elelem-ggml 0.1.0
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 +7 -0
- data/LICENSE.txt +21 -0
- data/exe/elelem-ggml +8 -0
- data/ext/elelem/ggml/elelem.c +466 -0
- data/ext/elelem/ggml/extconf.rb +52 -0
- data/lib/elelem/ggml/chat_ml.rb +23 -0
- data/lib/elelem/ggml/gguf.rb +82 -0
- data/lib/elelem/ggml/provider.rb +63 -0
- data/lib/elelem/ggml/tokenizer.rb +95 -0
- data/lib/elelem/ggml/version.rb +7 -0
- data/lib/elelem/ggml.rb +37 -0
- data/vendor/ggml/AUTHORS +335 -0
- data/vendor/ggml/CMakeLists.txt +499 -0
- data/vendor/ggml/CONTRIBUTING.md +3 -0
- data/vendor/ggml/LICENSE +21 -0
- data/vendor/ggml/README.md +49 -0
- data/vendor/ggml/cmake/FindNCCL.cmake +36 -0
- data/vendor/ggml/cmake/GitVars.cmake +22 -0
- data/vendor/ggml/cmake/common.cmake +50 -0
- data/vendor/ggml/cmake/ggml-config.cmake.in +216 -0
- data/vendor/ggml/ggml.pc.in +10 -0
- data/vendor/ggml/include/ggml-alloc.h +86 -0
- data/vendor/ggml/include/ggml-backend.h +437 -0
- data/vendor/ggml/include/ggml-blas.h +25 -0
- data/vendor/ggml/include/ggml-cann.h +123 -0
- data/vendor/ggml/include/ggml-cpp.h +39 -0
- data/vendor/ggml/include/ggml-cpu.h +152 -0
- data/vendor/ggml/include/ggml-cuda.h +47 -0
- data/vendor/ggml/include/ggml-et.h +28 -0
- data/vendor/ggml/include/ggml-hexagon.h +19 -0
- data/vendor/ggml/include/ggml-metal.h +61 -0
- data/vendor/ggml/include/ggml-opencl.h +26 -0
- data/vendor/ggml/include/ggml-openvino.h +37 -0
- data/vendor/ggml/include/ggml-opt.h +256 -0
- data/vendor/ggml/include/ggml-rpc.h +35 -0
- data/vendor/ggml/include/ggml-sycl.h +57 -0
- data/vendor/ggml/include/ggml-virtgpu.h +14 -0
- data/vendor/ggml/include/ggml-vulkan.h +29 -0
- data/vendor/ggml/include/ggml-webgpu.h +19 -0
- data/vendor/ggml/include/ggml-zdnn.h +17 -0
- data/vendor/ggml/include/ggml-zendnn.h +22 -0
- data/vendor/ggml/include/ggml.h +2963 -0
- data/vendor/ggml/include/gguf.h +211 -0
- data/vendor/ggml/requirements.txt +12 -0
- data/vendor/ggml/scripts/gen-authors.sh +9 -0
- data/vendor/ggml/scripts/make-release-checks.sh +96 -0
- data/vendor/ggml/scripts/make-release-desc.sh +69 -0
- data/vendor/ggml/scripts/make-release-summary.txt +33 -0
- data/vendor/ggml/scripts/release.sh +195 -0
- data/vendor/ggml/scripts/sync-llama-am.sh +167 -0
- data/vendor/ggml/scripts/sync-llama.last +1 -0
- data/vendor/ggml/scripts/sync-llama.sh +21 -0
- data/vendor/ggml/scripts/sync-whisper-am.sh +138 -0
- data/vendor/ggml/scripts/sync-whisper.last +1 -0
- data/vendor/ggml/scripts/sync-whisper.sh +17 -0
- data/vendor/ggml/src/CMakeLists.txt +634 -0
- data/vendor/ggml/src/ggml-alloc.c +1249 -0
- data/vendor/ggml/src/ggml-backend-dl.cpp +48 -0
- data/vendor/ggml/src/ggml-backend-dl.h +45 -0
- data/vendor/ggml/src/ggml-backend-impl.h +291 -0
- data/vendor/ggml/src/ggml-backend-meta.cpp +2511 -0
- data/vendor/ggml/src/ggml-backend-reg.cpp +605 -0
- data/vendor/ggml/src/ggml-backend.cpp +2517 -0
- data/vendor/ggml/src/ggml-blas/CMakeLists.txt +101 -0
- data/vendor/ggml/src/ggml-blas/ggml-blas.cpp +530 -0
- data/vendor/ggml/src/ggml-cann/CMakeLists.txt +89 -0
- data/vendor/ggml/src/ggml-cann/acl_tensor.cpp +195 -0
- data/vendor/ggml/src/ggml-cann/acl_tensor.h +349 -0
- data/vendor/ggml/src/ggml-cann/aclnn_ops.cpp +4479 -0
- data/vendor/ggml/src/ggml-cann/aclnn_ops.h +1191 -0
- data/vendor/ggml/src/ggml-cann/common.h +651 -0
- data/vendor/ggml/src/ggml-cann/ggml-cann.cpp +3070 -0
- data/vendor/ggml/src/ggml-common.h +1911 -0
- data/vendor/ggml/src/ggml-cpu/CMakeLists.txt +685 -0
- data/vendor/ggml/src/ggml-cpu/amx/amx.cpp +249 -0
- data/vendor/ggml/src/ggml-cpu/amx/amx.h +8 -0
- data/vendor/ggml/src/ggml-cpu/amx/common.h +115 -0
- data/vendor/ggml/src/ggml-cpu/amx/mmq.cpp +2511 -0
- data/vendor/ggml/src/ggml-cpu/amx/mmq.h +10 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +41 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/quants.c +4319 -0
- data/vendor/ggml/src/ggml-cpu/arch/arm/repack.cpp +5156 -0
- data/vendor/ggml/src/ggml-cpu/arch/loongarch/quants.c +2309 -0
- data/vendor/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
- data/vendor/ggml/src/ggml-cpu/arch/powerpc/quants.c +2304 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/quants.c +6596 -0
- data/vendor/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1703 -0
- data/vendor/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
- data/vendor/ggml/src/ggml-cpu/arch/s390/quants.c +1465 -0
- data/vendor/ggml/src/ggml-cpu/arch/wasm/quants.c +1292 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/quants.c +4108 -0
- data/vendor/ggml/src/ggml-cpu/arch/x86/repack.cpp +6407 -0
- data/vendor/ggml/src/ggml-cpu/arch-fallback.h +353 -0
- data/vendor/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
- data/vendor/ggml/src/ggml-cpu/binary-ops.h +16 -0
- data/vendor/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
- data/vendor/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
- data/vendor/ggml/src/ggml-cpu/common.h +95 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu-impl.h +539 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu.c +3936 -0
- data/vendor/ggml/src/ggml-cpu/ggml-cpu.cpp +712 -0
- data/vendor/ggml/src/ggml-cpu/hbm.cpp +55 -0
- data/vendor/ggml/src/ggml-cpu/hbm.h +8 -0
- data/vendor/ggml/src/ggml-cpu/iqp.cpp +1253 -0
- data/vendor/ggml/src/ggml-cpu/iqp.h +39 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/CMakeLists.txt +14 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.cpp +1101 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.h +101 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +1919 -0
- data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
- data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4164 -0
- data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
- data/vendor/ggml/src/ggml-cpu/ops.cpp +12164 -0
- data/vendor/ggml/src/ggml-cpu/ops.h +124 -0
- data/vendor/ggml/src/ggml-cpu/quants.c +1339 -0
- data/vendor/ggml/src/ggml-cpu/quants.h +106 -0
- data/vendor/ggml/src/ggml-cpu/repack.cpp +4836 -0
- data/vendor/ggml/src/ggml-cpu/repack.h +245 -0
- data/vendor/ggml/src/ggml-cpu/simd-gemm.h +226 -0
- data/vendor/ggml/src/ggml-cpu/simd-mappings.h +1319 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime.cpp +1742 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime.h +21 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +1027 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/ime_kernels.h +189 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
- data/vendor/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
- data/vendor/ggml/src/ggml-cpu/traits.cpp +36 -0
- data/vendor/ggml/src/ggml-cpu/traits.h +38 -0
- data/vendor/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
- data/vendor/ggml/src/ggml-cpu/unary-ops.h +35 -0
- data/vendor/ggml/src/ggml-cpu/vec.cpp +613 -0
- data/vendor/ggml/src/ggml-cpu/vec.h +1570 -0
- data/vendor/ggml/src/ggml-cuda/CMakeLists.txt +267 -0
- data/vendor/ggml/src/ggml-cuda/acc.cu +61 -0
- data/vendor/ggml/src/ggml-cuda/acc.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/add-id.cu +58 -0
- data/vendor/ggml/src/ggml-cuda/add-id.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/allreduce.cu +971 -0
- data/vendor/ggml/src/ggml-cuda/allreduce.cuh +29 -0
- data/vendor/ggml/src/ggml-cuda/arange.cu +34 -0
- data/vendor/ggml/src/ggml-cuda/arange.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/argmax.cu +91 -0
- data/vendor/ggml/src/ggml-cuda/argmax.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/argsort.cu +292 -0
- data/vendor/ggml/src/ggml-cuda/argsort.cuh +20 -0
- data/vendor/ggml/src/ggml-cuda/binbcast.cu +574 -0
- data/vendor/ggml/src/ggml-cuda/binbcast.cuh +12 -0
- data/vendor/ggml/src/ggml-cuda/clamp.cu +45 -0
- data/vendor/ggml/src/ggml-cuda/clamp.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
- data/vendor/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/common.cuh +1678 -0
- data/vendor/ggml/src/ggml-cuda/concat.cu +242 -0
- data/vendor/ggml/src/ggml-cuda/concat.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cu +88 -0
- data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cu +115 -0
- data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/conv2d.cu +167 -0
- data/vendor/ggml/src/ggml-cuda/conv2d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/convert.cu +705 -0
- data/vendor/ggml/src/ggml-cuda/convert.cuh +66 -0
- data/vendor/ggml/src/ggml-cuda/count-equal.cu +64 -0
- data/vendor/ggml/src/ggml-cuda/count-equal.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/cp-async.cuh +57 -0
- data/vendor/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
- data/vendor/ggml/src/ggml-cuda/cpy.cu +617 -0
- data/vendor/ggml/src/ggml-cuda/cpy.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cu +177 -0
- data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/cumsum.cu +307 -0
- data/vendor/ggml/src/ggml-cuda/cumsum.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/dequantize.cuh +452 -0
- data/vendor/ggml/src/ggml-cuda/diag.cu +77 -0
- data/vendor/ggml/src/ggml-cuda/diag.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/diagmask.cu +40 -0
- data/vendor/ggml/src/ggml-cuda/diagmask.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/dsv4-hc.cu +294 -0
- data/vendor/ggml/src/ggml-cuda/dsv4-hc.cuh +6 -0
- data/vendor/ggml/src/ggml-cuda/fattn-common.cuh +1289 -0
- data/vendor/ggml/src/ggml-cuda/fattn-mma-f16.cuh +2136 -0
- data/vendor/ggml/src/ggml-cuda/fattn-swizzle.cuh +126 -0
- data/vendor/ggml/src/ggml-cuda/fattn-tile.cu +60 -0
- data/vendor/ggml/src/ggml-cuda/fattn-tile.cuh +1355 -0
- data/vendor/ggml/src/ggml-cuda/fattn-vec.cuh +611 -0
- data/vendor/ggml/src/ggml-cuda/fattn.cu +722 -0
- data/vendor/ggml/src/ggml-cuda/fattn.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/fill.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/fill.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/fwht.cu +101 -0
- data/vendor/ggml/src/ggml-cuda/fwht.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/gated_delta_net.cu +327 -0
- data/vendor/ggml/src/ggml-cuda/gated_delta_net.cuh +14 -0
- data/vendor/ggml/src/ggml-cuda/getrows.cu +490 -0
- data/vendor/ggml/src/ggml-cuda/getrows.cuh +15 -0
- data/vendor/ggml/src/ggml-cuda/ggml-cuda.cu +5781 -0
- data/vendor/ggml/src/ggml-cuda/gla.cu +93 -0
- data/vendor/ggml/src/ggml-cuda/gla.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/im2col.cu +267 -0
- data/vendor/ggml/src/ggml-cuda/im2col.cuh +6 -0
- data/vendor/ggml/src/ggml-cuda/lightning-indexer.cu +588 -0
- data/vendor/ggml/src/ggml-cuda/lightning-indexer.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/mean.cu +77 -0
- data/vendor/ggml/src/ggml-cuda/mean.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/mma.cuh +1456 -0
- data/vendor/ggml/src/ggml-cuda/mmf.cu +191 -0
- data/vendor/ggml/src/ggml-cuda/mmf.cuh +908 -0
- data/vendor/ggml/src/ggml-cuda/mmid.cu +180 -0
- data/vendor/ggml/src/ggml-cuda/mmid.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-ampere.cuh +383 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-blackwell.cuh +37 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-cdna.cuh +185 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-dp4a.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-older.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna2.cuh +273 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh +290 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3.cuh +274 -0
- data/vendor/ggml/src/ggml-cuda/mmq-config-rdna4.cuh +290 -0
- data/vendor/ggml/src/ggml-cuda/mmq-load-tiles.cuh +1768 -0
- data/vendor/ggml/src/ggml-cuda/mmq-vec-dot.cuh +1240 -0
- data/vendor/ggml/src/ggml-cuda/mmq.cu +386 -0
- data/vendor/ggml/src/ggml-cuda/mmq.cuh +1598 -0
- data/vendor/ggml/src/ggml-cuda/mmvf.cu +875 -0
- data/vendor/ggml/src/ggml-cuda/mmvf.cuh +14 -0
- data/vendor/ggml/src/ggml-cuda/mmvq.cu +1504 -0
- data/vendor/ggml/src/ggml-cuda/mmvq.cuh +18 -0
- data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cu +65 -0
- data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/norm.cu +698 -0
- data/vendor/ggml/src/ggml-cuda/norm.cuh +18 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
- data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/out-prod.cu +125 -0
- data/vendor/ggml/src/ggml-cuda/out-prod.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/pad.cu +106 -0
- data/vendor/ggml/src/ggml-cuda/pad.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
- data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pool1d.cu +85 -0
- data/vendor/ggml/src/ggml-cuda/pool1d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/pool2d.cu +94 -0
- data/vendor/ggml/src/ggml-cuda/pool2d.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/quantize.cu +697 -0
- data/vendor/ggml/src/ggml-cuda/quantize.cuh +70 -0
- data/vendor/ggml/src/ggml-cuda/reduce_rows.cuh +43 -0
- data/vendor/ggml/src/ggml-cuda/roll.cu +67 -0
- data/vendor/ggml/src/ggml-cuda/roll.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/rope.cu +941 -0
- data/vendor/ggml/src/ggml-cuda/rope.cuh +11 -0
- data/vendor/ggml/src/ggml-cuda/scale.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/scale.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/set-rows.cu +398 -0
- data/vendor/ggml/src/ggml-cuda/set-rows.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/set.cu +39 -0
- data/vendor/ggml/src/ggml-cuda/set.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/snake.cu +72 -0
- data/vendor/ggml/src/ggml-cuda/snake.cuh +8 -0
- data/vendor/ggml/src/ggml-cuda/softcap.cu +37 -0
- data/vendor/ggml/src/ggml-cuda/softcap.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/softmax.cu +480 -0
- data/vendor/ggml/src/ggml-cuda/softmax.cuh +7 -0
- data/vendor/ggml/src/ggml-cuda/solve_tri.cu +273 -0
- data/vendor/ggml/src/ggml-cuda/solve_tri.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/ssm-conv.cu +206 -0
- data/vendor/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/ssm-scan.cu +859 -0
- data/vendor/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/sum.cu +41 -0
- data/vendor/ggml/src/ggml-cuda/sum.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/sumrows.cu +47 -0
- data/vendor/ggml/src/ggml-cuda/sumrows.cuh +4 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +6 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +11 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +12 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +111 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
- data/vendor/ggml/src/ggml-cuda/top-k.cu +275 -0
- data/vendor/ggml/src/ggml-cuda/top-k.cuh +3 -0
- data/vendor/ggml/src/ggml-cuda/topk-moe.cu +442 -0
- data/vendor/ggml/src/ggml-cuda/topk-moe.cuh +31 -0
- data/vendor/ggml/src/ggml-cuda/tri.cu +136 -0
- data/vendor/ggml/src/ggml-cuda/tri.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/tsembd.cu +47 -0
- data/vendor/ggml/src/ggml-cuda/tsembd.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/unary.cu +721 -0
- data/vendor/ggml/src/ggml-cuda/unary.cuh +123 -0
- data/vendor/ggml/src/ggml-cuda/upscale.cu +293 -0
- data/vendor/ggml/src/ggml-cuda/upscale.cuh +5 -0
- data/vendor/ggml/src/ggml-cuda/vecdotq.cuh +1371 -0
- data/vendor/ggml/src/ggml-cuda/vendors/cuda.h +28 -0
- data/vendor/ggml/src/ggml-cuda/vendors/hip.h +301 -0
- data/vendor/ggml/src/ggml-cuda/vendors/musa.h +151 -0
- data/vendor/ggml/src/ggml-cuda/wkv.cu +253 -0
- data/vendor/ggml/src/ggml-cuda/wkv.cuh +7 -0
- data/vendor/ggml/src/ggml-et/CMakeLists.txt +246 -0
- data/vendor/ggml/src/ggml-et/cmake/embed_one_kernel.cmake +15 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.cpp.in +6 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.hpp.in +12 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.cpp.in +18 -0
- data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.h.in +13 -0
- data/vendor/ggml/src/ggml-et/et-kernels/CMakeLists.txt +137 -0
- data/vendor/ggml/src/ggml-et/et-kernels/scripts/check_unimplemented_instructions.sh +36 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/RunBackend.sh +23 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/block_ops.h +997 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/clamp_f32.c +120 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/concat_f32.c +175 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f16.c +107 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f32.c +248 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/conv_2d_f32_me.c +807 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cpy_f32_f16.c +110 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/crt.S +15 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/cumsum_f32.c +96 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/diag_f32.c +90 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/el_map_f32.c +377 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/fill_f32.c +87 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f16_me.c +1000 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f32.c +217 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/gated_delta_net_f32.c +346 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/get_rows_f32.c +612 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ggml_tensor.h +44 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/glu_f32.c +606 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/group_norm_f32.c +171 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/im2col.c +130 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/l2_norm_f32.c +237 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/linker.ld +85 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/math_fp.h +299 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mean_f32.c +220 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/memops.c +181 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0.c +358 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0_matrix_engine.c +368 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q8_0.c +413 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16.c +142 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16_matrix_engine.c +329 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32.c +137 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32_matrix_engine.c +155 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q4_0.c +169 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q8_0.c +160 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_f32.c +288 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/norm_f32.c +328 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/pad_f32.c +165 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/platform.h +545 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/quants.h +72 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/repeat_f32.c +118 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_f32.c +270 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_mul_f32.c +290 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rope_f32.c +656 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv6_f32.c +184 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv7_f32.c +272 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/scale_f32.c +94 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/set_f32.c +101 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/set_rows_f32.c +394 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/softmax_f32.c +698 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/solve_tri_f32.c +109 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/sqr_f32.c +88 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_conv_f32.c +129 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_scan_f32.c +282 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/sum_rows_f32.c +103 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/tensor.h +897 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/tri_f32.c +244 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/uberkernel.c +497 -0
