whispercpp 1.3.5 → 1.3.6
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/LICENSE +1 -1
- data/README.md +99 -2
- data/ext/extconf.rb +1 -0
- data/ext/ruby_whisper.c +20 -4
- data/ext/ruby_whisper.h +30 -2
- data/ext/ruby_whisper_context.c +216 -124
- data/ext/ruby_whisper_context_params.c +163 -0
- data/ext/ruby_whisper_model.c +0 -1
- data/ext/ruby_whisper_params.c +0 -1
- data/ext/ruby_whisper_segment.c +0 -1
- data/ext/ruby_whisper_token.c +29 -9
- data/ext/ruby_whisper_transcribe.cpp +4 -1
- data/ext/ruby_whisper_vad_context.c +48 -1
- data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
- data/ext/ruby_whisper_vad_params.c +0 -1
- data/ext/ruby_whisper_vad_segment.c +0 -1
- data/ext/ruby_whisper_vad_segments.c +0 -1
- data/ext/sources/CMakeLists.txt +1 -1
- data/ext/sources/bindings/javascript/package.json +1 -1
- data/ext/sources/cmake/whisper-config.cmake.in +5 -40
- data/ext/sources/examples/bench/bench.cpp +23 -18
- data/ext/sources/examples/cli/cli.cpp +8 -0
- data/ext/sources/examples/common-ggml.cpp +2 -0
- data/ext/sources/examples/miniaudio.h +4507 -2131
- data/ext/sources/examples/server/server.cpp +18 -4
- data/ext/sources/examples/talk-llama/CMakeLists.txt +3 -2
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +7 -13
- data/ext/sources/examples/talk-llama/llama-adapter.h +4 -3
- data/ext/sources/examples/talk-llama/llama-arch.cpp +335 -17
- data/ext/sources/examples/talk-llama/llama-arch.h +42 -0
- data/ext/sources/examples/talk-llama/llama-batch.cpp +3 -1
- data/ext/sources/examples/talk-llama/llama-chat.cpp +21 -1
- data/ext/sources/examples/talk-llama/llama-chat.h +1 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +508 -520
- data/ext/sources/examples/talk-llama/llama-context.h +27 -28
- data/ext/sources/examples/talk-llama/llama-cparams.h +5 -0
- data/ext/sources/examples/talk-llama/llama-ext.h +12 -0
- data/ext/sources/examples/talk-llama/llama-grammar.cpp +8 -8
- data/ext/sources/examples/talk-llama/llama-graph.cpp +583 -130
- data/ext/sources/examples/talk-llama/llama-graph.h +131 -10
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +57 -40
- data/ext/sources/examples/talk-llama/llama-hparams.h +79 -10
- data/ext/sources/examples/talk-llama/llama-impl.cpp +4 -4
- data/ext/sources/examples/talk-llama/llama-impl.h +13 -1
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +3 -1
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +274 -89
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +2 -3
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +275 -0
- data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +140 -0
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +11 -13
- data/ext/sources/examples/talk-llama/llama-mmap.cpp +28 -11
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +527 -119
- data/ext/sources/examples/talk-llama/llama-model-loader.h +35 -5
- data/ext/sources/examples/talk-llama/llama-model-saver.cpp +60 -46
- data/ext/sources/examples/talk-llama/llama-model-saver.h +5 -2
- data/ext/sources/examples/talk-llama/llama-model.cpp +1365 -647
- data/ext/sources/examples/talk-llama/llama-model.h +72 -19
- data/ext/sources/examples/talk-llama/llama-quant.cpp +578 -346
- data/ext/sources/examples/talk-llama/{llama-sampling.cpp → llama-sampler.cpp} +190 -76
- data/ext/sources/examples/talk-llama/{llama-sampling.h → llama-sampler.h} +0 -2
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +118 -48
- data/ext/sources/examples/talk-llama/llama-vocab.h +5 -0
- data/ext/sources/examples/talk-llama/llama.cpp +76 -22
- data/ext/sources/examples/talk-llama/llama.h +63 -30
- data/ext/sources/examples/talk-llama/models/afmoe.cpp +2 -3
- data/ext/sources/examples/talk-llama/models/apertus.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/arcee.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/arctic.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/baichuan.cpp +4 -3
- data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +1 -2
- data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +3 -5
- data/ext/sources/examples/talk-llama/models/bert.cpp +13 -7
- data/ext/sources/examples/talk-llama/models/bitnet.cpp +9 -24
- data/ext/sources/examples/talk-llama/models/bloom.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/chameleon.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/chatglm.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/codeshell.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/cogvlm.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/command-r.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/dbrx.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/deci.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/deepseek.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/deepseek2.cpp +24 -21
- data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +445 -0
- data/ext/sources/examples/talk-llama/models/dots1.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/dream.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +4 -6
- data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/eurobert.cpp +97 -0
- data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +145 -0
- data/ext/sources/examples/talk-llama/models/exaone.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/exaone4.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +2 -4
- data/ext/sources/examples/talk-llama/models/falcon.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma3.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +7 -7
- data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/glm4.cpp +14 -7
- data/ext/sources/examples/talk-llama/models/gpt2.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/gptneox.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/granite.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/grok.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/grovemoe.cpp +5 -7
- data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/internlm2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/jais.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/jais2.cpp +123 -0
- data/ext/sources/examples/talk-llama/models/jamba.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +381 -0
- data/ext/sources/examples/talk-llama/models/lfm2.cpp +145 -124
- data/ext/sources/examples/talk-llama/models/llada-moe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/llada.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/llama.cpp +18 -11
- data/ext/sources/examples/talk-llama/models/maincoder.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/{graph-context-mamba.cpp → mamba-base.cpp} +9 -3
- data/ext/sources/examples/talk-llama/models/mamba.cpp +1 -2
- data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +11 -5
- data/ext/sources/examples/talk-llama/models/minicpm3.cpp +14 -13
- data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +4 -5
- data/ext/sources/examples/talk-llama/models/mistral3.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/models.h +181 -46
- data/ext/sources/examples/talk-llama/models/modern-bert.cpp +2 -9
- data/ext/sources/examples/talk-llama/models/mpt.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +26 -14
- data/ext/sources/examples/talk-llama/models/nemotron.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/neo-bert.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/olmo.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/olmo2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/olmoe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +1 -1
