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.
Files changed (1367) hide show
  1. checksums.yaml +7 -0
  2. data/LICENSE.txt +21 -0
  3. data/exe/elelem-ggml +8 -0
  4. data/ext/elelem/ggml/elelem.c +466 -0
  5. data/ext/elelem/ggml/extconf.rb +52 -0
  6. data/lib/elelem/ggml/chat_ml.rb +23 -0
  7. data/lib/elelem/ggml/gguf.rb +82 -0
  8. data/lib/elelem/ggml/provider.rb +63 -0
  9. data/lib/elelem/ggml/tokenizer.rb +95 -0
  10. data/lib/elelem/ggml/version.rb +7 -0
  11. data/lib/elelem/ggml.rb +37 -0
  12. data/vendor/ggml/AUTHORS +335 -0
  13. data/vendor/ggml/CMakeLists.txt +499 -0
  14. data/vendor/ggml/CONTRIBUTING.md +3 -0
  15. data/vendor/ggml/LICENSE +21 -0
  16. data/vendor/ggml/README.md +49 -0
  17. data/vendor/ggml/cmake/FindNCCL.cmake +36 -0
  18. data/vendor/ggml/cmake/GitVars.cmake +22 -0
  19. data/vendor/ggml/cmake/common.cmake +50 -0
  20. data/vendor/ggml/cmake/ggml-config.cmake.in +216 -0
  21. data/vendor/ggml/ggml.pc.in +10 -0
  22. data/vendor/ggml/include/ggml-alloc.h +86 -0
  23. data/vendor/ggml/include/ggml-backend.h +437 -0
  24. data/vendor/ggml/include/ggml-blas.h +25 -0
  25. data/vendor/ggml/include/ggml-cann.h +123 -0
  26. data/vendor/ggml/include/ggml-cpp.h +39 -0
  27. data/vendor/ggml/include/ggml-cpu.h +152 -0
  28. data/vendor/ggml/include/ggml-cuda.h +47 -0
  29. data/vendor/ggml/include/ggml-et.h +28 -0
  30. data/vendor/ggml/include/ggml-hexagon.h +19 -0
  31. data/vendor/ggml/include/ggml-metal.h +61 -0
  32. data/vendor/ggml/include/ggml-opencl.h +26 -0
  33. data/vendor/ggml/include/ggml-openvino.h +37 -0
  34. data/vendor/ggml/include/ggml-opt.h +256 -0
  35. data/vendor/ggml/include/ggml-rpc.h +35 -0
  36. data/vendor/ggml/include/ggml-sycl.h +57 -0
  37. data/vendor/ggml/include/ggml-virtgpu.h +14 -0
  38. data/vendor/ggml/include/ggml-vulkan.h +29 -0
  39. data/vendor/ggml/include/ggml-webgpu.h +19 -0
  40. data/vendor/ggml/include/ggml-zdnn.h +17 -0
  41. data/vendor/ggml/include/ggml-zendnn.h +22 -0
  42. data/vendor/ggml/include/ggml.h +2963 -0
  43. data/vendor/ggml/include/gguf.h +211 -0
  44. data/vendor/ggml/requirements.txt +12 -0
  45. data/vendor/ggml/scripts/gen-authors.sh +9 -0
  46. data/vendor/ggml/scripts/make-release-checks.sh +96 -0
  47. data/vendor/ggml/scripts/make-release-desc.sh +69 -0
  48. data/vendor/ggml/scripts/make-release-summary.txt +33 -0
  49. data/vendor/ggml/scripts/release.sh +195 -0
  50. data/vendor/ggml/scripts/sync-llama-am.sh +167 -0
  51. data/vendor/ggml/scripts/sync-llama.last +1 -0
  52. data/vendor/ggml/scripts/sync-llama.sh +21 -0
  53. data/vendor/ggml/scripts/sync-whisper-am.sh +138 -0
  54. data/vendor/ggml/scripts/sync-whisper.last +1 -0
  55. data/vendor/ggml/scripts/sync-whisper.sh +17 -0
  56. data/vendor/ggml/src/CMakeLists.txt +634 -0
  57. data/vendor/ggml/src/ggml-alloc.c +1249 -0
  58. data/vendor/ggml/src/ggml-backend-dl.cpp +48 -0
  59. data/vendor/ggml/src/ggml-backend-dl.h +45 -0
  60. data/vendor/ggml/src/ggml-backend-impl.h +291 -0
  61. data/vendor/ggml/src/ggml-backend-meta.cpp +2511 -0
  62. data/vendor/ggml/src/ggml-backend-reg.cpp +605 -0
  63. data/vendor/ggml/src/ggml-backend.cpp +2517 -0
  64. data/vendor/ggml/src/ggml-blas/CMakeLists.txt +101 -0
  65. data/vendor/ggml/src/ggml-blas/ggml-blas.cpp +530 -0
  66. data/vendor/ggml/src/ggml-cann/CMakeLists.txt +89 -0
  67. data/vendor/ggml/src/ggml-cann/acl_tensor.cpp +195 -0
  68. data/vendor/ggml/src/ggml-cann/acl_tensor.h +349 -0
  69. data/vendor/ggml/src/ggml-cann/aclnn_ops.cpp +4479 -0
  70. data/vendor/ggml/src/ggml-cann/aclnn_ops.h +1191 -0
  71. data/vendor/ggml/src/ggml-cann/common.h +651 -0
  72. data/vendor/ggml/src/ggml-cann/ggml-cann.cpp +3070 -0
  73. data/vendor/ggml/src/ggml-common.h +1911 -0
  74. data/vendor/ggml/src/ggml-cpu/CMakeLists.txt +685 -0
  75. data/vendor/ggml/src/ggml-cpu/amx/amx.cpp +249 -0
  76. data/vendor/ggml/src/ggml-cpu/amx/amx.h +8 -0
  77. data/vendor/ggml/src/ggml-cpu/amx/common.h +115 -0
  78. data/vendor/ggml/src/ggml-cpu/amx/mmq.cpp +2511 -0
  79. data/vendor/ggml/src/ggml-cpu/amx/mmq.h +10 -0
  80. data/vendor/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +41 -0
  81. data/vendor/ggml/src/ggml-cpu/arch/arm/quants.c +4319 -0
  82. data/vendor/ggml/src/ggml-cpu/arch/arm/repack.cpp +5156 -0
  83. data/vendor/ggml/src/ggml-cpu/arch/loongarch/quants.c +2309 -0
  84. data/vendor/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
  85. data/vendor/ggml/src/ggml-cpu/arch/powerpc/quants.c +2304 -0
  86. data/vendor/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
  87. data/vendor/ggml/src/ggml-cpu/arch/riscv/quants.c +6596 -0
  88. data/vendor/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1703 -0
  89. data/vendor/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
  90. data/vendor/ggml/src/ggml-cpu/arch/s390/quants.c +1465 -0
  91. data/vendor/ggml/src/ggml-cpu/arch/wasm/quants.c +1292 -0
  92. data/vendor/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
  93. data/vendor/ggml/src/ggml-cpu/arch/x86/quants.c +4108 -0
  94. data/vendor/ggml/src/ggml-cpu/arch/x86/repack.cpp +6407 -0
  95. data/vendor/ggml/src/ggml-cpu/arch-fallback.h +353 -0
  96. data/vendor/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
  97. data/vendor/ggml/src/ggml-cpu/binary-ops.h +16 -0
  98. data/vendor/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
  99. data/vendor/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  100. data/vendor/ggml/src/ggml-cpu/common.h +95 -0
  101. data/vendor/ggml/src/ggml-cpu/ggml-cpu-impl.h +539 -0
  102. data/vendor/ggml/src/ggml-cpu/ggml-cpu.c +3936 -0
  103. data/vendor/ggml/src/ggml-cpu/ggml-cpu.cpp +712 -0
  104. data/vendor/ggml/src/ggml-cpu/hbm.cpp +55 -0
  105. data/vendor/ggml/src/ggml-cpu/hbm.h +8 -0
  106. data/vendor/ggml/src/ggml-cpu/iqp.cpp +1253 -0
  107. data/vendor/ggml/src/ggml-cpu/iqp.h +39 -0
  108. data/vendor/ggml/src/ggml-cpu/kleidiai/CMakeLists.txt +14 -0
  109. data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.cpp +1101 -0
  110. data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.h +101 -0
  111. data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +1919 -0
  112. data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
  113. data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4164 -0
  114. data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
  115. data/vendor/ggml/src/ggml-cpu/ops.cpp +12164 -0
  116. data/vendor/ggml/src/ggml-cpu/ops.h +124 -0
  117. data/vendor/ggml/src/ggml-cpu/quants.c +1339 -0
  118. data/vendor/ggml/src/ggml-cpu/quants.h +106 -0
  119. data/vendor/ggml/src/ggml-cpu/repack.cpp +4836 -0
  120. data/vendor/ggml/src/ggml-cpu/repack.h +245 -0
  121. data/vendor/ggml/src/ggml-cpu/simd-gemm.h +226 -0
  122. data/vendor/ggml/src/ggml-cpu/simd-mappings.h +1319 -0
  123. data/vendor/ggml/src/ggml-cpu/spacemit/ime.cpp +1742 -0
  124. data/vendor/ggml/src/ggml-cpu/spacemit/ime.h +21 -0
  125. data/vendor/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +1027 -0
  126. data/vendor/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  127. data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  128. data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  129. data/vendor/ggml/src/ggml-cpu/spacemit/ime_kernels.h +189 -0
  130. data/vendor/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  131. data/vendor/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  132. data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  133. data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  134. data/vendor/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  135. data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  136. data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  137. data/vendor/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  138. data/vendor/ggml/src/ggml-cpu/traits.cpp +36 -0
  139. data/vendor/ggml/src/ggml-cpu/traits.h +38 -0