- data/vendor/ggml/src/ggml-et/et-kernels/src/unary_f32.c +705 -0
- data/vendor/ggml/src/ggml-et/ggml-et-common.h +86 -0
- data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.cpp +502 -0
- data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.h +54 -0
- data/vendor/ggml/src/ggml-et/ggml-et-kernels.cpp +508 -0
- data/vendor/ggml/src/ggml-et/ggml-et-kernels.h +48 -0
- data/vendor/ggml/src/ggml-et/ggml-et-memops.cpp +36 -0
- data/vendor/ggml/src/ggml-et/ggml-et-memops.h +18 -0
- data/vendor/ggml/src/ggml-et/ggml-et-ops.cpp +2584 -0
- data/vendor/ggml/src/ggml-et/ggml-et-ops.h +393 -0
- data/vendor/ggml/src/ggml-et/ggml-et-uberkernel-common.h +17 -0
- data/vendor/ggml/src/ggml-et/ggml-et.cpp +1880 -0
- data/vendor/ggml/src/ggml-feats.h +166 -0
- data/vendor/ggml/src/ggml-hexagon/CMakeLists.txt +113 -0
- data/vendor/ggml/src/ggml-hexagon/ggml-hexagon.cpp +6523 -0
- data/vendor/ggml/src/ggml-hexagon/htp/CMakeLists.txt +58 -0
- data/vendor/ggml/src/ggml-hexagon/htp/act-ops.c +585 -0
- data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.c +398 -0
- data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.h +40 -0
- data/vendor/ggml/src/ggml-hexagon/htp/argsort-ops.c +514 -0
- data/vendor/ggml/src/ggml-hexagon/htp/binary-ops.c +873 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +155 -0
- data/vendor/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cpy-ops.c +354 -0
- data/vendor/ggml/src/ggml-hexagon/htp/cumsum-ops.c +265 -0
- data/vendor/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
- data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.c +104 -0
- data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.h +408 -0
- data/vendor/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
- data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +2479 -0
- data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +344 -0
- data/vendor/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1157 -0
- data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.c +260 -0
- data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.h +77 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-bitmap.h +24 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-dma.h +2 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hex-utils.h +65 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1378 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.c +170 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.h +160 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-ctx.h +146 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-ops.h +242 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.c +300 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.h +29 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
- data/vendor/ggml/src/ggml-hexagon/htp/htp_iface.idl +23 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-arith.h +636 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-base.h +292 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-exp.h +255 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +303 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-log.h +118 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-norm.h +454 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-quant.h +165 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-reduce.h +336 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-scale.h +199 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +181 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +189 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/vendor/ggml/src/ggml-hexagon/htp/hvx-utils.h +24 -0
- data/vendor/ggml/src/ggml-hexagon/htp/im2col-ops.c +306 -0
- data/vendor/ggml/src/ggml-hexagon/htp/main.c +1245 -0
- data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.c +4078 -0
- data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.h +778 -0
- data/vendor/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
- data/vendor/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
- data/vendor/ggml/src/ggml-hexagon/htp/rope-ops.c +733 -0
- data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.c +230 -0
- data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.h +74 -0
- data/vendor/ggml/src/ggml-hexagon/htp/softmax-ops.c +407 -0
- data/vendor/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
- data/vendor/ggml/src/ggml-hexagon/htp/ssm-conv.c +433 -0
- data/vendor/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.c +1264 -0
- data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.h +176 -0
- data/vendor/ggml/src/ggml-hexagon/htp/work-queue.c +244 -0
- data/vendor/ggml/src/ggml-hexagon/htp/work-queue.h +38 -0
- data/vendor/ggml/src/ggml-hexagon/htp-drv.cpp +425 -0
- data/vendor/ggml/src/ggml-hexagon/htp-drv.h +123 -0
- data/vendor/ggml/src/ggml-hexagon/htp-opnode.h +372 -0
- data/vendor/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/vendor/ggml/src/ggml-hexagon/libggml-htp.inf +36 -0
- data/vendor/ggml/src/ggml-hip/CMakeLists.txt +153 -0
- data/vendor/ggml/src/ggml-impl.h +783 -0
- data/vendor/ggml/src/ggml-metal/CMakeLists.txt +240 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-common.cpp +474 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-common.h +56 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-context.m +759 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.cpp +2330 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.h +355 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-device.m +2473 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-impl.h +1345 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-ops.cpp +5610 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-ops.h +106 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.cpp +3732 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.h +77 -0
- data/vendor/ggml/src/ggml-metal/ggml-metal.cpp +1006 -0
- data/vendor/ggml/src/ggml-metal/kernels/argsort.metal +337 -0
- data/vendor/ggml/src/ggml-metal/kernels/binbcast.metal +228 -0
- data/vendor/ggml/src/ggml-metal/kernels/common.h +126 -0
- data/vendor/ggml/src/ggml-metal/kernels/conv.metal +724 -0
- data/vendor/ggml/src/ggml-metal/kernels/dequantize.h +735 -0
- data/vendor/ggml/src/ggml-metal/kernels/fa.metal +2445 -0
- data/vendor/ggml/src/ggml-metal/kernels/gated_delta_net.metal +250 -0
- data/vendor/ggml/src/ggml-metal/kernels/misc.metal +595 -0
- data/vendor/ggml/src/ggml-metal/kernels/mul_mm.metal +853 -0
- data/vendor/ggml/src/ggml-metal/kernels/mul_mv.metal +3225 -0
- data/vendor/ggml/src/ggml-metal/kernels/norm.metal +308 -0
- data/vendor/ggml/src/ggml-metal/kernels/pool.metal +148 -0
- data/vendor/ggml/src/ggml-metal/kernels/quantize.h +262 -0
- data/vendor/ggml/src/ggml-metal/kernels/quantize.metal +480 -0
- data/vendor/ggml/src/ggml-metal/kernels/reduce.metal +228 -0
- data/vendor/ggml/src/ggml-metal/kernels/rope.metal +333 -0
- data/vendor/ggml/src/ggml-metal/kernels/softmax.metal +223 -0
- data/vendor/ggml/src/ggml-metal/kernels/solve_tri.metal +75 -0
- data/vendor/ggml/src/ggml-metal/kernels/ssm.metal +473 -0
- data/vendor/ggml/src/ggml-metal/kernels/tri.metal +69 -0
- data/vendor/ggml/src/ggml-metal/kernels/unary.metal +400 -0
- data/vendor/ggml/src/ggml-metal/kernels/upscale.metal +179 -0
- data/vendor/ggml/src/ggml-metal/kernels/wkv.metal +179 -0
- data/vendor/ggml/src/ggml-musa/CMakeLists.txt +123 -0
- data/vendor/ggml/src/ggml-musa/mudnn.cu +112 -0
- data/vendor/ggml/src/ggml-musa/mudnn.cuh +12 -0
- data/vendor/ggml/src/ggml-opencl/CMakeLists.txt +246 -0
- data/vendor/ggml/src/ggml-opencl/cl-program-cache.cpp +453 -0
- data/vendor/ggml/src/ggml-opencl/cl-program-cache.h +75 -0
- data/vendor/ggml/src/ggml-opencl/fa_tune.h +92 -0
- data/vendor/ggml/src/ggml-opencl/ggml-opencl.cpp +26982 -0
- data/vendor/ggml/src/ggml-opencl/kernels/abs.cl +113 -0
- data/vendor/ggml/src/ggml-opencl/kernels/add.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
- data/vendor/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
- data/vendor/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
- data/vendor/ggml/src/ggml-opencl/kernels/concat.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
- data/vendor/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cpy.cl +288 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/vendor/ggml/src/ggml-opencl/kernels/cvt.cl +2663 -0
- data/vendor/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/vendor/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
- data/vendor/ggml/src/ggml-opencl/kernels/div.cl +138 -0
- data/vendor/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
- data/vendor/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
- data/vendor/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +421 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +422 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +2656 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +2011 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1840 -0
- data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +376 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_q8_1_dp4a.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_q8_1_dp4a.cl +169 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_q8_1_dp4a.cl +209 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_q8_1_dp4a.cl +200 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_0_f32_ns.cl +221 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_1_dp4a.cl +226 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_q8_1_dp4a.cl +143 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_f32.cl +139 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_q8_1_dp4a.cl +127 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +490 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_q8_1_dp4a.cl +281 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_q8_1_dp4a.cl +235 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_q8_1_dp4a.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +245 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32_tiled.cl +136 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_q8_1_dp4a.cl +144 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_q8_1_dp4a.cl +212 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +257 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +266 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +307 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl +383 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +384 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +823 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_o4.cl +349 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_tiled.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +296 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +299 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +453 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +409 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_o4.cl +372 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_tiled.cl +196 -0
- data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +280 -0
- data/vendor/ggml/src/ggml-opencl/kernels/get_rows.cl +193 -0
- data/vendor/ggml/src/ggml-opencl/kernels/glu.cl +443 -0
- data/vendor/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
- data/vendor/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
- data/vendor/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
- data/vendor/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_add_id_glu.cl +76 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_combine.cl +79 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_quant_a_q8_1.cl +64 -0
- data/vendor/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +155 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +196 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +189 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +101 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1237 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_mrow.cl +306 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +182 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +205 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +244 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +272 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
- data/vendor/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/vendor/ggml/src/ggml-opencl/kernels/norm.cl +164 -0
- data/vendor/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
- data/vendor/ggml/src/ggml-opencl/kernels/quant_a_q8_1.cl +42 -0
- data/vendor/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
- data/vendor/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
- data/vendor/ggml/src/ggml-opencl/kernels/rms_norm.cl +369 -0
- data/vendor/ggml/src/ggml-opencl/kernels/rope.cl +759 -0
- data/vendor/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
- data/vendor/ggml/src/ggml-opencl/kernels/set_rows.cl +708 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
- data/vendor/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
- data/vendor/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
- data/vendor/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
- data/vendor/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
- data/vendor/ggml/src/ggml-opencl/kernels/ssm_scan.cl +216 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
- data/vendor/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
- data/vendor/ggml/src/ggml-opencl/kernels/transpose.cl +143 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/vendor/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
- data/vendor/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
- data/vendor/ggml/src/ggml-opencl/libdl.h +79 -0
- data/vendor/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/vendor/ggml/src/ggml-openvino/ggml-decoder.cpp +2057 -0
- data/vendor/ggml/src/ggml-openvino/ggml-decoder.h +417 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +493 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.h +216 -0
- data/vendor/ggml/src/ggml-openvino/ggml-openvino.cpp +1566 -0
- data/vendor/ggml/src/ggml-openvino/ggml-quants.cpp +1273 -0
- data/vendor/ggml/src/ggml-openvino/ggml-quants.h +186 -0
- data/vendor/ggml/src/ggml-openvino/model-cache.cpp +272 -0
- data/vendor/ggml/src/ggml-openvino/model-cache.h +56 -0
- data/vendor/ggml/src/ggml-openvino/openvino/decoder.h +133 -0
- data/vendor/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/vendor/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/node_context.h +171 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/add.cpp +45 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cont.cpp +40 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cpy.cpp +295 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/cumsum.cpp +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/diag.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/fill.cpp +34 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +241 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +327 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/gather_matmul.hpp +43 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +142 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +75 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu_quick.cpp +64 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +110 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +61 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +283 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +100 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/permute.cpp +147 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/pool_2d.cpp +53 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/repeat.cpp +48 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/reshape.cpp +113 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +81 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/roll.cpp +36 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/rope.cpp +303 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/scale.cpp +86 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/set.cpp +76 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +95 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/softmax.cpp +166 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/solve_tri.cpp +108 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/sqr.cpp +35 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +61 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/transpose.cpp +54 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/tri.cpp +82 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op/view.cpp +245 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op_table.cpp +88 -0