- data/ext/sources/examples/talk-llama/models/openelm.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/orion.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/paddleocr.cpp +122 -0
- data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/phi2.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/phi3.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/plamo.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/plamo2.cpp +9 -5
- data/ext/sources/examples/talk-llama/models/plamo3.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/plm.cpp +15 -14
- data/ext/sources/examples/talk-llama/models/qwen.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/qwen2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/qwen3.cpp +12 -9
- data/ext/sources/examples/talk-llama/models/qwen35.cpp +381 -0
- data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +422 -0
- data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +15 -8
- data/ext/sources/examples/talk-llama/models/qwen3next.cpp +84 -432
- data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +9 -18
- data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +8 -17
- data/ext/sources/examples/talk-llama/models/refact.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/rnd1.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +2 -0
- data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +2 -0
- data/ext/sources/examples/talk-llama/models/seed-oss.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/smallthinker.cpp +4 -4
- data/ext/sources/examples/talk-llama/models/smollm3.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/stablelm.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/starcoder.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/starcoder2.cpp +3 -3
- data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +165 -0
- data/ext/sources/examples/talk-llama/models/t5-dec.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/t5-enc.cpp +2 -2
- data/ext/sources/examples/talk-llama/models/xverse.cpp +3 -3
- data/ext/sources/examples/talk-llama/unicode.cpp +21 -65
- data/ext/sources/ggml/CMakeLists.txt +9 -3
- data/ext/sources/ggml/include/ggml-backend.h +1 -1
- data/ext/sources/ggml/include/ggml-cann.h +1 -1
- data/ext/sources/ggml/include/ggml-cpu.h +5 -0
- data/ext/sources/ggml/include/ggml-openvino.h +37 -0
- data/ext/sources/ggml/include/ggml-opt.h +1 -1
- data/ext/sources/ggml/include/ggml-rpc.h +6 -1
- data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
- data/ext/sources/ggml/include/ggml.h +56 -9
- data/ext/sources/ggml/src/CMakeLists.txt +3 -0
- data/ext/sources/ggml/src/ggml-alloc.c +4 -9
- data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
- data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +28 -86
- data/ext/sources/ggml/src/ggml-backend.cpp +5 -2
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +6 -2
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +348 -189
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +40 -85
- data/ext/sources/ggml/src/ggml-cann/common.h +3 -4
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +44 -62
- data/ext/sources/ggml/src/ggml-common.h +11 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +16 -11
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +42 -19
- data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
- data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +85 -85
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +85 -1
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2744 -548
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +1653 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1391 -0
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -10
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +9 -9
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +118 -18
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +107 -26
- data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
- data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +59 -12
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +15 -0
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +965 -252
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +584 -197
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +903 -188
- data/ext/sources/ggml/src/ggml-cpu/ops.h +1 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.c +40 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2890 -679
- data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
- data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +136 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +111 -3
- data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +17 -0
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-cuda/argsort.cu +19 -10
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +32 -30
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +134 -18
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +41 -27
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +6 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +78 -64
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +384 -143
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +36 -22
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +3 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +26 -5
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +127 -12
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +263 -0
- data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +595 -200
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +9 -8
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +173 -6
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +158 -85
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +34 -22
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +127 -67
- data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +157 -65
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +18 -76
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +2 -16
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +233 -133
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
- data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +56 -32
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +3 -3
- data/ext/sources/ggml/src/ggml-cuda/top-k.cu +0 -1
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +199 -135
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +55 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +31 -17
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +10 -0
- data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +82 -45
- data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +334 -160
- data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +7 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +328 -197
- data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +765 -234
- data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +252 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +412 -265
- data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +23 -23
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +1 -1
- data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.h → hex-dma.h} +28 -3
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +1 -1
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +27 -37
- data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +6 -35