  140. data/vendor/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
  141. data/vendor/ggml/src/ggml-cpu/unary-ops.h +35 -0
  142. data/vendor/ggml/src/ggml-cpu/vec.cpp +613 -0
  143. data/vendor/ggml/src/ggml-cpu/vec.h +1570 -0
  144. data/vendor/ggml/src/ggml-cuda/CMakeLists.txt +267 -0
  145. data/vendor/ggml/src/ggml-cuda/acc.cu +61 -0
  146. data/vendor/ggml/src/ggml-cuda/acc.cuh +5 -0
  147. data/vendor/ggml/src/ggml-cuda/add-id.cu +58 -0
  148. data/vendor/ggml/src/ggml-cuda/add-id.cuh +3 -0
  149. data/vendor/ggml/src/ggml-cuda/allreduce.cu +971 -0
  150. data/vendor/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  151. data/vendor/ggml/src/ggml-cuda/arange.cu +34 -0
  152. data/vendor/ggml/src/ggml-cuda/arange.cuh +5 -0
  153. data/vendor/ggml/src/ggml-cuda/argmax.cu +91 -0
  154. data/vendor/ggml/src/ggml-cuda/argmax.cuh +3 -0
  155. data/vendor/ggml/src/ggml-cuda/argsort.cu +292 -0
  156. data/vendor/ggml/src/ggml-cuda/argsort.cuh +20 -0
  157. data/vendor/ggml/src/ggml-cuda/binbcast.cu +574 -0
  158. data/vendor/ggml/src/ggml-cuda/binbcast.cuh +12 -0
  159. data/vendor/ggml/src/ggml-cuda/clamp.cu +45 -0
  160. data/vendor/ggml/src/ggml-cuda/clamp.cuh +5 -0
  161. data/vendor/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
  162. data/vendor/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
  163. data/vendor/ggml/src/ggml-cuda/common.cuh +1678 -0
  164. data/vendor/ggml/src/ggml-cuda/concat.cu +242 -0
  165. data/vendor/ggml/src/ggml-cuda/concat.cuh +5 -0
  166. data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cu +88 -0
  167. data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
  168. data/vendor/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
  169. data/vendor/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
  170. data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cu +115 -0
  171. data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cuh +5 -0
  172. data/vendor/ggml/src/ggml-cuda/conv2d.cu +167 -0
  173. data/vendor/ggml/src/ggml-cuda/conv2d.cuh +5 -0
  174. data/vendor/ggml/src/ggml-cuda/convert.cu +705 -0
  175. data/vendor/ggml/src/ggml-cuda/convert.cuh +66 -0
  176. data/vendor/ggml/src/ggml-cuda/count-equal.cu +64 -0
  177. data/vendor/ggml/src/ggml-cuda/count-equal.cuh +5 -0
  178. data/vendor/ggml/src/ggml-cuda/cp-async.cuh +57 -0
  179. data/vendor/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
  180. data/vendor/ggml/src/ggml-cuda/cpy.cu +617 -0
  181. data/vendor/ggml/src/ggml-cuda/cpy.cuh +7 -0
  182. data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cu +177 -0
  183. data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
  184. data/vendor/ggml/src/ggml-cuda/cumsum.cu +307 -0
  185. data/vendor/ggml/src/ggml-cuda/cumsum.cuh +5 -0
  186. data/vendor/ggml/src/ggml-cuda/dequantize.cuh +452 -0
  187. data/vendor/ggml/src/ggml-cuda/diag.cu +77 -0
  188. data/vendor/ggml/src/ggml-cuda/diag.cuh +5 -0
  189. data/vendor/ggml/src/ggml-cuda/diagmask.cu +40 -0
  190. data/vendor/ggml/src/ggml-cuda/diagmask.cuh +5 -0
  191. data/vendor/ggml/src/ggml-cuda/dsv4-hc.cu +294 -0
  192. data/vendor/ggml/src/ggml-cuda/dsv4-hc.cuh +6 -0
  193. data/vendor/ggml/src/ggml-cuda/fattn-common.cuh +1289 -0
  194. data/vendor/ggml/src/ggml-cuda/fattn-mma-f16.cuh +2136 -0
  195. data/vendor/ggml/src/ggml-cuda/fattn-swizzle.cuh +126 -0
  196. data/vendor/ggml/src/ggml-cuda/fattn-tile.cu +60 -0
  197. data/vendor/ggml/src/ggml-cuda/fattn-tile.cuh +1355 -0
  198. data/vendor/ggml/src/ggml-cuda/fattn-vec.cuh +611 -0
  199. data/vendor/ggml/src/ggml-cuda/fattn.cu +722 -0
  200. data/vendor/ggml/src/ggml-cuda/fattn.cuh +7 -0
  201. data/vendor/ggml/src/ggml-cuda/fill.cu +37 -0
  202. data/vendor/ggml/src/ggml-cuda/fill.cuh +3 -0
  203. data/vendor/ggml/src/ggml-cuda/fwht.cu +101 -0
  204. data/vendor/ggml/src/ggml-cuda/fwht.cuh +4 -0
  205. data/vendor/ggml/src/ggml-cuda/gated_delta_net.cu +327 -0
  206. data/vendor/ggml/src/ggml-cuda/gated_delta_net.cuh +14 -0
  207. data/vendor/ggml/src/ggml-cuda/getrows.cu +490 -0
  208. data/vendor/ggml/src/ggml-cuda/getrows.cuh +15 -0
  209. data/vendor/ggml/src/ggml-cuda/ggml-cuda.cu +5781 -0
  210. data/vendor/ggml/src/ggml-cuda/gla.cu +93 -0
  211. data/vendor/ggml/src/ggml-cuda/gla.cuh +3 -0
  212. data/vendor/ggml/src/ggml-cuda/im2col.cu +267 -0
  213. data/vendor/ggml/src/ggml-cuda/im2col.cuh +6 -0
  214. data/vendor/ggml/src/ggml-cuda/lightning-indexer.cu +588 -0
  215. data/vendor/ggml/src/ggml-cuda/lightning-indexer.cuh +4 -0
  216. data/vendor/ggml/src/ggml-cuda/mean.cu +77 -0
  217. data/vendor/ggml/src/ggml-cuda/mean.cuh +3 -0
  218. data/vendor/ggml/src/ggml-cuda/mma.cuh +1456 -0
  219. data/vendor/ggml/src/ggml-cuda/mmf.cu +191 -0
  220. data/vendor/ggml/src/ggml-cuda/mmf.cuh +908 -0
  221. data/vendor/ggml/src/ggml-cuda/mmid.cu +180 -0
  222. data/vendor/ggml/src/ggml-cuda/mmid.cuh +5 -0
  223. data/vendor/ggml/src/ggml-cuda/mmq-config-ampere.cuh +383 -0
  224. data/vendor/ggml/src/ggml-cuda/mmq-config-blackwell.cuh +37 -0
  225. data/vendor/ggml/src/ggml-cuda/mmq-config-cdna.cuh +185 -0
  226. data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-dp4a.cuh +273 -0
  227. data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-older.cuh +273 -0
  228. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna2.cuh +273 -0
  229. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh +290 -0
  230. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3.cuh +274 -0
  231. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna4.cuh +290 -0
  232. data/vendor/ggml/src/ggml-cuda/mmq-load-tiles.cuh +1768 -0
  233. data/vendor/ggml/src/ggml-cuda/mmq-vec-dot.cuh +1240 -0
  234. data/vendor/ggml/src/ggml-cuda/mmq.cu +386 -0
  235. data/vendor/ggml/src/ggml-cuda/mmq.cuh +1598 -0
  236. data/vendor/ggml/src/ggml-cuda/mmvf.cu +875 -0
  237. data/vendor/ggml/src/ggml-cuda/mmvf.cuh +14 -0
  238. data/vendor/ggml/src/ggml-cuda/mmvq.cu +1504 -0
  239. data/vendor/ggml/src/ggml-cuda/mmvq.cuh +18 -0
  240. data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cu +65 -0
  241. data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cuh +7 -0
  242. data/vendor/ggml/src/ggml-cuda/norm.cu +698 -0
  243. data/vendor/ggml/src/ggml-cuda/norm.cuh +18 -0
  244. data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
  245. data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
  246. data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
  247. data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
  248. data/vendor/ggml/src/ggml-cuda/out-prod.cu +125 -0
  249. data/vendor/ggml/src/ggml-cuda/out-prod.cuh +3 -0
  250. data/vendor/ggml/src/ggml-cuda/pad.cu +106 -0
  251. data/vendor/ggml/src/ggml-cuda/pad.cuh +5 -0
  252. data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
  253. data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
  254. data/vendor/ggml/src/ggml-cuda/pool1d.cu +85 -0
  255. data/vendor/ggml/src/ggml-cuda/pool1d.cuh +5 -0
  256. data/vendor/ggml/src/ggml-cuda/pool2d.cu +94 -0
  257. data/vendor/ggml/src/ggml-cuda/pool2d.cuh +5 -0
  258. data/vendor/ggml/src/ggml-cuda/quantize.cu +697 -0
  259. data/vendor/ggml/src/ggml-cuda/quantize.cuh +70 -0
  260. data/vendor/ggml/src/ggml-cuda/reduce_rows.cuh +43 -0
  261. data/vendor/ggml/src/ggml-cuda/roll.cu +67 -0
  262. data/vendor/ggml/src/ggml-cuda/roll.cuh +5 -0
  263. data/vendor/ggml/src/ggml-cuda/rope.cu +941 -0
  264. data/vendor/ggml/src/ggml-cuda/rope.cuh +11 -0
  265. data/vendor/ggml/src/ggml-cuda/scale.cu +37 -0
  266. data/vendor/ggml/src/ggml-cuda/scale.cuh +5 -0
  267. data/vendor/ggml/src/ggml-cuda/set-rows.cu +398 -0
  268. data/vendor/ggml/src/ggml-cuda/set-rows.cuh +7 -0
  269. data/vendor/ggml/src/ggml-cuda/set.cu +39 -0
  270. data/vendor/ggml/src/ggml-cuda/set.cuh +7 -0
  271. data/vendor/ggml/src/ggml-cuda/snake.cu +72 -0
  272. data/vendor/ggml/src/ggml-cuda/snake.cuh +8 -0