- data/vendor/ggml/src/ggml-openvino/openvino/op_table.h +67 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp +212 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_dequantization_subgraph.h +44 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/vendor/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
- data/vendor/ggml/src/ggml-openvino/openvino/translate_session.cpp +432 -0
- data/vendor/ggml/src/ggml-openvino/openvino/translate_session.h +29 -0
- data/vendor/ggml/src/ggml-openvino/openvino/utils.cpp +886 -0
- data/vendor/ggml/src/ggml-openvino/openvino/utils.h +88 -0
- data/vendor/ggml/src/ggml-openvino/utils.cpp +1496 -0
- data/vendor/ggml/src/ggml-openvino/utils.h +195 -0
- data/vendor/ggml/src/ggml-opt.cpp +1094 -0
- data/vendor/ggml/src/ggml-quants.c +5667 -0
- data/vendor/ggml/src/ggml-quants.h +115 -0
- data/vendor/ggml/src/ggml-rpc/CMakeLists.txt +49 -0
- data/vendor/ggml/src/ggml-rpc/ggml-rpc.cpp +2351 -0
- data/vendor/ggml/src/ggml-rpc/transport-apple.cpp +481 -0
- data/vendor/ggml/src/ggml-rpc/transport-apple.h +27 -0
- data/vendor/ggml/src/ggml-rpc/transport.cpp +739 -0
- data/vendor/ggml/src/ggml-rpc/transport.h +38 -0
- data/vendor/ggml/src/ggml-sycl/CMakeLists.txt +218 -0
- data/vendor/ggml/src/ggml-sycl/add-id.cpp +81 -0
- data/vendor/ggml/src/ggml-sycl/add-id.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/backend.hpp +52 -0
- data/vendor/ggml/src/ggml-sycl/base.hpp +36 -0
- data/vendor/ggml/src/ggml-sycl/binbcast.cpp +650 -0
- data/vendor/ggml/src/ggml-sycl/binbcast.hpp +69 -0
- data/vendor/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
- data/vendor/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/common.cpp +153 -0
- data/vendor/ggml/src/ggml-sycl/common.hpp +1029 -0
- data/vendor/ggml/src/ggml-sycl/concat.cpp +509 -0
- data/vendor/ggml/src/ggml-sycl/concat.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/conv.cpp +101 -0
- data/vendor/ggml/src/ggml-sycl/conv.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-dw.cpp +170 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
- data/vendor/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv2d.cpp +150 -0
- data/vendor/ggml/src/ggml-sycl/conv2d.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/conv3d.cpp +224 -0
- data/vendor/ggml/src/ggml-sycl/conv3d.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/convert.cpp +877 -0
- data/vendor/ggml/src/ggml-sycl/convert.hpp +64 -0
- data/vendor/ggml/src/ggml-sycl/count-equal.cpp +79 -0
- data/vendor/ggml/src/ggml-sycl/count-equal.hpp +9 -0
- data/vendor/ggml/src/ggml-sycl/cpy.cpp +1433 -0
- data/vendor/ggml/src/ggml-sycl/cpy.hpp +537 -0
- data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
- data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
- data/vendor/ggml/src/ggml-sycl/cumsum.cpp +148 -0
- data/vendor/ggml/src/ggml-sycl/cumsum.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/dequantize.hpp +1681 -0
- data/vendor/ggml/src/ggml-sycl/diag.cpp +67 -0
- data/vendor/ggml/src/ggml-sycl/diag.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/dmmv.cpp +2227 -0
- data/vendor/ggml/src/ggml-sycl/dmmv.hpp +27 -0
- data/vendor/ggml/src/ggml-sycl/dpct/helper.hpp +3782 -0
- data/vendor/ggml/src/ggml-sycl/dsv4-hc.cpp +280 -0
- data/vendor/ggml/src/ggml-sycl/dsv4-hc.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/element_wise.cpp +1432 -0
- data/vendor/ggml/src/ggml-sycl/element_wise.hpp +135 -0
- data/vendor/ggml/src/ggml-sycl/esimd.hpp +589 -0
- data/vendor/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
- data/vendor/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
- data/vendor/ggml/src/ggml-sycl/fattn-common.hpp +1185 -0
- data/vendor/ggml/src/ggml-sycl/fattn-mkl.cpp +689 -0
- data/vendor/ggml/src/ggml-sycl/fattn-onednn.cpp +450 -0
- data/vendor/ggml/src/ggml-sycl/fattn-onednn.hpp +18 -0
- data/vendor/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
- data/vendor/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
- data/vendor/ggml/src/ggml-sycl/fattn-vec.hpp +684 -0
- data/vendor/ggml/src/ggml-sycl/fattn.cpp +452 -0
- data/vendor/ggml/src/ggml-sycl/fattn.hpp +42 -0
- data/vendor/ggml/src/ggml-sycl/fill.cpp +55 -0
- data/vendor/ggml/src/ggml-sycl/fill.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/fusion.cpp +212 -0
- data/vendor/ggml/src/ggml-sycl/fusion.hpp +17 -0
- data/vendor/ggml/src/ggml-sycl/fwht.cpp +119 -0
- data/vendor/ggml/src/ggml-sycl/fwht.hpp +12 -0
- data/vendor/ggml/src/ggml-sycl/gated_delta_net.cpp +366 -0
- data/vendor/ggml/src/ggml-sycl/gated_delta_net.hpp +19 -0
- data/vendor/ggml/src/ggml-sycl/gemm.hpp +93 -0
- data/vendor/ggml/src/ggml-sycl/getrows.cpp +368 -0
- data/vendor/ggml/src/ggml-sycl/getrows.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/ggml-sycl.cpp +7028 -0
- data/vendor/ggml/src/ggml-sycl/gla.cpp +106 -0
- data/vendor/ggml/src/ggml-sycl/gla.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/im2col.cpp +400 -0
- data/vendor/ggml/src/ggml-sycl/im2col.hpp +23 -0
- data/vendor/ggml/src/ggml-sycl/lightning-indexer.cpp +197 -0
- data/vendor/ggml/src/ggml-sycl/lightning-indexer.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/mem.cpp +162 -0
- data/vendor/ggml/src/ggml-sycl/mem.hpp +16 -0
- data/vendor/ggml/src/ggml-sycl/mmq.cpp +3030 -0
- data/vendor/ggml/src/ggml-sycl/mmq.hpp +33 -0
- data/vendor/ggml/src/ggml-sycl/mmvq.cpp +3033 -0
- data/vendor/ggml/src/ggml-sycl/mmvq.hpp +76 -0
- data/vendor/ggml/src/ggml-sycl/norm.cpp +963 -0
- data/vendor/ggml/src/ggml-sycl/norm.hpp +32 -0
- data/vendor/ggml/src/ggml-sycl/opt-step.cpp +131 -0
- data/vendor/ggml/src/ggml-sycl/opt-step.hpp +6 -0
- data/vendor/ggml/src/ggml-sycl/outprod.cpp +83 -0
- data/vendor/ggml/src/ggml-sycl/outprod.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/pad.cpp +97 -0
- data/vendor/ggml/src/ggml-sycl/pad.hpp +24 -0
- data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.cpp +100 -0
- data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.hpp +10 -0
- data/vendor/ggml/src/ggml-sycl/pool.cpp +185 -0
- data/vendor/ggml/src/ggml-sycl/pool.hpp +22 -0
- data/vendor/ggml/src/ggml-sycl/presets.hpp +79 -0
- data/vendor/ggml/src/ggml-sycl/quantize.hpp +133 -0
- data/vendor/ggml/src/ggml-sycl/quants.hpp +204 -0
- data/vendor/ggml/src/ggml-sycl/repeat_back.cpp +76 -0
- data/vendor/ggml/src/ggml-sycl/repeat_back.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/roll.cpp +122 -0
- data/vendor/ggml/src/ggml-sycl/roll.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/rope.cpp +651 -0
- data/vendor/ggml/src/ggml-sycl/rope.hpp +26 -0
- data/vendor/ggml/src/ggml-sycl/set.cpp +73 -0
- data/vendor/ggml/src/ggml-sycl/set.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/set_rows.cpp +577 -0
- data/vendor/ggml/src/ggml-sycl/set_rows.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/softmax.cpp +425 -0
- data/vendor/ggml/src/ggml-sycl/softmax.hpp +24 -0
- data/vendor/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
- data/vendor/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
- data/vendor/ggml/src/ggml-sycl/ssm_conv.cpp +136 -0
- data/vendor/ggml/src/ggml-sycl/ssm_conv.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/ssm_scan.cpp +170 -0
- data/vendor/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
- data/vendor/ggml/src/ggml-sycl/sycl_hw.cpp +67 -0
- data/vendor/ggml/src/ggml-sycl/sycl_hw.hpp +38 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
- data/vendor/ggml/src/ggml-sycl/topk-moe.cpp +620 -0
- data/vendor/ggml/src/ggml-sycl/topk-moe.hpp +12 -0
- data/vendor/ggml/src/ggml-sycl/tsembd.cpp +73 -0
- data/vendor/ggml/src/ggml-sycl/tsembd.hpp +20 -0
- data/vendor/ggml/src/ggml-sycl/type.hpp +112 -0
- data/vendor/ggml/src/ggml-sycl/upscale.cpp +410 -0
- data/vendor/ggml/src/ggml-sycl/upscale.hpp +9 -0
- data/vendor/ggml/src/ggml-sycl/vecdotq.hpp +1726 -0
- data/vendor/ggml/src/ggml-sycl/wkv.cpp +293 -0
- data/vendor/ggml/src/ggml-sycl/wkv.hpp +10 -0
- data/vendor/ggml/src/ggml-threading.cpp +12 -0
- data/vendor/ggml/src/ggml-threading.h +14 -0
- data/vendor/ggml/src/ggml-version.h.in +4 -0
- data/vendor/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/vendor/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +149 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-backend.cpp +72 -0
- data/vendor/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/vendor/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/vendor/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/vendor/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +194 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +54 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
- data/vendor/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
- data/vendor/ggml/src/ggml-vulkan/CMakeLists.txt +254 -0
- data/vendor/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
- data/vendor/ggml/src/ggml-vulkan/ggml-vulkan.cpp +20349 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +69 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp +28 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp +20 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +86 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp +114 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +53 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +473 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +424 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +25 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +350 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp +67 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose_02.comp +61 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +135 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp +78 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp +75 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +83 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +741 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +1466 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.glsl +13 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +49 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +40 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_0.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_tq2_0.comp +31 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +28 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fa_types.glsl +55 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/bfloat16.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e2m1.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e4m3.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/integer_dot.comp +7 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +760 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +223 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +647 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +487 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +146 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +121 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.glsl +11 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +75 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +51 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gla.comp +82 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +138 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +124 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/lightning_indexer.comp +151 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +264 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +230 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +132 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +95 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +105 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +93 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +124 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +156 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +128 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +134 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +165 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_tq2_0.comp +102 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +141 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +564 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +475 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +690 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +670 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +89 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +317 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +488 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +99 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +194 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/out_prod.comp +59 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +64 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp +43 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool1d.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +127 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +150 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +218 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +35 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +17 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +195 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +54 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp +62 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp +79 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp +65 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large_common.glsl +53 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp +81 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +60 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp +134 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +47 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl +19 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_clamp.comp +12 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp +118 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +221 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp +246 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_radix_select.comp +144 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +42 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +1938 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +178 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/utils.glsl +37 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +1367 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
- data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
- data/vendor/ggml/src/ggml-webgpu/CMakeLists.txt +83 -0
- data/vendor/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3457 -0
- data/vendor/ggml/src/ggml-webgpu/ggml-webgpu.cpp +4790 -0
- data/vendor/ggml/src/ggml-webgpu/pre_wgsl.hpp +808 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +102 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +967 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +92 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +127 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d_dw.wgsl +100 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +84 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +89 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +408 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_decls.tmpl +134 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_staging.tmpl +136 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +240 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +503 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +794 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +178 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +77 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +1052 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +159 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +210 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +167 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1560 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +70 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +151 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rope.wgsl +228 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +152 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +65 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +108 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.wgsl +210 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +120 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +247 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
- data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
- data/vendor/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
- data/vendor/ggml/src/ggml-zdnn/common.hpp +59 -0
- data/vendor/ggml/src/ggml-zdnn/ggml-zdnn.cpp +638 -0
- data/vendor/ggml/src/ggml-zdnn/mmf.cpp +80 -0
- data/vendor/ggml/src/ggml-zdnn/mmf.hpp +12 -0
- data/vendor/ggml/src/ggml-zdnn/utils.cpp +79 -0
- data/vendor/ggml/src/ggml-zdnn/utils.hpp +19 -0
- data/vendor/ggml/src/ggml-zendnn/CMakeLists.txt +91 -0
- data/vendor/ggml/src/ggml-zendnn/ggml-zendnn.cpp +838 -0
- data/vendor/ggml/src/ggml.c +8088 -0
- data/vendor/ggml/src/ggml.cpp +26 -0
- data/vendor/ggml/src/gguf.cpp +1706 -0
- metadata +1459 -0
|
@@ -0,0 +1,1000 @@
|
|
|
1
|
+
//******************************************************************************
|
|
2
|
+
// Flash Attention with TensorFMA16A32 for QK^T
|
|
3
|
+
//
|
|
4
|
+
// Uses the matrix engine for the QK^T dot products (F16×F16→F32),
|
|
5
|
+
// scalar code for online softmax and V accumulation.