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +240 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +251 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +20 -1347
- data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +211 -13
- data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +1119 -952
- data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +254 -244
- data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +36 -36
- data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +155 -138
- data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +339 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +209 -114
- data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
- data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
- data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
- data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +6 -0
- data/ext/sources/ggml/src/ggml-impl.h +62 -0
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +274 -73
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +22 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +102 -36
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +174 -23
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +580 -280
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +5 -4
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +320 -107
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1068 -825
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +19 -1
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +3108 -636
- data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +41 -99
- data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +45 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +204 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
- data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
- data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
- data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
- data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +975 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +373 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1110 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +884 -0
- data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +123 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +46 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +39 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +123 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +293 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +226 -0
- data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +85 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.cpp +823 -0
- data/ext/sources/ggml/src/ggml-openvino/utils.h +123 -0
- data/ext/sources/ggml/src/ggml-quants.c +96 -5
- data/ext/sources/ggml/src/ggml-quants.h +3 -0
- data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +15 -88
- data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +1 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
- data/ext/sources/ggml/src/ggml-sycl/common.hpp +315 -10
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +69 -1
- data/ext/sources/ggml/src/ggml-sycl/convert.hpp +22 -1
- data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +78 -68
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1179 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +55 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1338 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +667 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +225 -0
- data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +309 -0
- data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +316 -51
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
- data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +1 -1
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
- data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +7 -0
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +13 -0
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
- data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
- data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
- data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
- data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
- data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
- data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
- data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
- data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1250 -465
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +374 -170
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +66 -22
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +389 -201
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +106 -58
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +128 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +8 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +3 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +36 -63
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +10 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +16 -10
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +55 -35
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1314 -109
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1660 -1371
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +141 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +65 -72
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +75 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +6 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +40 -5
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +105 -60
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +53 -259
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat.tmpl.wgsl → mul_mat.wgsl} +68 -257
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +692 -23
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_reg_tile.tmpl.wgsl → mul_mat_reg_tile.wgsl} +28 -128
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +31 -137
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +480 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +9 -36
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +40 -12
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +6 -1
- data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
- data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +9 -6
- data/ext/sources/ggml/src/ggml.c +167 -33
- data/ext/sources/ggml/src/gguf.cpp +229 -44
- data/ext/sources/src/whisper.cpp +6 -28
- data/sig/whisper.rbs +43 -2
- data/test/test_context_params.rb +82 -0
- data/test/test_token.rb +11 -0
- data/test/test_vad_context.rb +58 -8
- data/test/test_whisper.rb +20 -0
- data/whispercpp.gemspec +1 -1
- metadata +240 -28
- data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
- data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
- data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
- data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
|
@@ -1,221 +0,0 @@
|
|
|
1
|
-
#ifndef HTP_UTILS_H
|
|
2
|
-
#define HTP_UTILS_H
|
|
3
|
-
|
|
4
|
-
#ifdef __cplusplus
|
|
5
|
-
extern "C" {
|
|
6
|
-
#endif
|
|
7
|
-
|
|
8
|
-
#include <AEEStdErr.h>
|
|
9
|
-
#include <inttypes.h>
|
|
10
|
-
#include <remote.h>
|
|
11
|
-
#include <rpcmem.h>
|
|
12
|
-
#include <stdbool.h>
|
|
13
|
-
|
|
14
|
-
/* Offset to differentiate HLOS and Hexagon error codes.
