
|
+
}
|
103
|
+
tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1);
|
104
|
+
tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
|
105
|
+
tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
|
106
|
+
|
107
|
+
coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
|
108
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
|
109
|
+
|
110
|
+
uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
|
111
|
+
coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
|
112
|
+
|
113
|
+
Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
|
114
|
+
Qf16 *= float16_t(p.scale);
|
115
|
+
|
116
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
|
117
|
+
|
118
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
|
119
|
+
|
120
|
+
// Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
|
121
|
+
const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
|
122
|
+
|
123
|
+
L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
|
124
|
+
M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2);
|
125
|
+
|
126
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);
|
127
|
+
|
128
|
+
// ALiBi
|
129
|
+
if (p.max_bias > 0.0f) {
|
130
|
+
coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
|
131
|
+
}
|
132
|
+
|
133
|
+
[[dont_unroll]]
|
134
|
+
for (uint32_t j = start_j; j < end_j; ++j) {
|
135
|
+
|
136
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
|
137
|
+
|
138
|
+
coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T;
|
139
|
+
|
140
|
+
uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
|
141
|
+
coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
|
142
|
+
S = coopMatMulAdd(Qf16, K_T, S);
|
143
|
+
|
144
|
+
if (p.logit_softcap != 0.0f) {
|
145
|
+
[[unroll]]
|
146
|
+
for (int k = 0; k < S.length(); ++k) {
|
147
|
+
S[k] = ACC_TYPE(p.logit_softcap)*tanh(S[k]);
|
148
|
+
}
|
149
|
+
}
|
150
|
+
|
151
|
+
if (p.mask != 0) {
|
152
|
+
tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
|
153
|
+
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
|
154
|
+
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
|
155
|
+
|
156
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
|
157
|
+
|
158
|
+
coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
|
159
|
+
|
160
|
+
S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
|
161
|
+
}
|
162
|
+
|
163
|
+
// Clear padding elements to -inf, so they don't contribute to rowmax
|
164
|
+
if (Clamp != 0 &&
|
165
|
+
((j + 1) * Bc > KV ||
|
166
|
+
(i + 1) * Br > N)) {
|
167
|
+
|
168
|
+
uint R = ((i + 1) * Br > N) ? (N % Br) : Br;
|
169
|
+
uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;
|
170
|
+
|
171
|
+
coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(NEG_FLT_MAX_OVER_2), R, C);
|
172
|
+
}
|
173
|
+
|
174
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> rowmax, P, rowsum, eM;
|
175
|
+
|
176
|
+
coopMatReduceNV(rowmax, S, gl_CooperativeMatrixReduceRowNV, maxReduce);
|
177
|
+
|
178
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> Mold = M;
|
179
|
+
|
180
|
+
// M = max(rowmax, Mold)
|
181
|
+
// P = e^(S - M)
|
182
|
+
// eM = e^(Mold - M)
|
183
|
+
coopMatPerElementNV(M, rowmax, Max, Mold);
|
184
|
+
coopMatPerElementNV(P, S - M, Exp);
|
185
|
+
coopMatPerElementNV(eM, Mold - M, Exp);
|
186
|
+
|
187
|
+
// Clear padding elements to 0, so they don't contribute to rowsum
|
188
|
+
if (Clamp != 0 &&
|
189
|
+
((j + 1) * Bc > KV ||
|
190
|
+
(i + 1) * Br > N)) {
|
191
|
+
|
192
|
+
uint R = ((i + 1) * Br > N) ? (N % Br) : Br;
|
193
|
+
uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;
|
194
|
+
|
195
|
+
coopMatPerElementNV(P, P, replacePadding, ACC_TYPE(0.0), R, C);
|
196
|
+
}
|
197
|
+
|
198
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA> P_A = coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA>(P);
|
199
|
+
|
200
|
+
// compute rowsum by multiplying by matrix of all ones.
|
201
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB> One = coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB>(1.0);
|
202
|
+
|
203
|
+
rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
|
204
|
+
rowsum = coopMatMulAdd(P_A, One, rowsum);
|
205
|
+
|
206
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
|
207
|
+
uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
|
208
|
+
coopMatLoadTensorNV(V, data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
|
209
|
+
|
210
|
+
L = eM*L + rowsum;
|
211
|
+
|
212
|
+
// This is the "diagonal" matrix in the paper, but since we do componentwise
|
213
|
+
// multiply rather than matrix multiply it has the diagonal element smeared
|
214
|
+
// across the row
|
215
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
|
216
|
+
|
217
|
+
// resize eM by using smear/reduce
|
218
|
+
coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
|
219
|
+
|
220
|
+
// multiply with fp16 accumulation, then add to O.