|
|
6
|
+
//
|
|
7
|
+
// Hart 0: tensor engine (Q load, K load from SCP, FMA, softmax, V accum)
|
|
8
|
+
// Hart 1: pack K into double-buffered L2 SCP panels, flush for tensor_load
|
|
9
|
+
//
|
|
10
|
+
// Requirements:
|
|
11
|
+
// - Q: F32 (converted to F16 internally)
|
|
12
|
+
// - K, V: F16
|
|
13
|
+
// - dk must be a multiple of 32 (TensorFMA16A32 K-tile)
|
|
14
|
+
// - dv ≤ 512 (accumulator in shire-local L2 SCP)
|
|
15
|
+
//
|
|
16
|
+
// Parallelization: each minion independently processes one (qpos, head, batch)
|
|
17
|
+
// row, round-robin across all minion hart-0s. Hart 1 assists with K packing.
|
|
18
|
+
//******************************************************************************
|
|
19
|
+
|
|
20
|
+
#include "ggml_tensor.h"
|
|
21
|
+
#include "math_fp.h"
|
|
22
|
+
#include "platform.h"
|
|
23
|
+
#include "tensor.h"
|
|
24
|
+
|
|
25
|
+
#include <etsoc/common/utils.h>
|
|
26
|
+
#include <stdint.h>
|
|
27
|
+
#include <string.h>
|
|
28
|
+
|
|
29
|
+
#define NUM_COMPUTE_SHIRES 32
|
|
30
|
+
#define MINIONS_PER_SHIRE 32
|
|
31
|
+
|
|
32
|
+
// QK^T tiles: 16 KV positions at a time, K in chunks of 32 F16
|
|
33
|
+
#define TILE_KV 16
|
|
34
|
+
#define TILE_K 32
|
|
35
|
+
|
|
36
|
+
// L1 scratchpad layout: A (Q) in lines 0-15, B (K interleaved) in lines 16-31
|
|
37
|
+
#define A_L1_START 0
|
|
38
|
+
#define B_L1_START 16
|
|
39
|
+
|
|
40
|
+
// Max head dimensions
|
|
41
|
+
#define FA_DV_MAX 512 // max value head dim (dv)
|
|
42
|
+
#define FA_DK_MAX 512 // max key head dim (dk) - some models use hsk > hsv
|
|
43
|
+
|
|
44
|
+
typedef uint16_t et_fp16_t;
|
|
45
|
+
|
|
46
|
+
#define ET_NEG_INF_F (-3.402823466e+38f)
|
|
47
|
+
|
|
48
|
+
// L2 SCP layout per minion:
|
|
49
|
+
// [0..2047] accumulator (FA_DV_MAX * sizeof(float))
|
|
50
|
+
// [2048..4095] kpanel buffer 0 (32 × 32 × 2 = 2048 bytes)
|
|
51
|
+
// [4096..6143] kpanel buffer 1 (2048 bytes)
|
|
52
|
+
// [6144..6207] stats line - (M_p at +0, S_p at +4), own cache line
|
|
53
|
+
// Double-buffering ensures hart 0 finishes buf[N%2] before hart 1
|
|
54
|
+
// overwrites it at chunk N+2.
|
|
55
|
+
//
|
|
56
|
+
// The stats line reserves a cache-line-aligned slot for split-KV softmax
|
|
57
|
+
// partials (M_p, S_p). With k_splits=1 the slot is currently unused; step 2
|
|
58
|
+
// will populate it and use peer minions' slots during the reduction.
|
|
59
|
+
#define SCP_ACC_OFF 0
|
|
60
|
+
#define SCP_ACC_STRIDE (FA_DV_MAX * sizeof(float)) // 2048
|
|
61
|
+
#define SCP_KPANEL_SIZE (32 * 32 * sizeof(et_fp16_t)) // 2048
|
|
62
|
+
#define SCP_KP0_OFF SCP_ACC_STRIDE // 2048
|
|
63
|
+
#define SCP_KP1_OFF (SCP_KP0_OFF + SCP_KPANEL_SIZE) // 4096
|
|
64
|
+
#define SCP_STATS_OFF (SCP_KP1_OFF + SCP_KPANEL_SIZE) // 6144
|
|
65
|
+
#define SCP_STATS_SIZE 64 // own cache line
|
|
66
|
+
#define SCP_PER_MINION (SCP_STATS_OFF + SCP_STATS_SIZE) // 6208
|
|
67
|
+
|
|
68
|
+
struct ggml_et_flash_attn_ext_params {
|
|
69
|
+
struct ggml_tensor src0; // Q (F32)
|
|
70
|
+
struct ggml_tensor src1; // K (F16)
|
|
71
|
+
struct ggml_tensor src2; // V (F16)
|
|
72
|
+
struct ggml_tensor mask; // mask (F16 or F32), zeroed when absent
|
|
73
|
+
struct ggml_tensor dst; // Output (F32)
|
|
74
|
+
float scale;
|
|
75
|
+
int32_t has_mask;
|
|
76
|
+
};
|
|
77
|
+
|
|
78
|
+
static inline float get_mask_val(const struct ggml_tensor * mask, int64_t iq1, int64_t ik1, int64_t iq2, int64_t iq3) {
|
|
79
|
+
const char * base = (const char *) mask->data + iq1 * mask->nb[1] + (iq2 % mask->ne[2]) * mask->nb[2] +
|
|
80
|
+
(iq3 % mask->ne[3]) * mask->nb[3];
|
|
81
|
+
|
|
82
|
+
if (mask->type == GGML_TYPE_F32) {
|
|
83
|
+
return *(const float *) (base + ik1 * mask->nb[0]);
|
|
84
|
+
}
|
|
85
|
+
return fp16_to_fp32(*(const uint16_t *) (base + ik1 * mask->nb[0]));
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
static inline const char * get_mask_row_base(const struct ggml_tensor * mask, int64_t iq1, int64_t iq2, int64_t iq3) {
|
|
89
|
+
return (const char *) mask->data + iq1 * mask->nb[1] + (iq2 % mask->ne[2]) * mask->nb[2] +
|
|
90
|
+
(iq3 % mask->ne[3]) * mask->nb[3];
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
static inline float get_mask_val_from_base(const struct ggml_tensor * mask, const char * base, int64_t ik1) {
|
|
94
|
+
if (mask->type == GGML_TYPE_F32) {
|
|
95
|
+
return *(const float *) (base + ik1 * mask->nb[0]);
|
|
96
|
+
}
|
|
97
|
+
return fp16_to_fp32(*(const uint16_t *) (base + ik1 * mask->nb[0]));
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
// Pack K rows for TensorLoadTranspose16 (even/odd deinterleave)
|
|
101
|
+
static inline void __attribute__((always_inline)) pack_k_for_transpose16(et_fp16_t * out,
|
|
102
|
+
const char * k_base,
|
|
103
|
+
int64_t kv_start,
|
|
104
|
+
int64_t dk_start,
|
|
105
|
+
int64_t kv_count,
|
|
106
|
+
int64_t nb1_k) {
|
|
107
|
+
unsigned long old_mask;
|
|
108
|
+
__asm__ volatile(
|
|
109
|
+
"mova.x.m %[ms] \n\t"
|
|
110
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
111
|
+
: [ms] "=&r"(old_mask)
|
|
112
|
+
:
|
|
113
|
+
:);
|
|
114
|
+
|
|
115
|
+
for (int j = 0; j < (int) kv_count; ++j) {
|
|
116
|
+
const et_fp16_t * k_row = (const et_fp16_t *) (k_base + (kv_start + j) * nb1_k) + dk_start;
|
|
117
|
+
et_fp16_t * even_row = out + (j * 2) * 32;
|
|
118
|
+
et_fp16_t * odd_row = out + (j * 2 + 1) * 32;
|
|
119
|
+
__asm__ volatile(
|
|
120
|
+
"flw.ps f2, 0(%[src0]) \n\t" // load row[0..15]
|
|
121
|
+
"flw.ps f3, 0(%[src1]) \n\t" // load row[16..31]
|
|
122
|
+
"fpackreph.pi f4, f2 \n\t" // even_lo from src0
|
|
123
|
+
"fpackreph.pi f6, f3 \n\t" // even_lo from src1 (interleaved)
|
|
124
|
+
"fsrli.pi f5, f2, 16 \n\t" // shift src0 for odd
|
|
125
|
+
"fsrli.pi f7, f3, 16 \n\t" // shift src1 for odd (interleaved)
|
|
126
|
+
"fpackreph.pi f5, f5 \n\t" // odd from src0
|
|
127
|
+
"fpackreph.pi f7, f7 \n\t" // odd from src1
|
|
128
|
+
"mov.m.x m0, x0, 0x0F \n\t"
|
|
129
|
+
"fcmovm.ps f4, f4, f6 \n\t" // merge even halves
|
|
130
|
+
"fcmovm.ps f5, f5, f7 \n\t" // merge odd halves
|
|
131
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
132
|
+
"fsw.ps f4, 0(%[even]) \n\t"
|
|
133
|
+
"fsw.ps f5, 0(%[odd]) \n\t"
|
|
134
|
+
:
|
|
135
|
+
: [src0] "r"(k_row), [src1] "r"(k_row + 16), [even] "r"(even_row), [odd] "r"(odd_row)
|
|
136
|
+
: "f2", "f3", "f4", "f5", "f6", "f7", "memory");
|
|
137
|
+
}
|
|
138
|
+
|
|
139
|
+
__asm__ volatile("mova.m.x %[ms] \n\t" : : [ms] "r"(old_mask));
|
|
140
|
+
|
|
141
|
+
for (int j = (int) kv_count; j < TILE_KV; ++j) {
|
|
142
|
+
et_fp16_t * even_row = out + (j * 2) * 32;
|
|
143
|
+
et_fp16_t * odd_row = out + (j * 2 + 1) * 32;
|
|
144
|
+
for (int l = 0; l < TILE_K / 2; ++l) {
|
|
145
|
+
even_row[l] = 0;
|
|
146
|
+
odd_row[l] = 0;
|
|
147
|
+
}
|
|
148
|
+
}
|
|
149
|
+
}
|
|
150
|
+
|
|
151
|
+
// Build interleaved B panel for TensorFMA16A32 (weights @ V).
|
|
152
|
+
static inline void __attribute__((always_inline)) pack_v_interleaved(et_fp16_t * out,
|
|
153
|
+
const char * v_head,
|
|
154
|
+
int64_t kv_base,
|
|
155
|
+
int64_t dv_start,
|
|
156
|
+
int64_t kv_count,
|
|
157
|
+
int64_t nb1_v) {
|
|
158
|
+
for (int k = 0; k < TILE_KV; ++k) {
|
|
159
|
+
const int l = k >> 1;
|
|
160
|
+
const int r = k & 1;
|
|
161
|
+
et_fp16_t * const dst = out + l * 32 + r;
|
|
162
|
+
if (k < (int) kv_count) {
|
|
163
|
+
const et_fp16_t * v_row = (const et_fp16_t *) (v_head + (kv_base + k) * nb1_v) + dv_start;
|
|
164
|
+
for (int n = 0; n < 16; ++n) {
|
|
165
|
+
dst[n * 2] = v_row[n];
|
|
166
|
+
}
|
|
167
|
+
} else {
|
|
168
|
+
for (int n = 0; n < 16; ++n) {
|
|
169
|
+
dst[n * 2] = 0;
|
|
170
|
+
}
|
|
171
|
+
}
|
|
172
|
+
}
|
|
173
|
+
}
|
|
174
|
+
|
|
175
|
+
// Prefetch KV rows for one chunk into L2.