|
|
15
|
-
Stores the value of AEE_EOFFSET for Hexagon. */
|
|
16
|
-
#ifndef DSP_OFFSET
|
|
17
|
-
# define DSP_OFFSET 0x80000400
|
|
18
|
-
#endif
|
|
19
|
-
|
|
20
|
-
/* Errno for connection reset by peer. */
|
|
21
|
-
#ifndef ECONNRESET
|
|
22
|
-
# ifdef __hexagon__
|
|
23
|
-
# define ECONNRESET 104
|
|
24
|
-
# endif
|
|
25
|
-
#endif
|
|
26
|
-
|
|
27
|
-
/* Abstraction of different OS specific sleep APIs.
|
|
28
|
-
SLEEP accepts input in seconds. */
|
|
29
|
-
#ifndef SLEEP
|
|
30
|
-
# ifdef __hexagon__
|
|
31
|
-
# define SLEEP(x) \
|
|
32
|
-
{ /* Do nothing for simulator. */ \
|
|
33
|
-
}
|
|
34
|
-
# else
|
|
35
|
-
# ifdef _WINDOWS
|
|
36
|
-
# define SLEEP(x) Sleep(1000 * x) /* Sleep accepts input in milliseconds. */
|
|
37
|
-
# else
|
|
38
|
-
# define SLEEP(x) sleep(x) /* sleep accepts input in seconds. */
|
|
39
|
-
# endif
|
|
40
|
-
# endif
|
|
41
|
-
#endif
|
|
42
|
-
|
|
43
|
-
/* Include windows specific header files. */
|
|
44
|
-
#ifdef _WINDOWS
|
|
45
|
-
# include <sysinfoapi.h>
|
|
46
|
-
# include <windows.h>
|
|
47
|
-
# define _CRT_SECURE_NO_WARNINGS 1
|
|
48
|
-
# define _WINSOCK_DEPRECATED_NO_WARNINGS 1
|
|
49
|
-
/* Including this file for custom implementation of getopt function. */
|
|
50
|
-
# include "getopt_custom.h"
|
|
51
|
-
#endif
|
|
52
|
-
|
|
53
|
-
/* Includes and defines for all HLOS except windows */
|
|
54
|
-
#if !defined(__hexagon__) && !defined(_WINDOWS)
|
|
55
|
-
# include "unistd.h"
|
|
56
|
-
|
|
57
|
-
# include <sys/time.h>
|
|
58
|
-
#endif
|
|
59
|
-
|
|
60
|
-
/* Includes and defines for Hexagon and all HLOS except Windows. */
|
|
61
|
-
#if !defined(_WINDOWS)
|
|
62
|
-
/* Weak reference to remote symbol for compilation. */
|
|
63
|
-
# pragma weak remote_session_control
|
|
64
|
-
# pragma weak remote_handle_control
|
|
65
|
-
# pragma weak remote_handle64_control
|
|
66
|
-
# pragma weak fastrpc_mmap
|
|
67
|
-
# pragma weak fastrpc_munmap
|
|
68
|
-
# pragma weak rpcmem_alloc2
|
|
69
|
-
#endif
|
|
70
|
-
|
|
71
|
-
#if !defined(_WINDOWS)
|
|
72
|
-
# pragma weak remote_system_request
|
|
73
|
-
#endif
|
|
74
|
-
/**
|
|
75
|
-
* Wrapper for FastRPC Capability API: query DSP support.
|
|
76
|
-
*
|
|
77
|
-
* @param[out] domain pointer to supported domain.
|
|
78
|
-
* @return 0 if query is successful.
|
|
79
|
-
* non-zero if error, return value points to the error.
|
|
80
|
-
*/
|
|
81
|
-
int get_dsp_support(int * domain);
|
|
82
|
-
|
|
83
|
-
/**
|
|
84
|
-
* Wrapper for FastRPC Capability API: query VTCM information.
|
|
85
|
-
*
|
|
86
|
-
* @param[in] domain value of domain in the queried.
|
|
87
|
-
* @param[out] capability capability value of the attribute queried.
|
|
88
|
-
* @param[in] attr value of the attribute to the queried.
|
|
89
|
-
* @return 0 if query is successful.
|
|
90
|
-
* non-zero if error, return value points to the error.
|
|
91
|
-
*/
|
|
92
|
-
int get_vtcm_info(int domain, uint32_t * capability, uint32_t attr);
|
|
93
|
-
|
|
94
|
-
/**
|
|
95
|
-
* Wrapper for FastRPC Capability API: query unsigned pd support on CDSP domain.
|
|
96
|
-
*
|
|
97
|
-
* @return true if unsigned pd is supported.