  273. data/vendor/ggml/src/ggml-cuda/softcap.cu +37 -0
  274. data/vendor/ggml/src/ggml-cuda/softcap.cuh +5 -0
  275. data/vendor/ggml/src/ggml-cuda/softmax.cu +480 -0
  276. data/vendor/ggml/src/ggml-cuda/softmax.cuh +7 -0
  277. data/vendor/ggml/src/ggml-cuda/solve_tri.cu +273 -0
  278. data/vendor/ggml/src/ggml-cuda/solve_tri.cuh +3 -0
  279. data/vendor/ggml/src/ggml-cuda/ssm-conv.cu +206 -0
  280. data/vendor/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
  281. data/vendor/ggml/src/ggml-cuda/ssm-scan.cu +859 -0
  282. data/vendor/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
  283. data/vendor/ggml/src/ggml-cuda/sum.cu +41 -0
  284. data/vendor/ggml/src/ggml-cuda/sum.cuh +5 -0
  285. data/vendor/ggml/src/ggml-cuda/sumrows.cu +47 -0
  286. data/vendor/ggml/src/ggml-cuda/sumrows.cuh +4 -0
  287. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +6 -0
  288. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
  289. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +12 -0
  290. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
  291. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +11 -0
  292. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +12 -0
  293. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +6 -0
  294. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
  295. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +12 -0
  296. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +12 -0
  297. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
  298. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +11 -0
  299. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +6 -0
  300. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +11 -0
  301. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +12 -0
  302. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +12 -0
  303. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
  304. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
  305. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +11 -0
  306. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +12 -0
  307. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +12 -0
  308. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu +5 -0
  309. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu +5 -0
  310. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  311. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu +5 -0
  312. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  313. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu +5 -0
  314. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  315. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu +5 -0
  316. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu +5 -0
  317. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu +5 -0
  318. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu +5 -0
  319. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu +5 -0
  320. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  321. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  322. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  323. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  324. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  325. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  326. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  327. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  328. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
  329. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
  330. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
  331. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
  332. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
  333. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
  334. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  335. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
  336. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
  337. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
  338. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
  339. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
  340. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
  341. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  342. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
  343. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
  344. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
  345. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
  346. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
  347. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
  348. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  349. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
  350. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
  351. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
  352. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
  353. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
  354. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
  355. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  356. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
  357. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
  358. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
  359. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
  360. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
  361. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
  362. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  363. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
  364. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
  365. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
  366. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
  367. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
  368. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
  369. data/vendor/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +111 -0
  370. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
  371. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
  372. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
  373. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
  374. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
  375. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
  376. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
  377. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
  378. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
  379. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
  380. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
  381. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
  382. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
  383. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
  384. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
  385. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
  386. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
  387. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
  388. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
  389. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
  390. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
  391. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
  392. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
  393. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
  394. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
  395. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  396. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  397. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_0.cu +5 -0
  398. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
  399. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
  400. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
  401. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
  402. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