|
|
176
|
+
static inline void __attribute__((always_inline)) prefetch_kv_to_l2(const char * head,
|
|
177
|
+
int64_t kv_start,
|
|
178
|
+
int64_t d_start,
|
|
179
|
+
int64_t kv_count,
|
|
180
|
+
int64_t nb1) {
|
|
181
|
+
const void * base = (const void *) (head + kv_start * nb1 + d_start * 2);
|
|
182
|
+
l2_prefetch(base, (uint64_t) kv_count, (uint64_t) nb1);
|
|
183
|
+
}
|
|
184
|
+
|
|
185
|
+
static inline void __attribute__((always_inline)) convert_q_row_f32_to_f16(et_fp16_t * dst,
|
|
186
|
+
const float * src,
|
|
187
|
+
int64_t n) {
|
|
188
|
+
static const int32_t __attribute__((aligned(32))) offsets[8] = { 0, 2, 4, 6, 8, 10, 12, 14 };
|
|
189
|
+
|
|
190
|
+
unsigned long old_mask;
|
|
191
|
+
__asm__ volatile(
|
|
192
|
+
"mova.x.m %[ms] \n\t"
|
|
193
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
194
|
+
"flw.ps f1, 0(%[offs]) \n\t"
|
|
195
|
+
: [ms] "=&r"(old_mask)
|
|
196
|
+
: [offs] "r"(offsets)
|
|
197
|
+
: "f1");
|
|
198
|
+
|
|
199
|
+
for (int64_t d = 0; d < n; d += 8) {
|
|
200
|
+
__asm__ volatile(
|
|
201
|
+
"flw.ps f2, 0(%[src]) \n\t"
|
|
202
|
+
"fcvt.f16.ps f3, f2 \n\t"
|
|
203
|
+
"fsch.ps f3, f1(%[dst]) \n\t"
|
|
204
|
+
:
|
|
205
|
+
: [src] "r"(src + d), [dst] "r"(dst + d)
|
|
206
|
+
: "f2", "f3", "memory");
|
|
207
|
+
}
|
|
208
|
+
|
|
209
|
+
__asm__ volatile("mova.m.x %[ms] \n\t" : : [ms] "r"(old_mask));
|
|
210
|
+
}
|
|
211
|
+
|
|
212
|
+
static inline void __attribute__((always_inline)) zero_acc_vec(float * acc, int64_t dv) {
|
|
213
|
+
const float zero = 0.0f;
|
|
214
|
+
unsigned long old_mask;
|
|
215
|
+
__asm__ volatile("mova.x.m %0" : "=r"(old_mask));
|
|
216
|
+
__asm__ volatile("mov.m.x m0, x0, 0xFF");
|
|
217
|
+
__asm__ volatile("fbc.ps f2, 0(%[z])" ::[z] "r"(&zero) : "f2");
|
|
218
|
+
|
|
219
|
+
for (int64_t d = 0; d < dv; d += 8) {
|
|
220
|
+
__asm__ volatile("fsw.ps f2, 0(%[a]) \n\t" ::[a] "r"(acc + d) : "f2", "memory");
|
|
221
|
+
}
|
|
222
|
+
|
|
223
|
+
__asm__ volatile("mova.m.x %0" ::"r"(old_mask));
|
|
224
|
+
}
|
|
225
|
+
|
|
226
|
+
static inline void __attribute__((always_inline)) scale_acc_vec(float * acc, int64_t dv, float scale) {
|
|
227
|
+
unsigned long old_mask;
|
|
228
|
+
__asm__ volatile("mova.x.m %0" : "=r"(old_mask));
|
|
229
|
+
__asm__ volatile("mov.m.x m0, x0, 0xFF");
|
|
230
|
+
|
|
231
|
+
for (int64_t d = 0; d < dv; d += 8) {
|
|
232
|
+
__asm__ volatile(
|
|
233
|
+
"fbc.ps f2, 0(%[s]) \n\t"
|
|
234
|
+
"flw.ps f3, 0(%[a]) \n\t"
|
|
235
|
+
"fmul.ps f3, f3, f2 \n\t"
|
|
236
|
+
"fsw.ps f3, 0(%[a]) \n\t"
|
|
237
|
+
:
|
|
238
|
+
: [s] "r"(&scale), [a] "r"(acc + d)
|
|
239
|
+
: "f2", "f3", "memory");
|
|
240
|
+
}
|
|
241
|
+
|
|
242
|
+
__asm__ volatile("mova.m.x %0" ::"r"(old_mask));
|
|
243
|
+
}
|
|
244
|
+
|
|
245
|
+
static inline void __attribute__((always_inline)) normalize_store_vec(float * out,
|
|
246
|
+
float * acc,
|
|
247
|
+
int64_t dv,
|
|
248
|
+
float inv,
|
|
249
|
+
int use_fast_store) {
|
|
250
|
+
unsigned long old_mask;
|
|
251
|
+
__asm__ volatile("mova.x.m %0" : "=r"(old_mask));
|
|
252
|
+
__asm__ volatile("mov.m.x m0, x0, 0xFF");
|
|
253
|
+
|
|
254
|
+
for (int64_t d = 0; d < dv; d += 8) {
|
|
255
|
+
__asm__ volatile(
|
|
256
|
+
"fbc.ps f2, 0(%[inv]) \n\t"
|
|
257
|
+
"flw.ps f3, 0(%[a]) \n\t"
|
|
258
|
+
"fmul.ps f3, f3, f2 \n\t"
|
|
259
|
+
"fsw.ps f3, 0(%[a]) \n\t"
|
|
260
|
+
:
|
|
261
|
+
: [inv] "r"(&inv), [a] "r"(acc + d)
|
|
262
|
+
: "f2", "f3", "memory");
|
|
263
|
+
if (use_fast_store) {
|
|
264
|
+
__asm__ volatile(
|
|
265
|
+
"flw.ps f4, 0(%[a]) \n\t"
|
|
266
|
+
"fsw.ps f4, 0(%[o]) \n\t"
|
|
267
|
+
:
|
|
268
|
+
: [a] "r"(acc + d), [o] "r"(out + d)
|
|
269
|
+
: "f4", "memory");
|
|
270
|
+
} else {
|
|
271
|
+
atomic_store_f32((volatile float *) &out[d + 0], acc[d + 0]);
|
|
272
|
+
atomic_store_f32((volatile float *) &out[d + 1], acc[d + 1]);
|
|
273
|
+
atomic_store_f32((volatile float *) &out[d + 2], acc[d + 2]);
|
|
274
|
+
atomic_store_f32((volatile float *) &out[d + 3], acc[d + 3]);
|
|
275
|
+
atomic_store_f32((volatile float *) &out[d + 4], acc[d + 4]);
|
|
276
|
+
atomic_store_f32((volatile float *) &out[d + 5], acc[d + 5]);
|
|
277
|
+
atomic_store_f32((volatile float *) &out[d + 6], acc[d + 6]);
|
|
278
|
+
atomic_store_f32((volatile float *) &out[d + 7], acc[d + 7]);
|
|
279
|
+
}
|
|
280
|
+
}
|
|
281
|
+
|
|
282
|
+
__asm__ volatile("mova.m.x %0" ::"r"(old_mask));
|
|
283
|
+
}
|
|
284
|
+
|
|
285
|
+
static inline size_t tensor_bytes_fa(const struct ggml_tensor * t) {
|
|
286
|
+
return (size_t) t->ne[0] * t->ne[1] * t->ne[2] * t->ne[3] * t->nb[0];
|
|
287
|
+
}
|
|
288
|
+
|
|
289
|
+
// Evict a byte range from L1D to L2 SCP, splitting into batches of ≤16
|
|
290
|
+
// cache lines (the hw limit for evict_to_l2). Use before a barrier when
|
|
291
|
+
// another minion in the shire needs to read the region, or after a barrier
|
|
292
|
+
// on the reader side to drop stale L1D copies before reading peer data.
|
|
293
|
+
static inline void __attribute__((always_inline)) evict_range_to_l2(const void * addr, int64_t bytes) {
|
|
294
|
+
if (bytes <= 0) {
|
|
295
|
+
return;
|
|
296
|
+
}
|
|
297
|
+
int64_t lines = (bytes + 63) / 64;
|
|
298
|
+
const char * p = (const char *) addr;
|
|
299
|
+
while (lines > 0) {
|
|
300
|
+
int64_t batch = lines > 16 ? 16 : lines;
|
|
301
|
+
evict_to_l2((const void *) p, (uint64_t) batch, 64);
|
|
302
|
+
p += batch * 64;
|
|
303
|
+
lines -= batch;
|
|
304
|
+
}
|
|
305
|
+
}
|
|
306
|
+
|
|
307
|
+
// Split-KV online merge inner loop:
|
|
308
|
+
//
|
|
309
|
+
// for d in [0, dv) step 8:
|
|
310
|
+
// acc[d..d+8] = alpha_own * acc[d..d+8] + alpha_peer * peer_acc[d..d+8]
|
|
311
|
+
//
|
|
312
|
+
// Runs on the reducer (k_split == 0) after all tensor_fma ops for the row are
|
|
313
|
+
// complete, so f0..f31 are dead at entry. We still bracket the loop in inline
|
|
314
|
+
// asm with explicit f2/f3/f4/f5 clobbers to lock register usage down — per the
|
|
315
|
+
// MM register lifetime rule, never let the compiler mingle FP ops into code
|
|
316
|
+
// that sits anywhere near a tensor engine output window.
|
|
317
|
+
static inline void __attribute__((always_inline)) merge_rescale_add_asm(float * acc,
|
|
318
|
+
const float * peer_acc,
|
|
319
|
+
int64_t dv,
|
|
320
|
+
float alpha_own,
|
|
321
|
+
float alpha_peer) {
|
|
322
|
+
unsigned long old_mask;
|
|
323
|
+
__asm__ volatile(
|
|
324
|
+
"mova.x.m %[ms] \n\t"
|
|
325
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
326
|
+
"fbc.ps f4, 0(%[ao]) \n\t" // broadcast alpha_own
|
|
327
|
+
"fbc.ps f5, 0(%[ap]) \n\t" // broadcast alpha_peer
|
|
328
|
+
: [ms] "=&r"(old_mask)
|
|
329
|
+
: [ao] "r"(&alpha_own), [ap] "r"(&alpha_peer)
|
|
330
|
+
: "f4", "f5");
|
|
331
|
+
|
|
332
|
+
for (int64_t d = 0; d < dv; d += 8) {
|
|
333
|
+
__asm__ volatile(
|
|
334
|
+
"flw.ps f2, 0(%[a]) \n\t" // own
|
|
335
|
+
"flw.ps f3, 0(%[p]) \n\t" // peer
|
|
336
|
+
"fmul.ps f2, f2, f4 \n\t" // own *= alpha_own
|
|
337
|
+
"fmul.ps f3, f3, f5 \n\t" // peer *= alpha_peer
|
|
338
|
+
"fadd.ps f2, f2, f3 \n\t"
|
|
339
|
+
"fsw.ps f2, 0(%[a]) \n\t"
|
|
340
|
+
:
|
|
341
|
+
: [a] "r"(acc + d), [p] "r"(peer_acc + d)
|
|
342
|
+
: "f2", "f3", "memory");
|
|
343
|
+
}
|
|
344
|
+
|
|
345
|
+
__asm__ volatile("mova.m.x %0" ::"r"(old_mask));
|
|
346
|
+
}
|
|
347
|
+
|
|
348
|
+
int entry_point(struct ggml_et_flash_attn_ext_params * params, void * env) {
|
|
349
|
+
(void) env;
|
|
350
|
+
|
|
351
|
+
uint64_t hart_id = get_hart_id();
|
|
352
|
+
uint64_t shire_id = get_shire_id();
|
|
353
|
+
|
|
354
|
+
if (shire_id >= NUM_COMPUTE_SHIRES) {
|
|
355
|
+
return 0;
|
|
356
|
+
}
|
|
357
|
+
|
|
358
|
+
const int is_hart1 = hart_id & 1;
|
|
359
|
+
uint64_t local_minion = (hart_id >> 1) & 0x1F;
|
|
360
|
+
|
|
361
|
+
struct ggml_tensor * q = ¶ms->src0;
|
|
362
|
+
struct ggml_tensor * k = ¶ms->src1;
|
|
363
|
+
struct ggml_tensor * v = ¶ms->src2;
|
|
364
|
+
struct ggml_tensor * dst = ¶ms->dst;
|
|
365
|
+
const int32_t has_mask = params->has_mask;
|
|
366
|
+
struct ggml_tensor * mask = has_mask ? ¶ms->mask : (struct ggml_tensor *) 0;
|
|
367
|
+
|
|
368
|
+
const char * q_data = (const char *) q->data;
|
|
369
|
+
const char * k_data = (const char *) k->data;
|
|
370
|
+
const char * v_data = (const char *) v->data;
|
|
371
|
+
char * dst_data = (char *) dst->data;
|
|
372
|
+
|
|
373
|
+
// et_barrier(ET_BARRIER_GLOBAL);
|
|
374
|
+
evict_region_past_l2(q->data, tensor_bytes_fa(q));
|
|
375
|
+
evict_region_past_l2(k->data, tensor_bytes_fa(k));
|
|
376
|
+
evict_region_past_l2(v->data, tensor_bytes_fa(v));
|
|
377
|
+
if (mask) {
|
|
378
|
+
evict_region_past_l2(mask->data, tensor_bytes_fa(mask));
|
|
379
|
+
}
|
|
380
|
+
et_barrier(ET_BARRIER_GLOBAL);
|
|
381
|
+
|
|
382
|
+
const int64_t dk = q->ne[0];
|
|
383
|
+
const int64_t nq = q->ne[1];
|
|
384
|
+
const int64_t nhq = q->ne[2];
|
|
385
|
+
const int64_t no = q->ne[3];
|
|
386
|
+
const int64_t nk = k->ne[1];
|
|
387
|
+
const int64_t nhk = k->ne[2];
|
|
388
|
+
const int64_t dv = v->ne[0];
|
|
389
|
+
|
|
390
|
+
if (dv > FA_DV_MAX || dk > FA_DK_MAX) {
|
|
391
|
+
return -1;
|
|
392
|
+
}
|
|
393
|
+
if (k->nb[0] != 2 || v->nb[0] != 2) {
|
|
394
|
+
return -1;
|
|
395
|
+
}
|
|
396
|
+
if ((dk % 8) != 0 || (dv % 16) != 0) {
|
|
397
|
+
return -1;
|
|
398
|
+
}
|
|
399
|
+
|
|
400
|
+
const int64_t gqa_ratio = nhq / nhk;
|
|
401
|
+
const int64_t total_rows = nq * nhq * no;
|
|
402
|
+
const float scale = params->scale;
|
|
403
|
+
const int use_fast_store = (dv % 16 == 0);
|
|
404
|
+
|
|
405
|
+
// Split-KV team layout (mirrors mul_mat_f16_matrix_engine.c)
|
|
406
|
+
//
|
|
407
|
+
// When total_rows is small compared to the total minion count (typical
|
|
408
|
+
// for decode: nq=1, nhq small), we group k_splits minions within the
|
|
409
|
+
// same shire into a team that cooperates on one row by splitting the
|
|
410
|
+
// KV dimension. Each team member computes a partial (M_p, S_p, acc_p)
|
|
411
|
+
// over its KV slab; the k_split==0 member merges the partials with the
|
|
412
|
+
// softmax combine rule.