|
|
98
|
-
* false if unsigned pd is not supported, capability query failed.
|
|
99
|
-
*/
|
|
100
|
-
|
|
101
|
-
bool get_unsignedpd_support(void);
|
|
102
|
-
|
|
103
|
-
/**
|
|
104
|
-
* Wrapper for FastRPC Capability API: query unsigned pd support.
|
|
105
|
-
*
|
|
106
|
-
* @param[in] domain value of domain in the queried.
|
|
107
|
-
* @return true if unsigned pd is supported.
|
|
108
|
-
* false if unsigned pd is not supported, capability query failed.
|
|
109
|
-
*/
|
|
110
|
-
|
|
111
|
-
bool is_unsignedpd_supported(int domain_id);
|
|
112
|
-
|
|
113
|
-
/**
|
|
114
|
-
* is_valid_domain_id API: query a domain id is valid.
|
|
115
|
-
*
|
|
116
|
-
* @param[in] domain value of domain in the queried.
|
|
117
|
-
* @param[in] compute_only value of domain is only compared with CDSP domains supported by the target when enabled.
|
|
118
|
-
* @return true if value of domain is valid.
|
|
119
|
-
* false if value of domain is not valid.
|
|
120
|
-
*/
|
|
121
|
-
|
|
122
|
-
bool is_valid_domain_id(int domain_id, int compute_only);
|
|
123
|
-
|
|
124
|
-
/**
|
|
125
|
-
* get_domain API: get domain struct from domain value.
|
|
126
|
-
*
|
|
127
|
-
* @param[in] domain value of a domain
|
|
128
|
-
* @return Returns domain struct of the domain if it is supported or else
|
|
129
|
-
* returns NULL.
|
|
130
|
-
*
|
|
131
|
-
*/
|
|
132
|
-
|
|
133
|
-
domain * get_domain(int domain_id);
|
|
134
|
-
|
|
135
|
-
/**
|
|
136
|
-
* get_domains_info API: get information for all the domains available on the device
|
|
137
|
-
*
|
|
138
|
-
* @param[in] domain_type pointer to domain type
|
|
139
|
-
* @param[in] num_domains pointer to number of domains
|
|
140
|
-
* @param[in] domains_info pointer to save discovered domains information.
|
|
141
|
-
* @return 0 if query is successful.
|
|
142
|
-
* non-zero if error, return value points to the error.
|
|
143
|
-
*
|
|
144
|
-
* It is user's responsibility to free the memory used to store the domains info whose address is present in domains_info before closing the application.
|
|
145
|
-
*
|
|
146
|
-
*/
|
|
147
|
-
|
|
148
|
-
int get_domains_info(char * domain_type, int * num_domains, fastrpc_domain ** domains_info);
|
|
149
|
-
|
|
150
|
-
/**
|
|
151
|
-
* get_effective_domain_id API: get effective domain id for given session id
|
|
152
|
-
*
|
|
153
|
-
* @param[in] domain_name pointer to domain name
|
|
154
|
-
* @param[in] session_id
|
|
155
|
-
* @param[in] effec_domain_id pointer to save obtained effective domain id.
|
|
156
|
-
* @return 0 if query is successful.
|
|
157
|
-
* non-zero if error, return value points to the error.
|
|
158
|
-
*
|
|
159
|
-
*/
|
|
160
|
-
|
|
161
|
-
int get_effective_domain_id(char * domain_name, int session_id, int * effec_domain_id);
|
|
162
|
-
|
|
163
|
-
/**
|
|
164
|
-
* is_async_fastrpc_supported API: query a domain id has async fastrpc supported or not
|
|
165
|
-
*
|
|
166
|
-
* @param[in] domain_id value of a domain
|
|
167
|
-
* @return Returns true or false stating support of Async FastRPC
|
|
168
|
-
*
|
|
169
|
-
*/
|
|
170
|
-
|
|
171
|
-
bool is_async_fastrpc_supported(int domain_id);
|
|
172
|
-
|
|
173
|
-
/**
|
|
174
|
-
* is_status_notification_supported API: query the DSP for STATUS_NOTIFICATION_SUPPORT information
|
|
175
|
-
*
|
|
176
|
-
* @param[in] domain_id value of a domain
|
|
177
|
-
* @return Returns true or false stating status notification support information
|
|
178
|
-
*
|
|
179
|
-
*/
|
|
180
|
-
bool is_status_notification_supported(int domain_id);
|
|
181
|
-
|
|
182
|
-
/**
|
|
183
|
-
* get_hmx_support_info API: query the DSP for HMX SUPPORT information
|
|
184
|
-
*
|
|
185
|
-
* @param[in] domain_id value of a domain
|
|
186
|
-
* @param[out] capability capability value of the attribute queried.