  403. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
  404. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
  405. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
  406. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
  407. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
  408. data/vendor/ggml/src/ggml-cuda/top-k.cu +275 -0
  409. data/vendor/ggml/src/ggml-cuda/top-k.cuh +3 -0
  410. data/vendor/ggml/src/ggml-cuda/topk-moe.cu +442 -0
  411. data/vendor/ggml/src/ggml-cuda/topk-moe.cuh +31 -0
  412. data/vendor/ggml/src/ggml-cuda/tri.cu +136 -0
  413. data/vendor/ggml/src/ggml-cuda/tri.cuh +5 -0
  414. data/vendor/ggml/src/ggml-cuda/tsembd.cu +47 -0
  415. data/vendor/ggml/src/ggml-cuda/tsembd.cuh +5 -0
  416. data/vendor/ggml/src/ggml-cuda/unary.cu +721 -0
  417. data/vendor/ggml/src/ggml-cuda/unary.cuh +123 -0
  418. data/vendor/ggml/src/ggml-cuda/upscale.cu +293 -0
  419. data/vendor/ggml/src/ggml-cuda/upscale.cuh +5 -0
  420. data/vendor/ggml/src/ggml-cuda/vecdotq.cuh +1371 -0
  421. data/vendor/ggml/src/ggml-cuda/vendors/cuda.h +28 -0
  422. data/vendor/ggml/src/ggml-cuda/vendors/hip.h +301 -0
  423. data/vendor/ggml/src/ggml-cuda/vendors/musa.h +151 -0
  424. data/vendor/ggml/src/ggml-cuda/wkv.cu +253 -0
  425. data/vendor/ggml/src/ggml-cuda/wkv.cuh +7 -0
  426. data/vendor/ggml/src/ggml-et/CMakeLists.txt +246 -0
  427. data/vendor/ggml/src/ggml-et/cmake/embed_one_kernel.cmake +15 -0
  428. data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.cpp.in +6 -0
  429. data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.hpp.in +12 -0
  430. data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.cpp.in +18 -0
  431. data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.h.in +13 -0
  432. data/vendor/ggml/src/ggml-et/et-kernels/CMakeLists.txt +137 -0
  433. data/vendor/ggml/src/ggml-et/et-kernels/scripts/check_unimplemented_instructions.sh +36 -0
  434. data/vendor/ggml/src/ggml-et/et-kernels/src/RunBackend.sh +23 -0
  435. data/vendor/ggml/src/ggml-et/et-kernels/src/block_ops.h +997 -0
  436. data/vendor/ggml/src/ggml-et/et-kernels/src/clamp_f32.c +120 -0
  437. data/vendor/ggml/src/ggml-et/et-kernels/src/concat_f32.c +175 -0
  438. data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f16.c +107 -0
  439. data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f32.c +248 -0
  440. data/vendor/ggml/src/ggml-et/et-kernels/src/conv_2d_f32_me.c +807 -0
  441. data/vendor/ggml/src/ggml-et/et-kernels/src/cpy_f32_f16.c +110 -0
  442. data/vendor/ggml/src/ggml-et/et-kernels/src/crt.S +15 -0
  443. data/vendor/ggml/src/ggml-et/et-kernels/src/cumsum_f32.c +96 -0
  444. data/vendor/ggml/src/ggml-et/et-kernels/src/diag_f32.c +90 -0
  445. data/vendor/ggml/src/ggml-et/et-kernels/src/el_map_f32.c +377 -0
  446. data/vendor/ggml/src/ggml-et/et-kernels/src/fill_f32.c +87 -0
  447. data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f16_me.c +1000 -0
  448. data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f32.c +217 -0
  449. data/vendor/ggml/src/ggml-et/et-kernels/src/gated_delta_net_f32.c +346 -0
  450. data/vendor/ggml/src/ggml-et/et-kernels/src/get_rows_f32.c +612 -0
  451. data/vendor/ggml/src/ggml-et/et-kernels/src/ggml_tensor.h +44 -0
  452. data/vendor/ggml/src/ggml-et/et-kernels/src/glu_f32.c +606 -0
  453. data/vendor/ggml/src/ggml-et/et-kernels/src/group_norm_f32.c +171 -0
  454. data/vendor/ggml/src/ggml-et/et-kernels/src/im2col.c +130 -0
  455. data/vendor/ggml/src/ggml-et/et-kernels/src/l2_norm_f32.c +237 -0
  456. data/vendor/ggml/src/ggml-et/et-kernels/src/linker.ld +85 -0
  457. data/vendor/ggml/src/ggml-et/et-kernels/src/math_fp.h +299 -0
  458. data/vendor/ggml/src/ggml-et/et-kernels/src/mean_f32.c +220 -0
  459. data/vendor/ggml/src/ggml-et/et-kernels/src/memops.c +181 -0
  460. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0.c +358 -0
  461. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0_matrix_engine.c +368 -0
  462. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q8_0.c +413 -0
  463. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16.c +142 -0
  464. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16_matrix_engine.c +329 -0
  465. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32.c +137 -0
  466. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32_matrix_engine.c +155 -0
  467. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q4_0.c +169 -0
  468. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q8_0.c +160 -0
  469. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_f32.c +288 -0
  470. data/vendor/ggml/src/ggml-et/et-kernels/src/norm_f32.c +328 -0
  471. data/vendor/ggml/src/ggml-et/et-kernels/src/pad_f32.c +165 -0
  472. data/vendor/ggml/src/ggml-et/et-kernels/src/platform.h +545 -0
  473. data/vendor/ggml/src/ggml-et/et-kernels/src/quants.h +72 -0
  474. data/vendor/ggml/src/ggml-et/et-kernels/src/repeat_f32.c +118 -0
  475. data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_f32.c +270 -0
  476. data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_mul_f32.c +290 -0
  477. data/vendor/ggml/src/ggml-et/et-kernels/src/rope_f32.c +656 -0
  478. data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv6_f32.c +184 -0
  479. data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv7_f32.c +272 -0
  480. data/vendor/ggml/src/ggml-et/et-kernels/src/scale_f32.c +94 -0
  481. data/vendor/ggml/src/ggml-et/et-kernels/src/set_f32.c +101 -0
  482. data/vendor/ggml/src/ggml-et/et-kernels/src/set_rows_f32.c +394 -0
  483. data/vendor/ggml/src/ggml-et/et-kernels/src/softmax_f32.c +698 -0
  484. data/vendor/ggml/src/ggml-et/et-kernels/src/solve_tri_f32.c +109 -0
  485. data/vendor/ggml/src/ggml-et/et-kernels/src/sqr_f32.c +88 -0
  486. data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_conv_f32.c +129 -0
  487. data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_scan_f32.c +282 -0
  488. data/vendor/ggml/src/ggml-et/et-kernels/src/sum_rows_f32.c +103 -0
  489. data/vendor/ggml/src/ggml-et/et-kernels/src/tensor.h +897 -0
  490. data/vendor/ggml/src/ggml-et/et-kernels/src/tri_f32.c +244 -0
  491. data/vendor/ggml/src/ggml-et/et-kernels/src/uberkernel.c +497 -0
  492. data/vendor/ggml/src/ggml-et/et-kernels/src/unary_f32.c +705 -0
  493. data/vendor/ggml/src/ggml-et/ggml-et-common.h +86 -0
  494. data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.cpp +502 -0
  495. data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.h +54 -0
  496. data/vendor/ggml/src/ggml-et/ggml-et-kernels.cpp +508 -0
  497. data/vendor/ggml/src/ggml-et/ggml-et-kernels.h +48 -0
  498. data/vendor/ggml/src/ggml-et/ggml-et-memops.cpp +36 -0
  499. data/vendor/ggml/src/ggml-et/ggml-et-memops.h +18 -0
  500. data/vendor/ggml/src/ggml-et/ggml-et-ops.cpp +2584 -0
  501. data/vendor/ggml/src/ggml-et/ggml-et-ops.h +393 -0
  502. data/vendor/ggml/src/ggml-et/ggml-et-uberkernel-common.h +17 -0
  503. data/vendor/ggml/src/ggml-et/ggml-et.cpp +1880 -0
  504. data/vendor/ggml/src/ggml-feats.h +166 -0
  505. data/vendor/ggml/src/ggml-hexagon/CMakeLists.txt +113 -0
  506. data/vendor/ggml/src/ggml-hexagon/ggml-hexagon.cpp +6523 -0
  507. data/vendor/ggml/src/ggml-hexagon/htp/CMakeLists.txt +58 -0
  508. data/vendor/ggml/src/ggml-hexagon/htp/act-ops.c +585 -0
  509. data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.c +398 -0
  510. data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.h +40 -0
  511. data/vendor/ggml/src/ggml-hexagon/htp/argsort-ops.c +514 -0
  512. data/vendor/ggml/src/ggml-hexagon/htp/binary-ops.c +873 -0
  513. data/vendor/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +155 -0
  514. data/vendor/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  515. data/vendor/ggml/src/ggml-hexagon/htp/cpy-ops.c +354 -0
  516. data/vendor/ggml/src/ggml-hexagon/htp/cumsum-ops.c +265 -0