|
|
413
|
+
//
|
|
414
|
+
// k_splits is a power of two, capped at MINIONS_PER_SHIRE (so a team
|
|
415
|
+
// never spans shires — L2 SCP is shire-local) and at nk_tiles (so each
|
|
416
|
+
// team member gets at least one KV tile).
|
|
417
|
+
const int64_t nk_tiles = (nk + TILE_KV - 1) / TILE_KV;
|
|
418
|
+
const int64_t total_minions = 2 * NUM_COMPUTE_SHIRES * MINIONS_PER_SHIRE;
|
|
419
|
+
int64_t k_splits = 1;
|
|
420
|
+
if (total_rows < total_minions) {
|
|
421
|
+
int64_t target = total_minions / total_rows;
|
|
422
|
+
int64_t ks = 1;
|
|
423
|
+
while (ks * 2 <= target && ks * 2 <= MINIONS_PER_SHIRE && ks * 2 <= nk_tiles) {
|
|
424
|
+
ks *= 2;
|
|
425
|
+
}
|
|
426
|
+
k_splits = ks;
|
|
427
|
+
}
|
|
428
|
+
|
|
429
|
+
const int64_t tiles_per_shire = MINIONS_PER_SHIRE / k_splits;
|
|
430
|
+
const int64_t k_split = (int64_t) local_minion % k_splits;
|
|
431
|
+
const int64_t local_tile_idx = (int64_t) local_minion / k_splits;
|
|
432
|
+
const int64_t tiles_stride = (int64_t) NUM_COMPUTE_SHIRES * tiles_per_shire;
|
|
433
|
+
|
|
434
|
+
// KV slab for this k_split. With k_splits=1 this is the full range.
|
|
435
|
+
const int64_t tiles_per_split_rounded = (nk_tiles + k_splits - 1) / k_splits;
|
|
436
|
+
const int64_t tile_start = k_split * tiles_per_split_rounded;
|
|
437
|
+
int64_t tile_end = tile_start + tiles_per_split_rounded;
|
|
438
|
+
if (tile_end > nk_tiles) {
|
|
439
|
+
tile_end = nk_tiles;
|
|
440
|
+
}
|
|
441
|
+
const int64_t kv_start = tile_start * TILE_KV;
|
|
442
|
+
int64_t kv_end = tile_end * TILE_KV;
|
|
443
|
+
if (kv_end > nk) {
|
|
444
|
+
kv_end = nk;
|
|
445
|
+
}
|
|
446
|
+
|
|
447
|
+
// L2 SCP pointers for this minion
|
|
448
|
+
uint64_t scp_base = local_minion * SCP_PER_MINION;
|
|
449
|
+
et_fp16_t * scp_kp[2] = {
|
|
450
|
+
(et_fp16_t *) et_shire_l2scp_local(scp_base + SCP_KP0_OFF),
|
|
451
|
+
(et_fp16_t *) et_shire_l2scp_local(scp_base + SCP_KP1_OFF),
|
|
452
|
+
};
|
|
453
|
+
|
|
454
|
+
// Hart 1 does K-panel packing
|
|
455
|
+
//
|
|
456
|
+
// When k_splits > 1, hart 1 must also participate in the two shire
|
|
457
|
+
// barriers that bracket the merge phase (one before and one after, so
|
|
458
|
+
// the reducer can read peer partials safely and the writers know when
|
|
459
|
+
// their acc/stats slab is free to reuse). Hart 1 has no useful work
|
|
460
|
+
// between those barriers.
|
|
461
|
+
//
|
|
462
|
+
// All teams in a shire must iterate the same number of times so the
|
|
463
|
+
// per-iter shire barriers stay balanced. Teams whose assigned row is
|
|
464
|
+
// past total_rows still call the barriers but skip the packing work.
|
|
465
|
+
et_barrier(ET_BARRIER_SHIRE);
|
|
466
|
+
// et_barrier(ET_BARRIER_GLOBAL);
|
|
467
|
+
if (is_hart1) {
|
|
468
|
+
uint32_t chunk_id = 0;
|
|
469
|
+
const int64_t row_base = (int64_t) shire_id + local_tile_idx * NUM_COMPUTE_SHIRES;
|
|
470
|
+
|
|
471
|
+
int64_t max_iters;
|
|
472
|
+
if (k_splits > 1) {
|
|
473
|
+
max_iters = (total_rows + tiles_stride - 1) / tiles_stride;
|
|
474
|
+
} else {
|
|
475
|
+
max_iters = (row_base >= total_rows) ? 0 : ((total_rows - row_base - 1) / tiles_stride + 1);
|
|
476
|
+
}
|
|
477
|
+
|
|
478
|
+
for (int64_t iter = 0; iter < max_iters; iter++) {
|
|
479
|
+
const int64_t row = row_base + iter * tiles_stride;
|
|
480
|
+
const int has_work = (row < total_rows);
|
|
481
|
+
|
|
482
|
+
if (has_work) {
|
|
483
|
+
const int64_t iq3 = row / (nhq * nq);
|
|
484
|
+
const int64_t rem = row % (nhq * nq);
|
|
485
|
+
const int64_t iq2 = rem / nq;
|
|
486
|
+
const int64_t ik2 = iq2 / gqa_ratio;
|
|
487
|
+
|
|
488
|
+
const char * k_head = k_data + ik2 * k->nb[2] + iq3 * k->nb[3];
|
|
489
|
+
|
|
490
|
+
for (int64_t kv_base = kv_start; kv_base < kv_end; kv_base += TILE_KV) {
|
|
491
|
+
const int64_t kv_count = (kv_base + TILE_KV <= nk) ? TILE_KV : (nk - kv_base);
|
|
492
|
+
|
|
493
|
+
for (int64_t dk_chunk = 0; dk_chunk < dk; dk_chunk += TILE_K) {
|
|
494
|
+
int buf = chunk_id & 1;
|
|
495
|
+
|
|
496
|
+
// Back-pressure: before overwriting buf[buf] on chunk N
|
|
497
|
+
// (which will displace chunk N-2), wait for hart 0 to
|
|
498
|
+
// post that it's done with chunk N-2. Gates both
|
|
499
|
+
// directions of double-buffering.
|
|
500
|
+
//
|
|
501
|
+
// NOTE: we use et_sem_* (FCC 0 only) rather than
|
|
502
|
+
// et_barrier(ET_BARRIER_MINION) here because the
|
|
503
|
+
// minion barrier for minion 0 shares FLB 0 with
|
|
504
|
+
// ET_BARRIER_SHIRE. Mixing them deadlocks. See
|
|
505
|
+
// feedback_flb_collision.
|
|
506
|
+
if (chunk_id >= 2) {
|
|
507
|
+
et_sem_wait(ET_BARRIER_MINION);
|
|
508
|
+
}
|
|
509
|
+
|
|
510
|
+
// Prefetch K data for this chunk
|
|
511
|
+
prefetch_kv_to_l2(k_head, kv_base, dk_chunk, kv_count, k->nb[1]);
|
|
512
|
+
|
|
513
|
+
pack_k_for_transpose16(scp_kp[buf], k_head, kv_base, dk_chunk, kv_count, k->nb[1]);
|
|
514
|
+
|
|
515
|
+
FENCE;
|
|
516
|
+
flush_to_l2(scp_kp[buf], 16, 64);
|
|
517
|
+
flush_to_l2((et_fp16_t *) ((char *) scp_kp[buf] + 1024), 16, 64);
|
|
518
|
+
WAIT_CACHEOPS;
|
|
519
|
+
|
|
520
|
+
// Signal: this buf is ready for hart 0 to consume.
|
|
521
|
+
et_sem_post(ET_BARRIER_MINION);
|
|
522
|
+
|
|
523
|
+
chunk_id++;
|
|
524
|
+
}
|
|
525
|
+
}
|
|
526
|
+
}
|
|
527
|
+
|
|
528
|
+
// Shire barriers for split-KV merge (hart 1 is a passive arrival).
|
|
529
|
+
if (k_splits > 1) {
|
|
530
|
+
et_barrier(ET_BARRIER_SHIRE); // A: team has written its partial
|
|
531
|
+
et_barrier(ET_BARRIER_SHIRE); // B: reducer has finished merge
|
|
532
|
+
}
|
|
533
|
+
}
|
|
534
|
+
|
|
535
|
+
// Self-drain phantom FCC 0 credits left by the wait-skip on the
|
|
536
|
+
// first 2 chunks. Hart 1 issued chunk_id posts but only
|
|
537
|
+
// (chunk_id - 2) waits (when chunk_id >= 2), so hart 1's FCC 0
|
|
538
|
+
// carries +min(chunk_id,2) credits from hart 0's matching posts
|
|
539
|
+
// that hart 1 never consumed.
|
|
540
|
+
uint32_t drain = (chunk_id < 2) ? chunk_id : 2;
|
|
541
|
+
for (uint32_t d = 0; d < drain; d++) {
|
|
542
|
+
et_sem_wait(ET_BARRIER_MINION);
|
|
543
|
+
}
|
|
544
|
+
|
|
545
|
+
// FENCE;
|
|
546
|
+
// et_barrier(ET_BARRIER_GLOBAL);
|
|
547
|
+
return 0;
|
|
548
|
+
}
|
|
549
|
+
|
|
550
|
+
// Hart 0: tensor engine compute
|
|
551
|
+
#ifndef UBERKERNEL_SUPPRESS_SCP_SETUP
|
|
552
|
+
setup_cache_scp();
|
|
553
|
+
#endif
|
|
554
|
+
CLEAR_TENSOR_ERROR;
|
|
555
|
+
|
|
556
|
+
// Q converted to F16 (one row at a time)
|
|
557
|
+
et_fp16_t q_f16[FA_DK_MAX] __attribute__((aligned(64)));
|
|
558
|
+
|
|
559
|
+
// Score buffer for QK^T output (16 scores per KV tile)
|
|
560
|
+
float scores[TILE_KV] __attribute__((aligned(64)));
|
|
561
|
+
|
|
562
|
+
// Small buffers for V accumulation
|
|
563
|
+
et_fp16_t w_f16_buf[32] __attribute__((aligned(64))); // 64 bytes
|
|
564
|
+
et_fp16_t vpanel_buf[8 * 32] __attribute__((aligned(64))); // 512 bytes
|
|
565
|
+
|
|
566
|
+
float * acc = (float *) et_shire_l2scp_local(scp_base + SCP_ACC_OFF);
|
|
567
|
+
|
|
568
|
+
uint32_t chunk_id = 0;
|
|
569
|
+
|
|
570
|
+
// Iter-based outer loop (matches hart 1). When k_splits > 1 all teams
|
|
571
|
+
// in a shire iterate the same number of times so the per-row shire
|
|
572
|
+
// barriers stay balanced; iterations with row >= total_rows skip the
|
|
573
|
+
// compute but still participate in the barriers.
|
|
574
|
+
const int64_t hart0_row_base = (int64_t) shire_id + local_tile_idx * NUM_COMPUTE_SHIRES;
|
|
575
|
+
int64_t hart0_max_iters;
|
|
576
|
+
if (k_splits > 1) {
|
|
577
|
+
hart0_max_iters = (total_rows + tiles_stride - 1) / tiles_stride;
|
|
578
|
+
} else {
|
|
579
|
+
hart0_max_iters = (hart0_row_base >= total_rows) ? 0 : ((total_rows - hart0_row_base - 1) / tiles_stride + 1);
|
|
580
|
+
}
|
|
581
|
+
|
|
582
|
+
for (int64_t iter = 0; iter < hart0_max_iters; iter++) {
|
|
583
|
+
const int64_t row = hart0_row_base + iter * tiles_stride;
|
|
584
|
+
if (row >= total_rows) {
|
|
585
|
+
// No-work iteration: only participate in barriers (k_splits > 1).