|
|
187
|
-
* @param[in] attr value of the attribute to the queried.
|
|
188
|
-
* @return 0 if query is successful.
|
|
189
|
-
* non-zero if error, return value points to the error.
|
|
190
|
-
*
|
|
191
|
-
*/
|
|
192
|
-
int get_hmx_support_info(int domain, uint32_t * capability, uint32_t attr);
|
|
193
|
-
|
|
194
|
-
/**
|
|
195
|
-
* get_hex_arch_ver API: query the Hexagon processor architecture version information
|
|
196
|
-
*
|
|
197
|
-
* @param[in] domain_id value of a domain
|
|
198
|
-
* @param[out] Arch version (73, 75, ...)
|
|
199
|
-
* @return 0 if query is successful.
|
|
200
|
-
* non-zero if error, return value points to the error.
|
|
201
|
-
*
|
|
202
|
-
*/
|
|
203
|
-
int get_hex_arch_ver(int domain, int * arch);
|
|
204
|
-
|
|
205
|
-
/**
|
|
206
|
-
* get_hvx_support_info API: query the DSP for HVX SUPPORT information
|
|
207
|
-
*
|
|
208
|
-
* @param[in] domain_id value of a domain
|
|
209
|
-
* @param[out] capability capability value of the attribute queried.
|
|
210
|
-
* @param[in] attr value of the attribute to the queried.
|
|
211
|
-
* @return 0 if query is successful.
|
|
212
|
-
* non-zero if error, return value points to the error.
|
|
213
|
-
*
|
|
214
|
-
*/
|
|
215
|
-
int get_hvx_support_info(int domain, uint32_t * capability, uint32_t attr);
|
|
216
|
-
|
|
217
|
-
#ifdef __cplusplus
|
|
218
|
-
}
|
|
219
|
-
#endif
|
|
220
|
-
|
|
221
|
-
#endif //DSP_CAPABILITIES_UTILS_H
|
|
@@ -1,188 +0,0 @@
|
|
|
1
|
-
#define(VARIANTS)
|
|
2
|
-
|
|
3
|
-
[
|
|
4
|
-
{
|
|
5
|
-
"SHADER_NAME": "add_f32",
|
|
6
|
-
"REPLS": {
|
|
7
|
-
"TYPE" : "f32",
|
|
8
|
-
"OP": "+"
|
|
9
|
-
},
|
|
10
|
-
"DECLS": ["NOT_INPLACE"]
|
|
11
|
-
},
|
|
12
|
-
{
|
|
13
|
-
"SHADER_NAME": "add_f16",
|
|
14
|
-
"REPLS": {
|
|
15
|
-
"TYPE" : "f16",
|
|
16
|
-
"OP": "+"
|
|
17
|
-
},
|
|
18
|
-
"DECLS": ["NOT_INPLACE"]
|
|
19
|
-
},
|
|
20
|
-
{
|
|
21
|
-
"SHADER_NAME": "add_f32_inplace",
|
|
22
|
-
"REPLS": {
|
|
23
|
-
"TYPE" : "f32",
|
|
24
|
-
"OP": "+"
|
|
25
|
-
},
|
|
26
|
-
"DECLS": ["INPLACE"]
|
|
27
|
-
},
|
|
28
|
-
{
|
|
29
|
-
"SHADER_NAME": "add_f16_inplace",
|
|
30
|
-
"REPLS": {
|
|
31
|
-
"TYPE" : "f16",
|
|
32
|
-
"OP": "+"
|
|
33
|
-
},
|
|
34
|
-
"DECLS": ["INPLACE"]
|
|
35
|
-
},
|
|
36
|
-
{
|
|
37
|
-
"SHADER_NAME": "mul_f32",
|
|
38
|
-
"REPLS": {
|
|
39
|
-
"TYPE" : "f32",
|
|
40
|
-
"OP": "*"
|
|
41
|
-
},
|
|
42
|
-
"DECLS": ["NOT_INPLACE"]
|
|
43
|
-
},
|
|
44
|
-
{
|
|
45
|
-
"SHADER_NAME": "mul_f16",
|
|
46
|
-
"REPLS": {
|
|
47
|
-
"TYPE" : "f16",
|
|
48
|
-
"OP": "*"