  517. data/vendor/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  518. data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.c +104 -0
  519. data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.h +408 -0
  520. data/vendor/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  521. data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +2479 -0
  522. data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +344 -0
  523. data/vendor/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1157 -0
  524. data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.c +260 -0
  525. data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.h +77 -0
  526. data/vendor/ggml/src/ggml-hexagon/htp/hex-bitmap.h +24 -0
  527. data/vendor/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
  528. data/vendor/ggml/src/ggml-hexagon/htp/hex-dma.h +2 -0
  529. data/vendor/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
  530. data/vendor/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  531. data/vendor/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
  532. data/vendor/ggml/src/ggml-hexagon/htp/hex-utils.h +65 -0
  533. data/vendor/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
  534. data/vendor/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1378 -0
  535. data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.c +170 -0
  536. data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.h +160 -0
  537. data/vendor/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
  538. data/vendor/ggml/src/ggml-hexagon/htp/htp-ctx.h +146 -0
  539. data/vendor/ggml/src/ggml-hexagon/htp/htp-ops.h +242 -0
  540. data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.c +300 -0
  541. data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.h +29 -0
  542. data/vendor/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
  543. data/vendor/ggml/src/ggml-hexagon/htp/htp_iface.idl +23 -0
  544. data/vendor/ggml/src/ggml-hexagon/htp/hvx-arith.h +636 -0
  545. data/vendor/ggml/src/ggml-hexagon/htp/hvx-base.h +292 -0
  546. data/vendor/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
  547. data/vendor/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
  548. data/vendor/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  549. data/vendor/ggml/src/ggml-hexagon/htp/hvx-exp.h +255 -0
  550. data/vendor/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +303 -0
  551. data/vendor/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  552. data/vendor/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  553. data/vendor/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  554. data/vendor/ggml/src/ggml-hexagon/htp/hvx-log.h +118 -0
  555. data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
  556. data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
  557. data/vendor/ggml/src/ggml-hexagon/htp/hvx-norm.h +454 -0
  558. data/vendor/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  559. data/vendor/ggml/src/ggml-hexagon/htp/hvx-quant.h +165 -0
  560. data/vendor/ggml/src/ggml-hexagon/htp/hvx-reduce.h +336 -0
  561. data/vendor/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  562. data/vendor/ggml/src/ggml-hexagon/htp/hvx-scale.h +199 -0
  563. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +181 -0
  564. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  565. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +189 -0
  566. data/vendor/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  567. data/vendor/ggml/src/ggml-hexagon/htp/hvx-utils.h +24 -0
  568. data/vendor/ggml/src/ggml-hexagon/htp/im2col-ops.c +306 -0
  569. data/vendor/ggml/src/ggml-hexagon/htp/main.c +1245 -0
  570. data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.c +4078 -0
  571. data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.h +778 -0
  572. data/vendor/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  573. data/vendor/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  574. data/vendor/ggml/src/ggml-hexagon/htp/rope-ops.c +733 -0
  575. data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.c +230 -0
  576. data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.h +74 -0
  577. data/vendor/ggml/src/ggml-hexagon/htp/softmax-ops.c +407 -0
  578. data/vendor/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  579. data/vendor/ggml/src/ggml-hexagon/htp/ssm-conv.c +433 -0
  580. data/vendor/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  581. data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.c +1264 -0
  582. data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.h +176 -0
  583. data/vendor/ggml/src/ggml-hexagon/htp/work-queue.c +244 -0
  584. data/vendor/ggml/src/ggml-hexagon/htp/work-queue.h +38 -0
  585. data/vendor/ggml/src/ggml-hexagon/htp-drv.cpp +425 -0
  586. data/vendor/ggml/src/ggml-hexagon/htp-drv.h +123 -0
  587. data/vendor/ggml/src/ggml-hexagon/htp-opnode.h +372 -0
  588. data/vendor/ggml/src/ggml-hexagon/libdl.h +79 -0
  589. data/vendor/ggml/src/ggml-hexagon/libggml-htp.inf +36 -0
  590. data/vendor/ggml/src/ggml-hip/CMakeLists.txt +153 -0
  591. data/vendor/ggml/src/ggml-impl.h +783 -0
  592. data/vendor/ggml/src/ggml-metal/CMakeLists.txt +240 -0
  593. data/vendor/ggml/src/ggml-metal/ggml-metal-common.cpp +474 -0
  594. data/vendor/ggml/src/ggml-metal/ggml-metal-common.h +56 -0
  595. data/vendor/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
  596. data/vendor/ggml/src/ggml-metal/ggml-metal-context.m +759 -0
  597. data/vendor/ggml/src/ggml-metal/ggml-metal-device.cpp +2330 -0
  598. data/vendor/ggml/src/ggml-metal/ggml-metal-device.h +355 -0
  599. data/vendor/ggml/src/ggml-metal/ggml-metal-device.m +2473 -0
  600. data/vendor/ggml/src/ggml-metal/ggml-metal-impl.h +1345 -0
  601. data/vendor/ggml/src/ggml-metal/ggml-metal-ops.cpp +5610 -0
  602. data/vendor/ggml/src/ggml-metal/ggml-metal-ops.h +106 -0
  603. data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.cpp +3732 -0
  604. data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.h +77 -0
  605. data/vendor/ggml/src/ggml-metal/ggml-metal.cpp +1006 -0
  606. data/vendor/ggml/src/ggml-metal/kernels/argsort.metal +337 -0
  607. data/vendor/ggml/src/ggml-metal/kernels/binbcast.metal +228 -0
  608. data/vendor/ggml/src/ggml-metal/kernels/common.h +126 -0
  609. data/vendor/ggml/src/ggml-metal/kernels/conv.metal +724 -0
  610. data/vendor/ggml/src/ggml-metal/kernels/dequantize.h +735 -0
  611. data/vendor/ggml/src/ggml-metal/kernels/fa.metal +2445 -0
  612. data/vendor/ggml/src/ggml-metal/kernels/gated_delta_net.metal +250 -0
  613. data/vendor/ggml/src/ggml-metal/kernels/misc.metal +595 -0
  614. data/vendor/ggml/src/ggml-metal/kernels/mul_mm.metal +853 -0
  615. data/vendor/ggml/src/ggml-metal/kernels/mul_mv.metal +3225 -0
  616. data/vendor/ggml/src/ggml-metal/kernels/norm.metal +308 -0
  617. data/vendor/ggml/src/ggml-metal/kernels/pool.metal +148 -0
  618. data/vendor/ggml/src/ggml-metal/kernels/quantize.h +262 -0
  619. data/vendor/ggml/src/ggml-metal/kernels/quantize.metal +480 -0
  620. data/vendor/ggml/src/ggml-metal/kernels/reduce.metal +228 -0
  621. data/vendor/ggml/src/ggml-metal/kernels/rope.metal +333 -0
  622. data/vendor/ggml/src/ggml-metal/kernels/softmax.metal +223 -0
  623. data/vendor/ggml/src/ggml-metal/kernels/solve_tri.metal +75 -0
  624. data/vendor/ggml/src/ggml-metal/kernels/ssm.metal +473 -0
  625. data/vendor/ggml/src/ggml-metal/kernels/tri.metal +69 -0
  626. data/vendor/ggml/src/ggml-metal/kernels/unary.metal +400 -0
  627. data/vendor/ggml/src/ggml-metal/kernels/upscale.metal +179 -0
  628. data/vendor/ggml/src/ggml-metal/kernels/wkv.metal +179 -0
  629. data/vendor/ggml/src/ggml-musa/CMakeLists.txt +123 -0
  630. data/vendor/ggml/src/ggml-musa/mudnn.cu +112 -0
  631. data/vendor/ggml/src/ggml-musa/mudnn.cuh +12 -0
  632. data/vendor/ggml/src/ggml-opencl/CMakeLists.txt +246 -0
  633. data/vendor/ggml/src/ggml-opencl/cl-program-cache.cpp +453 -0