|
|
586
|
+
if (k_splits > 1) {
|
|
587
|
+
et_barrier(ET_BARRIER_SHIRE); // A
|
|
588
|
+
et_barrier(ET_BARRIER_SHIRE); // B
|
|
589
|
+
}
|
|
590
|
+
continue;
|
|
591
|
+
}
|
|
592
|
+
|
|
593
|
+
const int64_t iq3 = row / (nhq * nq);
|
|
594
|
+
const int64_t rem = row % (nhq * nq);
|
|
595
|
+
const int64_t iq2 = rem / nq;
|
|
596
|
+
const int64_t iq1 = rem % nq;
|
|
597
|
+
const int64_t ik2 = iq2 / gqa_ratio;
|
|
598
|
+
|
|
599
|
+
// Read Q row (F32) and convert to F16
|
|
600
|
+
const float * pq = (const float *) (q_data + iq1 * q->nb[1] + iq2 * q->nb[2] + iq3 * q->nb[3]);
|
|
601
|
+
convert_q_row_f32_to_f16(q_f16, pq, dk);
|
|
602
|
+
|
|
603
|
+
// V base for this head + batch (K packing handled by hart 1)
|
|
604
|
+
const char * v_head = v_data + ik2 * v->nb[2] + iq3 * v->nb[3];
|
|
605
|
+
|
|
606
|
+
// Output pointer
|
|
607
|
+
float * out = (float *) (dst_data + iq2 * dst->nb[1] + iq1 * dst->nb[2] + iq3 * dst->nb[3]);
|
|
608
|
+
|
|
609
|
+
zero_acc_vec(acc, dv);
|
|
610
|
+
float M = ET_NEG_INF_F;
|
|
611
|
+
float S = 0.0f;
|
|
612
|
+
const char * mask_base = has_mask ? get_mask_row_base(mask, iq1, iq2, iq3) : (const char *) 0;
|
|
613
|
+
|
|
614
|
+
// Flush Q_f16 to L2 so tensor_load can see it
|
|
615
|
+
FENCE;
|
|
616
|
+
flush_to_l2(q_f16, (dk * 2 + 63) / 64, 64);
|
|
617
|
+
WAIT_CACHEOPS;
|
|
618
|
+
|
|
619
|
+
for (int64_t kv_base = kv_start; kv_base < kv_end; kv_base += TILE_KV) {
|
|
620
|
+
const int64_t kv_count = (kv_base + TILE_KV <= nk) ? TILE_KV : (nk - kv_base);
|
|
621
|
+
|
|
622
|
+
// Set tensor_mask for partial tiles
|
|
623
|
+
if (kv_count < TILE_KV) {
|
|
624
|
+
uint64_t tmask = (1ULL << kv_count) - 1;
|
|
625
|
+
__asm__ __volatile__("csrw 0x805, %0" : : "r"(tmask));
|
|
626
|
+
}
|
|
627
|
+
|
|
628
|
+
// ============================================================
|
|
629
|
+
// QK^T via TensorFMA16A32
|
|
630
|
+
// ============================================================
|
|
631
|
+
|
|
632
|
+
// Pipelined QK^T:
|
|
633
|
+
// - Q for the whole row is preloaded once into A_L1[0..n-1].
|
|
634
|
+
// Each FMA picks its chunk via scp_loc_a = chunk_idx.
|
|
635
|
+
// - K is double-buffered in L1: K_BUFS[0]=lines 16..31,
|
|
636
|
+
// K_BUFS[1]=lines 32..47.
|
|
637
|
+
// - In iteration i (1..N-1), the K[i] load runs concurrently
|
|
638
|
+
// with the FMA on chunk i-1: they touch disjoint L1 regions
|
|
639
|
+
// (FMA reads K_BUFS[(i-1)&1], load writes K_BUFS[i&1]; FMA
|
|
640
|
+
// reads A_L1[i-1], load doesn't touch A_L1).
|
|
641
|
+
//
|
|
642
|
+
// L1 footprint: max dk=512 → Q uses 16 lines (0..15), K uses 32
|
|
643
|
+
// lines (16..47). Within ET-SoC-1 L1 SCP (≥128 lines per minion).
|
|
644
|
+
const int64_t n_dk_chunks = dk / TILE_K;
|
|
645
|
+
const uint64_t K_BUFS[2] = {
|
|
646
|
+
(uint64_t) B_L1_START, // 16..31
|
|
647
|
+
(uint64_t) (B_L1_START + 16), // 32..47
|
|
648
|
+
};
|
|
649
|
+
|
|
650
|
+
// Preload entire Q row into A_L1[0..n_dk_chunks-1] (one tensor_load,
|
|
651
|
+
// one wait, regardless of dk).
|
|
652
|
+
tensor_load(false, false, A_L1_START, TENSOR_LOAD_PLAIN, 0, (uint64_t) q_f16, 0,
|
|
653
|
+
(uint64_t) (n_dk_chunks - 1), 64, 0);
|
|
654
|
+
|
|
655
|
+
// Prologue: wait hart 1's K[0], issue K[0] load, wait both loads.
|
|
656
|
+
{
|
|
657
|
+
int buf = chunk_id & 1;
|
|
658
|
+
et_sem_wait(ET_BARRIER_MINION);
|
|
659
|
+
tensor_load(false, false, K_BUFS[0], TENSOR_LOAD_TRANSPOSE16, 0, (uint64_t) scp_kp[buf], 0, 15, 64, 1);
|
|
660
|
+
tensor_wait(TENSOR_LOAD_WAIT_0); // Q row complete
|
|
661
|
+
tensor_wait(TENSOR_LOAD_WAIT_1); // K[0] complete
|
|
662
|
+
et_sem_post(ET_BARRIER_MINION);
|
|
663
|
+
chunk_id++;
|
|
664
|
+
}
|
|
665
|
+
|
|
666
|
+
// Main loop: in iter i, issue K[i] load and FMA chunk i-1 in
|
|
667
|
+
// parallel. The matrix engine is busy on FMA[i-1] while the
|
|
668
|
+
// load unit fetches K[i] from L2 SCP.
|
|
669
|
+
//
|
|
670
|
+
// Order of waits matters: wait K[i] load first, then sem_post
|
|
671
|
+
// immediately (frees scp_kp[buf] for hart 1 to refill chunk i+2),
|
|
672
|
+
// then wait FMA. Putting sem_post after FMA wait would stall
|
|
673
|
+
// hart 1 by a full FMA latency — defeating the producer pipeline.
|
|
674
|
+
for (int64_t i = 1; i < n_dk_chunks; i++) {
|
|
675
|
+
int buf = chunk_id & 1;
|
|
676
|
+
int k_slot_prev = (int) ((i - 1) & 1);
|
|
677
|
+
int k_slot = (int) (i & 1);
|
|
678
|
+
|
|
679
|
+
et_sem_wait(ET_BARRIER_MINION);
|
|
680
|
+
tensor_load(false, false, K_BUFS[k_slot], TENSOR_LOAD_TRANSPOSE16, 0, (uint64_t) scp_kp[buf], 0, 15, 64,
|
|
681
|
+
1);
|
|
682
|
+
|
|
683
|
+
tensor_fma((kv_count < TILE_KV), 3, 0, 15, 0, false, false, false, false, K_BUFS[k_slot_prev],
|
|
684
|
+
(uint64_t) (i - 1), TENSOR_FMA_OP_FP16, (i == 1));
|
|
685
|
+
|
|
686
|
+
tensor_wait(TENSOR_LOAD_WAIT_1); // K[i] in L1
|
|
687
|
+
et_sem_post(ET_BARRIER_MINION); // release scp_kp[buf] EARLY
|
|
688
|
+
tensor_wait(TENSOR_FMA_WAIT); // then wait FMA[i-1]
|
|
689
|
+
chunk_id++;
|
|
690
|
+
}
|
|
691
|
+
|
|
692
|
+
// Epilogue: FMA on the last chunk (no overlapping load).
|
|
693
|
+
{
|
|
694
|
+
int k_slot_last = (int) ((n_dk_chunks - 1) & 1);
|
|
695
|
+
tensor_fma((kv_count < TILE_KV), 3, 0, 15, 0, false, false, false, false, K_BUFS[k_slot_last],
|
|
696
|
+
(uint64_t) (n_dk_chunks - 1), TENSOR_FMA_OP_FP16, (n_dk_chunks == 1));
|
|
697
|
+
tensor_wait(TENSOR_FMA_WAIT);
|
|
698
|
+
}
|
|
699
|
+
|
|
700
|
+
// Prefetch V rows for this tile.
|
|
701
|
+
// Only useful for the partial-tile path below
|
|
702
|
+
if (kv_count < TILE_KV) {
|
|
703
|
+
for (int64_t d = 0; d < dv; d += 32) {
|
|
704
|
+
prefetch_kv_to_l2(v_head, kv_base, d, kv_count, v->nb[1]);
|
|
705
|
+
}
|
|
706
|
+
}
|
|
707
|
+
|
|
708
|
+
// Extract QK^T scores from vector register file
|
|
709
|
+
__asm__ volatile("" ::: "f0", "f1");
|
|
710
|
+
{
|
|
711
|
+
unsigned long _ms;
|
|
712
|
+
__asm__ volatile(
|
|
713
|
+
"mova.x.m %[ms] \n\t"
|
|
714
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
715
|
+
"fbc.ps f2, 0(%[p_scale]) \n\t"
|
|
716
|
+
"fmul.ps f0, f0, f2 \n\t"
|
|
717
|
+
"fmul.ps f1, f1, f2 \n\t"
|
|
718
|
+
"fsw.ps f0, 0(%[dst]) \n\t"
|
|
719
|
+
"fsw.ps f1, 32(%[dst]) \n\t"
|
|
720
|
+
"mova.m.x %[ms] \n\t"
|
|
721
|
+
: [ms] "=&r"(_ms)
|
|
722
|
+
: [dst] "r"(scores), [p_scale] "r"(&scale)
|
|
723
|
+
: "f0", "f1", "f2", "memory");
|
|
724
|
+
}
|
|
725
|
+
|
|
726
|
+
// ============================================================
|
|
727
|
+
// Two-phase softmax + V accumulation
|
|
728
|
+
// ============================================================
|
|
729
|
+
|
|
730
|
+
float weights[TILE_KV] __attribute__((aligned(64)));
|
|
731
|
+
{
|
|
732
|
+
// A1: apply mask to scores, pad unused slots
|
|
733
|
+
for (int64_t j = 0; j < kv_count; ++j) {
|
|
734
|
+
if (has_mask) {
|
|
735
|
+
float mv = get_mask_val_from_base(mask, mask_base, kv_base + j);
|
|
736
|
+
if (mv == ET_NEG_INF_F || mv != mv) {
|
|
737
|
+
scores[j] = ET_NEG_INF_F;
|
|
738
|
+
} else {
|
|
739
|
+
scores[j] += mv;
|
|
740
|
+
}
|
|
741
|
+
}
|
|
742
|
+
}
|
|
743
|
+
for (int64_t j = kv_count; j < TILE_KV; ++j) {
|
|
744
|
+
scores[j] = ET_NEG_INF_F;
|
|
745
|
+
}
|
|
746
|
+
|
|
747
|
+
// A1b: SIMD horizontal max across all 16 scores
|
|
748
|
+
float tile_max;
|
|
749
|
+
{
|
|
750
|
+
unsigned long _ms;
|
|
751
|
+
__asm__ volatile(
|
|
752
|
+
"mova.x.m %[ms] \n\t"
|
|
753
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
754
|
+
"flw.ps f2, 0(%[sc]) \n\t"
|
|
755
|
+
"flw.ps f3, 32(%[sc]) \n\t"
|
|
756
|
+
"fmax.ps f2, f2, f3 \n\t"
|
|
757
|
+
"fswizz.ps f3, f2, 0xB1 \n\t"
|
|
758
|
+
"fmax.ps f2, f2, f3 \n\t"
|
|
759
|
+
"fswizz.ps f3, f2, 0x4E \n\t"
|
|
760
|
+
"fmax.ps f2, f2, f3 \n\t"
|
|
761
|
+
"fmvz.x.ps t0, f2, 4 \n\t"
|
|
762
|
+
"fbcx.ps f3, t0 \n\t"
|
|
763
|
+
"fmax.ps %[tm], f2, f3 \n\t"
|
|
764
|
+
"mova.m.x %[ms] \n\t"
|
|
765
|
+
: [ms] "=&r"(_ms), [tm] "=f"(tile_max)
|
|
766
|
+
: [sc] "r"(scores)
|
|
767
|
+
: "f2", "f3", "t0", "memory");
|
|
768
|
+
}
|
|
769
|
+
|
|
770
|
+
if (tile_max > ET_NEG_INF_F) {
|
|
771
|
+
// A2: rescale accumulator if this tile has a new global max
|
|
772
|
+
if (tile_max > M) {
|
|
773
|
+
float rescale = et_exp2f((M - tile_max) * 1.4426950408889634f);
|
|
774
|
+
scale_acc_vec(acc, dv, rescale);
|
|
775
|
+
S *= rescale;
|
|
776
|
+
M = tile_max;
|
|
777
|
+
}
|
|
778
|
+
|
|
779
|
+
// A3: SIMD exp2 + horizontal sum
|
|
780
|
+
// Interleaved: f2/f3 chains alternate to hide ALU latency.
|
|
781
|
+
// fexp.ps has multi-cycle latency — the two independent
|
|
782
|
+
// exp2 calls naturally pipeline.