|
|
49
|
-
},
|
|
50
|
-
"DECLS": ["NOT_INPLACE"]
|
|
51
|
-
},
|
|
52
|
-
{
|
|
53
|
-
"SHADER_NAME": "mul_f32_inplace",
|
|
54
|
-
"REPLS": {
|
|
55
|
-
"TYPE" : "f32",
|
|
56
|
-
"OP": "*"
|
|
57
|
-
},
|
|
58
|
-
"DECLS": ["INPLACE"]
|
|
59
|
-
},
|
|
60
|
-
{
|
|
61
|
-
"SHADER_NAME": "mul_f16_inplace",
|
|
62
|
-
"REPLS": {
|
|
63
|
-
"TYPE" : "f16",
|
|
64
|
-
"OP": "*"
|
|
65
|
-
},
|
|
66
|
-
"DECLS": ["INPLACE"]
|
|
67
|
-
},
|
|
68
|
-
{
|
|
69
|
-
"SHADER_NAME": "sub_f32",
|
|
70
|
-
"REPLS": {
|
|
71
|
-
"TYPE" : "f32",
|
|
72
|
-
"OP": "-"
|
|
73
|
-
},
|
|
74
|
-
"DECLS": ["NOT_INPLACE"]
|
|
75
|
-
},
|
|
76
|
-
{
|
|
77
|
-
"SHADER_NAME": "sub_f16",
|
|
78
|
-
"REPLS": {
|
|
79
|
-
"TYPE" : "f16",
|
|
80
|
-
"OP": "-"
|
|
81
|
-
},
|
|
82
|
-
"DECLS": ["NOT_INPLACE"]
|
|
83
|
-
},
|
|
84
|
-
{
|
|
85
|
-
"SHADER_NAME": "sub_f32_inplace",
|
|
86
|
-
"REPLS": {
|
|
87
|
-
"TYPE" : "f32",
|
|
88
|
-
"OP": "-"
|
|
89
|
-
},
|
|
90
|
-
"DECLS": ["INPLACE"]
|
|
91
|
-
},
|
|
92
|
-
{
|
|
93
|
-
"SHADER_NAME": "sub_f16_inplace",
|
|
94
|
-
"REPLS": {
|
|
95
|
-
"TYPE" : "f16",
|
|
96
|
-
"OP": "-"
|
|
97
|
-
},
|
|
98
|
-
"DECLS": ["INPLACE"]
|
|
99
|
-
},
|
|
100
|
-
{
|
|
101
|
-
"SHADER_NAME": "div_f32",
|
|
102
|
-
"REPLS": {
|
|
103
|
-
"TYPE" : "f32",
|
|
104
|
-
"OP": "/"
|
|
105
|
-
},
|
|
106
|
-
"DECLS": ["NOT_INPLACE"]
|
|
107
|
-
},
|
|
108
|
-
{
|
|
109
|
-
"SHADER_NAME": "div_f16",
|
|
110
|
-
"REPLS": {
|
|
111
|
-
"TYPE" : "f16",
|
|
112
|
-
"OP": "/"
|
|
113
|
-
},
|
|
114
|
-
"DECLS": ["NOT_INPLACE"]
|
|
115
|
-
},
|
|
116
|
-
{
|
|
117
|
-
"SHADER_NAME": "div_f32_inplace",
|
|
118
|
-
"REPLS": {
|
|
119
|
-
"TYPE" : "f32",
|
|
120
|
-
"OP": "/"
|
|
121
|
-
},
|
|
122
|
-
"DECLS": ["INPLACE"]
|
|
123
|
-
},
|
|
124
|
-
{
|
|
125
|
-
"SHADER_NAME": "div_f16_inplace",
|
|
126
|
-
"REPLS": {
|
|
127
|
-
"TYPE" : "f16",
|
|
128
|
-
"OP": "/"
|
|
129
|
-
},
|
|
130
|
-
"DECLS": ["INPLACE"]
|
|
131
|
-
}
|
|
132
|
-
]
|
|
133
|
-
|
|
134
|
-
#end(VARIANTS)
|
|
135
|
-
|
|
136
|
-
#define(DECLS)
|
|
137
|
-
|
|
138
|
-
#decl(NOT_INPLACE)
|
|
139
|
-
|
|
140
|
-
fn update(dst_i: u32, src0_i: u32, src1_i: u32) {
|
|
141
|
-
dst[dst_i] = src0[src0_i] {{OP}} src1[src1_i];
|
|
142
|
-
}
|
|
143
|
-
|
|
144
|
-
@group(0) @binding(2)
|
|
145
|
-
var<storage, read_write> dst: array<{{TYPE}}>;
|
|
146
|
-
|
|
147
|
-
@group(0) @binding(3)
|
|
148
|
-
var<uniform> params: Params;
|
|
149
|
-
|
|
150
|
-
#enddecl(NOT_INPLACE)
|
|
151
|
-
|
|
152
|
-
#decl(INPLACE)
|
|
153
|
-
|
|
154
|
-