  634. data/vendor/ggml/src/ggml-opencl/cl-program-cache.h +75 -0
  635. data/vendor/ggml/src/ggml-opencl/fa_tune.h +92 -0
  636. data/vendor/ggml/src/ggml-opencl/ggml-opencl.cpp +26982 -0
  637. data/vendor/ggml/src/ggml-opencl/kernels/abs.cl +113 -0
  638. data/vendor/ggml/src/ggml-opencl/kernels/add.cl +190 -0
  639. data/vendor/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  640. data/vendor/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
  641. data/vendor/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
  642. data/vendor/ggml/src/ggml-opencl/kernels/concat.cl +118 -0
  643. data/vendor/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
  644. data/vendor/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
  645. data/vendor/ggml/src/ggml-opencl/kernels/cpy.cl +288 -0
  646. data/vendor/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  647. data/vendor/ggml/src/ggml-opencl/kernels/cvt.cl +2663 -0
  648. data/vendor/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  649. data/vendor/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
  650. data/vendor/ggml/src/ggml-opencl/kernels/div.cl +138 -0
  651. data/vendor/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
  652. data/vendor/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  653. data/vendor/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
  654. data/vendor/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
  655. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +421 -0
  656. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +422 -0
  657. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +2656 -0
  658. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +2011 -0
  659. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1840 -0
  660. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
  661. data/vendor/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  662. data/vendor/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
  663. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
  664. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +376 -0
  665. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_q8_1_dp4a.cl +190 -0
  666. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
  667. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_q8_1_dp4a.cl +169 -0
  668. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
  669. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
  670. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_q8_1_dp4a.cl +209 -0
  671. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
  672. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
  673. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
  674. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
  675. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_q8_1_dp4a.cl +200 -0
  676. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_0_f32_ns.cl +221 -0
  677. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_1_dp4a.cl +226 -0
  678. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  679. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_q8_1_dp4a.cl +143 -0
  680. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
  681. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_f32.cl +139 -0
  682. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_q8_1_dp4a.cl +127 -0
  683. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  684. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +490 -0
  685. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_q8_1_dp4a.cl +281 -0
  686. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  687. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_q8_1_dp4a.cl +235 -0
  688. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  689. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  690. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_q8_1_dp4a.cl +164 -0
  691. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +245 -0
  692. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32_tiled.cl +136 -0
  693. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_q8_1_dp4a.cl +144 -0
  694. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
  695. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_q8_1_dp4a.cl +212 -0
  696. data/vendor/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  697. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
  698. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +257 -0
  699. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  700. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  701. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +266 -0
  702. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  703. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  704. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  705. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  706. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +307 -0
  707. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +125 -0
  708. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl +383 -0
  709. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl +272 -0
  710. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +384 -0
  711. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +823 -0
  712. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_o4.cl +349 -0
  713. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_tiled.cl +118 -0
  714. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +296 -0
  715. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +299 -0
  716. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +453 -0
  717. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +409 -0
  718. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_o4.cl +372 -0
  719. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_tiled.cl +196 -0
  720. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +280 -0
  721. data/vendor/ggml/src/ggml-opencl/kernels/get_rows.cl +193 -0
  722. data/vendor/ggml/src/ggml-opencl/kernels/glu.cl +443 -0
  723. data/vendor/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
  724. data/vendor/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
  725. data/vendor/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
  726. data/vendor/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  727. data/vendor/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
  728. data/vendor/ggml/src/ggml-opencl/kernels/moe_add_id_glu.cl +76 -0
  729. data/vendor/ggml/src/ggml-opencl/kernels/moe_combine.cl +79 -0
  730. data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  731. data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_quant_a_q8_1.cl +64 -0
  732. data/vendor/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +155 -0
  733. data/vendor/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
  734. data/vendor/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
  735. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
  736. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
  737. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +196 -0
  738. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  739. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
  740. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  741. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  742. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +189 -0
  743. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  744. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  745. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +202 -0
  746. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  747. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
  748. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +125 -0
  749. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +125 -0
  750. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +101 -0