|
|
783
|
+
{
|
|
784
|
+
const float log2e = 1.4426950408889634f;
|
|
785
|
+
float S_tile;
|
|
786
|
+
unsigned long _ms;
|
|
787
|
+
__asm__ volatile(
|
|
788
|
+
"mova.x.m %[ms] \n\t"
|
|
789
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
790
|
+
"flw.ps f2, 0(%[sc]) \n\t"
|
|
791
|
+
"fbc.ps f4, 0(%[pM]) \n\t"
|
|
792
|
+
"flw.ps f3, 32(%[sc]) \n\t"
|
|
793
|
+
"fbc.ps f5, 0(%[pL]) \n\t"
|
|
794
|
+
"fsub.ps f2, f2, f4 \n\t"
|
|
795
|
+
"fsub.ps f3, f3, f4 \n\t"
|
|
796
|
+
"fmul.ps f2, f2, f5 \n\t"
|
|
797
|
+
"fmul.ps f3, f3, f5 \n\t"
|
|
798
|
+
"fexp.ps f2, f2 \n\t"
|
|
799
|
+
"fexp.ps f3, f3 \n\t"
|
|
800
|
+
"fsw.ps f2, 0(%[wt]) \n\t"
|
|
801
|
+
"fsw.ps f3, 32(%[wt]) \n\t"
|
|
802
|
+
"fadd.ps f2, f2, f3, rne \n\t"
|
|
803
|
+
"fswizz.ps f3, f2, 0xB1 \n\t"
|
|
804
|
+
"fadd.ps f2, f2, f3, rne \n\t"
|
|
805
|
+
"fswizz.ps f3, f2, 0x4E \n\t"
|
|
806
|
+
"fadd.ps f2, f2, f3, rne \n\t"
|
|
807
|
+
"fmvz.x.ps t0, f2, 4 \n\t"
|
|
808
|
+
"fbcx.ps f3, t0 \n\t"
|
|
809
|
+
"fadd.ps %[st], f2, f3, rne \n\t"
|
|
810
|
+
"mova.m.x %[ms] \n\t"
|
|
811
|
+
: [ms] "=&r"(_ms), [st] "=f"(S_tile)
|
|
812
|
+
: [pM] "r"(&M), [pL] "r"(&log2e), [sc] "r"(scores), [wt] "r"(weights)
|
|
813
|
+
: "f2", "f3", "f4", "f5", "t0", "memory");
|
|
814
|
+
S += S_tile;
|
|
815
|
+
}
|
|
816
|
+
|
|
817
|
+
// Phase B: weights @ V via TensorFMA16A32
|
|
818
|
+
{
|
|
819
|
+
// B1: convert weights F32 → F16
|
|
820
|
+
convert_q_row_f32_to_f16(w_f16_buf, weights, TILE_KV);
|
|
821
|
+
|
|
822
|
+
FENCE;
|
|
823
|
+
flush_to_l2(w_f16_buf, 1, 64);
|
|
824
|
+
WAIT_CACHEOPS;
|
|
825
|
+
|
|
826
|
+
// Issue weights load (wait_id=0) and the first V chunk
|
|
827
|
+
// load (wait_id=1) concurrently. Weights comes from
|
|
828
|
+
// L2 SCP (just flushed); V[0] comes from DRAM via
|
|
829
|
+
// INTERLEAVE16 — running them in parallel hides the
|
|
830
|
+
// shorter load behind the longer one. For partial
|
|
831
|
+
// tiles, V is software-packed below — we only kick
|
|
832
|
+
// off the early V load on the full-tile fast path.
|
|
833
|
+
tensor_load(false, false, A_L1_START, TENSOR_LOAD_PLAIN, 0, (uint64_t) w_f16_buf, 0, 0, 64, 0);
|
|
834
|
+
|
|
835
|
+
const int v_full_tile = (kv_count == TILE_KV);
|
|
836
|
+
const uintptr_t v_base = (uintptr_t) v_head + kv_base * v->nb[1];
|
|
837
|
+
const uint64_t nb1_v = (uint64_t) v->nb[1];
|
|
838
|
+
uint64_t b_cur = 8;
|
|
839
|
+
|
|
840
|
+
if (v_full_tile) {
|
|
841
|
+
tensor_load(false, false, b_cur, TENSOR_LOAD_INTERLEAVE16, 0, (uint64_t) v_base, 0, 7,
|
|
842
|
+
nb1_v, 1);
|
|
843
|
+
}
|
|
844
|
+
|
|
845
|
+
tensor_wait(TENSOR_LOAD_WAIT_0); // weights in A_L1
|
|
846
|
+
if (v_full_tile) {
|
|
847
|
+
tensor_wait(TENSOR_LOAD_WAIT_1); // V[0] in b_cur
|
|
848
|
+
}
|
|
849
|
+
|
|
850
|
+
// B2: process dv in chunks of 16
|
|
851
|
+
if (v_full_tile) {
|
|
852
|
+
for (int64_t dv_off = 0; dv_off < dv; dv_off += 16) {
|
|
853
|
+
const uint64_t b_nxt = b_cur ^ 24;
|
|
854
|
+
|
|
855
|
+
if (dv_off + 16 < dv) {
|
|
856
|
+
tensor_load(false, false, b_nxt, TENSOR_LOAD_INTERLEAVE16, 0,
|
|
857
|
+
(uint64_t) (v_base + (dv_off + 16) * 2), 0, 7, nb1_v, 1);
|
|
858
|
+
}
|
|
859
|
+
|
|
860
|
+
tensor_fma(false, 3, 0, 7, 0, false, false, false, false, b_cur, A_L1_START,
|
|
861
|
+
TENSOR_FMA_OP_FP16, true);
|
|
862
|
+
tensor_wait(TENSOR_FMA_WAIT);
|
|
863
|
+
|
|
864
|
+
__asm__ volatile("" ::: "f0", "f1");
|
|
865
|
+
{
|
|
866
|
+
unsigned long _ms;
|
|
867
|
+
__asm__ volatile(
|
|
868
|
+
"mova.x.m %[ms] \n\t"
|
|
869
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
870
|
+
"flw.ps f2, 0(%[pa]) \n\t"
|
|
871
|
+
"flw.ps f3, 32(%[pa]) \n\t"
|
|
872
|
+
"fadd.ps f0, f0, f2 \n\t"
|
|
873
|
+
"fadd.ps f1, f1, f3 \n\t"
|
|
874
|
+
"fsw.ps f0, 0(%[pa]) \n\t"
|
|
875
|
+
"fsw.ps f1, 32(%[pa]) \n\t"
|
|
876
|
+
"mova.m.x %[ms] \n\t"
|
|
877
|
+
: [ms] "=&r"(_ms)
|
|
878
|
+
: [pa] "r"(acc + dv_off)
|
|
879
|
+
: "f0", "f1", "f2", "f3", "memory");
|
|
880
|
+
}
|
|
881
|
+
|
|
882
|
+
if (dv_off + 16 < dv) {
|
|
883
|
+
tensor_wait(TENSOR_LOAD_WAIT_1);
|
|
884
|
+
b_cur = b_nxt;
|
|
885
|
+
}
|
|
886
|
+
}
|
|
887
|
+
} else {
|
|
888
|
+
// Partial tile: software pack, no pipeline
|
|
889
|
+
for (int64_t dv_off = 0; dv_off < dv; dv_off += 16) {
|
|
890
|
+
pack_v_interleaved(vpanel_buf, v_head, kv_base, dv_off, kv_count, v->nb[1]);
|
|
891
|
+
FENCE;
|
|
892
|
+
flush_to_l2(vpanel_buf, 8, 64);
|
|
893
|
+
WAIT_CACHEOPS;
|
|
894
|
+
tensor_load(false, false, B_L1_START, TENSOR_LOAD_PLAIN, 0, (uint64_t) vpanel_buf, 0, 7,
|
|
895
|
+
64, 0);
|
|
896
|
+
tensor_wait(TENSOR_LOAD_WAIT_0);
|
|
897
|
+
|
|
898
|
+
tensor_fma(false, 3, 0, 7, 0, false, false, false, false, B_L1_START, A_L1_START,
|
|
899
|
+
TENSOR_FMA_OP_FP16, true);
|
|
900
|
+
tensor_wait(TENSOR_FMA_WAIT);
|
|
901
|
+
|
|
902
|
+
__asm__ volatile("" ::: "f0", "f1");
|
|
903
|
+
{
|
|
904
|
+
unsigned long _ms;
|
|
905
|
+
__asm__ volatile(
|
|
906
|
+
"mova.x.m %[ms] \n\t"
|
|
907
|
+
"mov.m.x m0, x0, 0xFF \n\t"
|
|
908
|
+
"flw.ps f2, 0(%[pa]) \n\t"
|
|
909
|
+
"flw.ps f3, 32(%[pa]) \n\t"
|
|
910
|
+
"fadd.ps f0, f0, f2 \n\t"
|
|
911
|
+
"fadd.ps f1, f1, f3 \n\t"
|
|
912
|
+
"fsw.ps f0, 0(%[pa]) \n\t"
|
|
913
|
+
"fsw.ps f1, 32(%[pa]) \n\t"
|
|
914
|
+
"mova.m.x %[ms] \n\t"
|
|
915
|
+
: [ms] "=&r"(_ms)
|
|
916
|
+
: [pa] "r"(acc + dv_off)
|
|
917
|
+
: "f0", "f1", "f2", "f3", "memory");
|
|
918
|
+
}
|
|
919
|
+
}
|
|
920
|
+
}
|
|
921
|
+
}
|
|
922
|
+
}
|
|
923
|
+
}
|
|
924
|
+
}
|
|
925
|
+
|
|
926
|
+
// Finalize row
|
|
927
|
+
//
|
|
928
|
+
// k_splits == 1: this minion computed the full row. Normalize in
|
|
929
|
+
// place and store to DRAM.
|
|
930
|
+
//
|
|
931
|
+
// k_splits > 1: this minion computed a KV slab. Publish the
|
|
932
|
+
// partial (M, S, acc) to L2 SCP, sync with the
|
|
933
|
+
// team, and let the k_split==0 member do the
|
|
934
|
+
// softmax combine and the final store. All tensor
|
|
935
|
+
// engine ops are complete before this block, so
|
|
936
|
+
// f0..f31 are free to use.
|
|
937
|
+
if (k_splits > 1) {
|
|
938
|
+
// Publish our partial.
|
|
939
|
+
volatile float * my_stats = (volatile float *) et_shire_l2scp_local(scp_base + SCP_STATS_OFF);
|
|
940
|
+
my_stats[0] = M;
|
|
941
|
+
my_stats[1] = S;
|
|
942
|
+
FENCE;
|
|
943
|
+
evict_range_to_l2(acc, (int64_t) dv * (int64_t) sizeof(float));
|
|
944
|
+
evict_to_l2((const void *) my_stats, 1, 64);
|
|
945
|
+
WAIT_CACHEOPS;
|
|
946
|
+
|
|
947
|
+
// A: team members have all written their partials.
|
|
948
|
+
et_barrier(ET_BARRIER_SHIRE);
|
|
949
|
+
|
|
950
|
+
if (k_split == 0) {
|
|
951
|
+
// Online softmax merge: fold peers 1..k_splits-1 into our
|
|
952
|
+
// own (M_running, S_running, acc). For each peer p:
|
|
953
|
+
// M_new = max(M_running, M_p)
|
|
954
|
+
// α_own = exp2((M_running - M_new) * log2e)
|
|
955
|
+
// α_p = exp2((M_p - M_new) * log2e)
|
|
956
|
+
// acc[d] = α_own * acc[d] + α_p * peer_acc[d]
|
|
957
|
+
// S_running = α_own * S_running + α_p * S_p
|
|
958
|
+
float M_running = M;
|
|
959
|
+
float S_running = S;
|
|
960
|
+
const float log2e = 1.4426950408889634f;
|
|
961
|
+
|
|
962
|
+
for (int64_t p = 1; p < k_splits; p++) {
|
|
963
|
+
uint64_t peer_scp = (local_tile_idx * k_splits + p) * SCP_PER_MINION;
|
|
964
|
+
volatile float * peer_stats = (volatile float *) et_shire_l2scp_local(peer_scp + SCP_STATS_OFF);
|
|
965
|
+
float * peer_acc = (float *) et_shire_l2scp_local(peer_scp + SCP_ACC_OFF);
|
|
966
|
+
|
|
967
|
+
// Drop stale L1D copies before reading peer's data.
|
|
968
|
+
evict_to_l2((const void *) peer_stats, 1, 64);
|
|
969
|
+
evict_range_to_l2(peer_acc, (int64_t) dv * (int64_t) sizeof(float));
|
|
970
|
+
WAIT_CACHEOPS;
|
|
971
|
+
|
|
972
|
+
const float M_p = peer_stats[0];
|
|
973
|
+
const float S_p = peer_stats[1];
|
|
974
|
+
|
|
975
|
+
const float M_new = (M_p > M_running) ? M_p : M_running;
|
|
976
|
+
const float alpha_own = (M_running == ET_NEG_INF_F) ? 0.0f : et_exp2f((M_running - M_new) * log2e);
|
|
977
|
+
const float alpha_p = (M_p == ET_NEG_INF_F) ? 0.0f : et_exp2f((M_p - M_new) * log2e);
|
|
978
|
+
|
|
979
|
+
merge_rescale_add_asm(acc, peer_acc, dv, alpha_own, alpha_p);
|
|
980
|
+
|
|
981
|
+
S_running = alpha_own * S_running + alpha_p * S_p;
|
|
982
|
+
M_running = M_new;
|
|
983
|
+
}
|
|
984
|
+
|
|
985
|
+
const float S_inv = (S_running == 0.0f) ? 0.0f : et_fdiv(1.0f, S_running);
|
|
986
|
+
normalize_store_vec(out, acc, dv, S_inv, use_fast_store);
|
|
987
|
+
}
|
|
988
|
+
|
|
989
|
+
// B: reducer is done, team may reuse its acc/stats slabs.
|
|
990
|
+
et_barrier(ET_BARRIER_SHIRE);
|
|
991
|
+
} else {
|
|
992
|
+
// k_splits == 1 fast path — this minion owns the full row.
|
|
993
|
+
const float S_inv = S == 0.0f ? 0.0f : et_fdiv(1.0f, S);
|
|
994
|
+
normalize_store_vec(out, acc, dv, S_inv, use_fast_store);
|
|
995
|
+
}
|
|
996
|
+
}
|
|
997
|
+
|
|
998
|
+
FENCE;
|
|
999
|
+
return 0;
|
|
1000
|
+
}
|