fn update(dst_i: u32, src0_i: u32, src1_i: u32) {
|
|
155
|
-
src0[dst_i] = src0[src0_i] {{OP}} src1[src1_i];
|
|
156
|
-
}
|
|
157
|
-
|
|
158
|
-
@group(0) @binding(2)
|
|
159
|
-
var<uniform> params: Params;
|
|
160
|
-
|
|
161
|
-
#enddecl(INPLACE)
|
|
162
|
-
|
|
163
|
-
#end(DECLS)
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
#define(SHADER)
|
|
167
|
-
|
|
168
|
-
enable f16;
|
|
169
|
-
|
|
170
|
-
#include "binary_head.tmpl"
|
|
171
|
-
|
|
172
|
-
@group(0) @binding(0)
|
|
173
|
-
var<storage, read_write> src0: array<{{TYPE}}>;
|
|
174
|
-
|
|
175
|
-
@group(0) @binding(1)
|
|
176
|
-
var<storage, read_write> src1: array<{{TYPE}}>;
|
|
177
|
-
|
|
178
|
-
DECLS
|
|
179
|
-
|
|
180
|
-
override wg_size: u32;
|
|
181
|
-
@compute @workgroup_size(wg_size)
|
|
182
|
-
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
183
|
-
if (gid.x < params.ne) {
|
|
184
|
-
update(params.offset_dst + gid.x, params.offset_src0 + gid.x, params.offset_src1 + src1_index(gid.x));
|
|
185
|
-
}
|
|
186
|
-
}
|
|
187
|
-
|
|
188
|
-
#end(SHADER)
|
|
@@ -1,45 +0,0 @@
|
|
|
1
|
-
struct Params {
|
|
2
|
-
ne: u32,
|
|
3
|
-
|
|
4
|
-
// offsets in elements
|
|
5
|
-
offset_src0: u32,
|
|
6
|
-
offset_src1: u32,
|
|
7
|
-
offset_dst: u32,
|
|
8
|
-
|
|
9
|
-
stride_src1_0: u32,
|
|
10
|
-
stride_src1_1: u32,
|
|
11
|
-
stride_src1_2: u32,
|
|
12
|
-
stride_src1_3: u32,
|
|
13
|
-
|
|
14
|
-
a_ne0: u32,
|
|
15
|
-
a_ne1: u32,
|
|
16
|
-
a_ne2: u32,
|
|
17
|
-
|
|
18
|
-
b_ne0: u32,
|
|
19
|
-
b_ne1: u32,
|
|
20
|
-
b_ne2: u32,
|
|
21
|
-
b_ne3: u32,
|
|
22
|
-
};
|
|
23
|
-
|
|
24
|
-
fn src1_index(_i: u32) -> u32 {
|
|
25
|
-
var i = _i;
|
|
26
|
-
let a_i3 = i / (params.a_ne2 * params.a_ne1 * params.a_ne0);
|
|
27
|
-
i = i % (params.a_ne2 * params.a_ne1 * params.a_ne0);
|
|
28
|
-
let a_i2 = i / (params.a_ne1 * params.a_ne0);
|
|
29
|
-
i = i % (params.a_ne1 * params.a_ne0);
|
|
30
|
-
let a_i1 = i / params.a_ne0;
|
|
31
|
-
let a_i0 = i % params.a_ne0;
|
|
32
|
-
|
|
33
|
-
// handle repetition of b
|
|
34
|
-
// index loops back to the beginning and repeats after elements are exhausted = modulo
|
|
35
|
-
let b_i0 = a_i0 % params.b_ne0;
|
|
36
|
-
let b_i1 = a_i1 % params.b_ne1;
|
|
37
|
-
let b_i2 = a_i2 % params.b_ne2;
|
|
38
|
-
let b_i3 = a_i3 % params.b_ne3;
|
|
39
|
-
|
|
40
|
-
// compute index for position in b's flat array
|
|
41
|
-
return b_i0 * params.stride_src1_0 +
|
|
42
|
-
b_i1 * params.stride_src1_1 +
|
|
43
|
-
b_i2 * params.stride_src1_2 +
|
|
44
|
-
b_i3 * params.stride_src1_3;
|
|
45
|
-
}
|