  751. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1237 -0
  752. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_mrow.cl +306 -0
  753. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
  754. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  755. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
  756. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
  757. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
  758. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
  759. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  760. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  761. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  762. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
  763. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
  764. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
  765. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
  766. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
  767. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
  768. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
  769. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
  770. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  771. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  772. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +182 -0
  773. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +205 -0
  774. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  775. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  776. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  777. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  778. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  779. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +244 -0
  780. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
  781. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +272 -0
  782. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
  783. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
  784. data/vendor/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  785. data/vendor/ggml/src/ggml-opencl/kernels/norm.cl +164 -0
  786. data/vendor/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
  787. data/vendor/ggml/src/ggml-opencl/kernels/quant_a_q8_1.cl +42 -0
  788. data/vendor/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
  789. data/vendor/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
  790. data/vendor/ggml/src/ggml-opencl/kernels/rms_norm.cl +369 -0
  791. data/vendor/ggml/src/ggml-opencl/kernels/rope.cl +759 -0
  792. data/vendor/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
  793. data/vendor/ggml/src/ggml-opencl/kernels/set_rows.cl +708 -0
  794. data/vendor/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
  795. data/vendor/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
  796. data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
  797. data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
  798. data/vendor/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
  799. data/vendor/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
  800. data/vendor/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
  801. data/vendor/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  802. data/vendor/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
  803. data/vendor/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
  804. data/vendor/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
  805. data/vendor/ggml/src/ggml-opencl/kernels/ssm_scan.cl +216 -0
  806. data/vendor/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
  807. data/vendor/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
  808. data/vendor/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
  809. data/vendor/ggml/src/ggml-opencl/kernels/transpose.cl +143 -0
  810. data/vendor/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  811. data/vendor/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
  812. data/vendor/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
  813. data/vendor/ggml/src/ggml-opencl/libdl.h +79 -0
  814. data/vendor/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  815. data/vendor/ggml/src/ggml-openvino/ggml-decoder.cpp +2057 -0
  816. data/vendor/ggml/src/ggml-openvino/ggml-decoder.h +417 -0
  817. data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +493 -0
  818. data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.h +216 -0
  819. data/vendor/ggml/src/ggml-openvino/ggml-openvino.cpp +1566 -0
  820. data/vendor/ggml/src/ggml-openvino/ggml-quants.cpp +1273 -0
  821. data/vendor/ggml/src/ggml-openvino/ggml-quants.h +186 -0
  822. data/vendor/ggml/src/ggml-openvino/model-cache.cpp +272 -0
  823. data/vendor/ggml/src/ggml-openvino/model-cache.h +56 -0
  824. data/vendor/ggml/src/ggml-openvino/openvino/decoder.h +133 -0
  825. data/vendor/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  826. data/vendor/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  827. data/vendor/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  828. data/vendor/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  829. data/vendor/ggml/src/ggml-openvino/openvino/node_context.h +171 -0
  830. data/vendor/ggml/src/ggml-openvino/openvino/op/add.cpp +45 -0
  831. data/vendor/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
  832. data/vendor/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
  833. data/vendor/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
  834. data/vendor/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
  835. data/vendor/ggml/src/ggml-openvino/openvino/op/cont.cpp +40 -0
  836. data/vendor/ggml/src/ggml-openvino/openvino/op/cpy.cpp +295 -0
  837. data/vendor/ggml/src/ggml-openvino/openvino/op/cumsum.cpp +29 -0
  838. data/vendor/ggml/src/ggml-openvino/openvino/op/diag.cpp +58 -0
  839. data/vendor/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
  840. data/vendor/ggml/src/ggml-openvino/openvino/op/fill.cpp +34 -0
  841. data/vendor/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +241 -0
  842. data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +327 -0
  843. data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
  844. data/vendor/ggml/src/ggml-openvino/openvino/op/gather_matmul.hpp +43 -0
  845. data/vendor/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +142 -0
  846. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +75 -0
  847. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu_quick.cpp +64 -0
  848. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +110 -0
  849. data/vendor/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
  850. data/vendor/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +61 -0
  851. data/vendor/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +283 -0
  852. data/vendor/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +100 -0
  853. data/vendor/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
  854. data/vendor/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
  855. data/vendor/ggml/src/ggml-openvino/openvino/op/permute.cpp +147 -0
  856. data/vendor/ggml/src/ggml-openvino/openvino/op/pool_2d.cpp +53 -0
  857. data/vendor/ggml/src/ggml-openvino/openvino/op/repeat.cpp +48 -0
  858. data/vendor/ggml/src/ggml-openvino/openvino/op/reshape.cpp +113 -0
  859. data/vendor/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +81 -0
  860. data/vendor/ggml/src/ggml-openvino/openvino/op/roll.cpp +36 -0
  861. data/vendor/ggml/src/ggml-openvino/openvino/op/rope.cpp +303 -0
  862. data/vendor/ggml/src/ggml-openvino/openvino/op/scale.cpp +86 -0
  863. data/vendor/ggml/src/ggml-openvino/openvino/op/set.cpp +76 -0
  864. data/vendor/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +95 -0
  865. data/vendor/ggml/src/ggml-openvino/openvino/op/softmax.cpp +166 -0
  866. data/vendor/ggml/src/ggml-openvino/openvino/op/solve_tri.cpp +108 -0