|
221
|
+
coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
|
222
|
+
PV = coopMatMulAdd(P_A, V, PV);
|
223
|
+
|
224
|
+
O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV);
|
225
|
+
}
|
226
|
+
|
227
|
+
// If there is split_k, then the split_k resolve shader does the final
|
228
|
+
// division by L. Store the intermediate O value and per-row m and L values.
|
229
|
+
if (p.k_num > 1) {
|
230
|
+
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
|
231
|
+
|
232
|
+
uint32_t o_offset = D * p.ne1 * split_k_index;
|
233
|
+
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
|
234
|
+
|
235
|
+
o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
|
236
|
+
coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
|
237
|
+
coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
|
238
|
+
return;
|
239
|
+
}
|
240
|
+
|
241
|
+
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
|
242
|
+
|
243
|
+
// resize L by using smear/reduce
|
244
|
+
coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
|
245
|
+
|
246
|
+
[[unroll]]
|
247
|
+
for (int k = 0; k < Ldiag.length(); ++k) {
|
248
|
+
Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
|
249
|
+
}
|
250
|
+
|
251
|
+
O = Ldiag*O;
|
252
|
+
|
253
|
+
uint32_t o_offset = iq3*p.ne2*p.ne1;
|
254
|
+
|
255
|
+
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
|
256
|
+
if (p.gqa_ratio > 1) {
|
257
|
+
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
|
258
|
+
} else {
|
259
|
+
tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
|
260
|
+
tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
|
261
|
+
|
262
|
+
// permute dimensions
|
263
|
+
tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
|
264
|
+
|
265
|
+
coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute);
|
266
|
+
}
|
267
|
+
}
|
@@ -0,0 +1,59 @@
|
|
1
|
+
#version 450
|
2
|
+
|
3
|
+
#extension GL_EXT_control_flow_attributes : enable
|
4
|
+
|
5
|
+
#define BLOCK_SIZE 32
|
6
|
+
|
7
|
+
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
|
8
|
+
|
9
|
+
layout (binding = 0) readonly buffer A {float data_a[];};
|
10
|
+
layout (binding = 1) writeonly buffer D {float data_d[];};
|
11
|
+
|
12
|
+
layout (push_constant) uniform parameter {
|
13
|
+
uint D;
|
14
|
+
uint N;
|
15
|
+
uint k_num;
|
16
|
+
} p;
|
17
|
+
|
18
|
+
void main() {
|
19
|
+
// Each workgroup handles a row
|
20
|
+
const uint n = gl_WorkGroupID.x;
|
21
|
+
const uint tid = gl_LocalInvocationID.x;
|
22
|
+
|
23
|
+
uint D = p.D;
|
24
|
+
uint N = p.N;
|
25
|
+
uint k_num = p.k_num;
|
26
|
+
|
27
|
+
uint l_offset = D * N * k_num + n;
|
28
|
+
uint m_offset = D * N * k_num + N + n;
|
29
|
+
uint lm_stride = N * 2;
|
30
|
+
|
31
|
+
// Compute the max m value for the row
|
32
|
+
float m_max = -1.0/0.0;
|
33
|
+
[[unroll]] for (uint k = 0; k < k_num; ++k) {
|
34
|
+
float m = data_a[m_offset + k * lm_stride];
|
35
|
+
m_max = max(m_max, m);
|
36
|
+
}
|
37
|
+
|
38
|
+
// Compute L based on m_max
|
39
|
+
float L = 0;
|
40
|
+
[[unroll]] for (uint k = 0; k < k_num; ++k) {
|
41
|
+
float l = data_a[l_offset + k * lm_stride];
|
42
|
+
float m = data_a[m_offset + k * lm_stride];
|
43
|
+
L += exp(m - m_max) * l;
|
44
|
+
}
|
45
|
+
|
46
|
+
L = 1.0 / L;
|
47
|
+
|
48
|
+
// Scale and sum the O contributions based on m_max and store the result to memory
|
49
|
+
for (uint d = tid; d < D; d += BLOCK_SIZE) {
|
50
|
+
float O = 0.0;
|
51
|
+
[[unroll]] for (uint k = 0; k < k_num; ++k) {
|
52
|
+
uint o_offset = D * N * k + D * n + d;
|
53
|
+
float m = data_a[m_offset + k * lm_stride];
|
54
|
+
O += exp(m - m_max) * data_a[o_offset];
|
55
|
+
}
|
56
|
+
O *= L;
|
57
|
+
data_d[D * n + d] = O;
|
58
|
+
}
|
59
|
+
}
|