  867. data/vendor/ggml/src/ggml-openvino/openvino/op/sqr.cpp +35 -0
  868. data/vendor/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +61 -0
  869. data/vendor/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
  870. data/vendor/ggml/src/ggml-openvino/openvino/op/transpose.cpp +54 -0
  871. data/vendor/ggml/src/ggml-openvino/openvino/op/tri.cpp +82 -0
  872. data/vendor/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  873. data/vendor/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
  874. data/vendor/ggml/src/ggml-openvino/openvino/op/view.cpp +245 -0
  875. data/vendor/ggml/src/ggml-openvino/openvino/op_table.cpp +88 -0
  876. data/vendor/ggml/src/ggml-openvino/openvino/op_table.h +67 -0
  877. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp +212 -0
  878. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h +17 -0
  879. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  880. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  881. data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  882. data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_dequantization_subgraph.h +44 -0
  883. data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  884. data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  885. data/vendor/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  886. data/vendor/ggml/src/ggml-openvino/openvino/translate_session.cpp +432 -0
  887. data/vendor/ggml/src/ggml-openvino/openvino/translate_session.h +29 -0
  888. data/vendor/ggml/src/ggml-openvino/openvino/utils.cpp +886 -0
  889. data/vendor/ggml/src/ggml-openvino/openvino/utils.h +88 -0
  890. data/vendor/ggml/src/ggml-openvino/utils.cpp +1496 -0
  891. data/vendor/ggml/src/ggml-openvino/utils.h +195 -0
  892. data/vendor/ggml/src/ggml-opt.cpp +1094 -0
  893. data/vendor/ggml/src/ggml-quants.c +5667 -0
  894. data/vendor/ggml/src/ggml-quants.h +115 -0
  895. data/vendor/ggml/src/ggml-rpc/CMakeLists.txt +49 -0
  896. data/vendor/ggml/src/ggml-rpc/ggml-rpc.cpp +2351 -0
  897. data/vendor/ggml/src/ggml-rpc/transport-apple.cpp +481 -0
  898. data/vendor/ggml/src/ggml-rpc/transport-apple.h +27 -0
  899. data/vendor/ggml/src/ggml-rpc/transport.cpp +739 -0
  900. data/vendor/ggml/src/ggml-rpc/transport.h +38 -0
  901. data/vendor/ggml/src/ggml-sycl/CMakeLists.txt +218 -0
  902. data/vendor/ggml/src/ggml-sycl/add-id.cpp +81 -0
  903. data/vendor/ggml/src/ggml-sycl/add-id.hpp +8 -0
  904. data/vendor/ggml/src/ggml-sycl/backend.hpp +52 -0
  905. data/vendor/ggml/src/ggml-sycl/base.hpp +36 -0
  906. data/vendor/ggml/src/ggml-sycl/binbcast.cpp +650 -0
  907. data/vendor/ggml/src/ggml-sycl/binbcast.hpp +69 -0
  908. data/vendor/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
  909. data/vendor/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
  910. data/vendor/ggml/src/ggml-sycl/common.cpp +153 -0
  911. data/vendor/ggml/src/ggml-sycl/common.hpp +1029 -0
  912. data/vendor/ggml/src/ggml-sycl/concat.cpp +509 -0
  913. data/vendor/ggml/src/ggml-sycl/concat.hpp +20 -0
  914. data/vendor/ggml/src/ggml-sycl/conv.cpp +101 -0
  915. data/vendor/ggml/src/ggml-sycl/conv.hpp +20 -0
  916. data/vendor/ggml/src/ggml-sycl/conv2d-dw.cpp +170 -0
  917. data/vendor/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
  918. data/vendor/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
  919. data/vendor/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
  920. data/vendor/ggml/src/ggml-sycl/conv2d.cpp +150 -0
  921. data/vendor/ggml/src/ggml-sycl/conv2d.hpp +10 -0
  922. data/vendor/ggml/src/ggml-sycl/conv3d.cpp +224 -0
  923. data/vendor/ggml/src/ggml-sycl/conv3d.hpp +8 -0
  924. data/vendor/ggml/src/ggml-sycl/convert.cpp +877 -0
  925. data/vendor/ggml/src/ggml-sycl/convert.hpp +64 -0
  926. data/vendor/ggml/src/ggml-sycl/count-equal.cpp +79 -0
  927. data/vendor/ggml/src/ggml-sycl/count-equal.hpp +9 -0
  928. data/vendor/ggml/src/ggml-sycl/cpy.cpp +1433 -0
  929. data/vendor/ggml/src/ggml-sycl/cpy.hpp +537 -0
  930. data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
  931. data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
  932. data/vendor/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  933. data/vendor/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  934. data/vendor/ggml/src/ggml-sycl/dequantize.hpp +1681 -0
  935. data/vendor/ggml/src/ggml-sycl/diag.cpp +67 -0
  936. data/vendor/ggml/src/ggml-sycl/diag.hpp +5 -0
  937. data/vendor/ggml/src/ggml-sycl/dmmv.cpp +2227 -0
  938. data/vendor/ggml/src/ggml-sycl/dmmv.hpp +27 -0
  939. data/vendor/ggml/src/ggml-sycl/dpct/helper.hpp +3782 -0
  940. data/vendor/ggml/src/ggml-sycl/dsv4-hc.cpp +280 -0
  941. data/vendor/ggml/src/ggml-sycl/dsv4-hc.hpp +10 -0
  942. data/vendor/ggml/src/ggml-sycl/element_wise.cpp +1432 -0
  943. data/vendor/ggml/src/ggml-sycl/element_wise.hpp +135 -0
  944. data/vendor/ggml/src/ggml-sycl/esimd.hpp +589 -0
  945. data/vendor/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  946. data/vendor/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  947. data/vendor/ggml/src/ggml-sycl/fattn-common.hpp +1185 -0
  948. data/vendor/ggml/src/ggml-sycl/fattn-mkl.cpp +689 -0
  949. data/vendor/ggml/src/ggml-sycl/fattn-onednn.cpp +450 -0
  950. data/vendor/ggml/src/ggml-sycl/fattn-onednn.hpp +18 -0
  951. data/vendor/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
  952. data/vendor/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
  953. data/vendor/ggml/src/ggml-sycl/fattn-vec.hpp +684 -0
  954. data/vendor/ggml/src/ggml-sycl/fattn.cpp +452 -0
  955. data/vendor/ggml/src/ggml-sycl/fattn.hpp +42 -0
  956. data/vendor/ggml/src/ggml-sycl/fill.cpp +55 -0
  957. data/vendor/ggml/src/ggml-sycl/fill.hpp +5 -0
  958. data/vendor/ggml/src/ggml-sycl/fusion.cpp +212 -0
  959. data/vendor/ggml/src/ggml-sycl/fusion.hpp +17 -0
  960. data/vendor/ggml/src/ggml-sycl/fwht.cpp +119 -0
  961. data/vendor/ggml/src/ggml-sycl/fwht.hpp +12 -0
  962. data/vendor/ggml/src/ggml-sycl/gated_delta_net.cpp +366 -0
  963. data/vendor/ggml/src/ggml-sycl/gated_delta_net.hpp +19 -0
  964. data/vendor/ggml/src/ggml-sycl/gemm.hpp +93 -0
  965. data/vendor/ggml/src/ggml-sycl/getrows.cpp +368 -0
  966. data/vendor/ggml/src/ggml-sycl/getrows.hpp +20 -0
  967. data/vendor/ggml/src/ggml-sycl/ggml-sycl.cpp +7028 -0
  968. data/vendor/ggml/src/ggml-sycl/gla.cpp +106 -0
  969. data/vendor/ggml/src/ggml-sycl/gla.hpp +8 -0
  970. data/vendor/ggml/src/ggml-sycl/im2col.cpp +400 -0
  971. data/vendor/ggml/src/ggml-sycl/im2col.hpp +23 -0
  972. data/vendor/ggml/src/ggml-sycl/lightning-indexer.cpp +197 -0
  973. data/vendor/ggml/src/ggml-sycl/lightning-indexer.hpp +8 -0
  974. data/vendor/ggml/src/ggml-sycl/mem.cpp +162 -0
  975. data/vendor/ggml/src/ggml-sycl/mem.hpp +16 -0
  976. data/vendor/ggml/src/ggml-sycl/mmq.cpp +3030 -0
  977. data/vendor/ggml/src/ggml-sycl/mmq.hpp +33 -0
  978. data/vendor/ggml/src/ggml-sycl/mmvq.cpp +3033 -0
  979. data/vendor/ggml/src/ggml-sycl/mmvq.hpp +76 -0
  980. data/vendor/ggml/src/ggml-sycl/norm.cpp +963 -0
  981. data/vendor/ggml/src/ggml-sycl/norm.hpp +32 -0
  982. data/vendor/ggml/src/ggml-sycl/opt-step.cpp +131 -0
  983. data/vendor/ggml/src/ggml-sycl/opt-step.hpp +6 -0
  984. data/vendor/ggml/src/ggml-sycl/outprod.cpp +83 -0
  985. data/vendor/ggml/src/ggml-sycl/outprod.hpp +10 -0
  986. data/vendor/ggml/src/ggml-sycl/pad.cpp +97 -0
  987. data/vendor/ggml/src/ggml-sycl/pad.hpp +24 -0
  988. data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.cpp +100 -0
  989. data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.hpp +10 -0
  990. data/vendor/ggml/src/ggml-sycl/pool.cpp +185 -0
  991. data/vendor/ggml/src/ggml-sycl/pool.hpp +22 -0
  992. data/vendor/ggml/src/ggml-sycl/presets.hpp +79 -0
  993. data/vendor/ggml/src/ggml-sycl/quantize.hpp +133 -0
  994. data/vendor/ggml/src/ggml-sycl/quants.hpp +204 -0
  995. data/vendor/ggml/src/ggml-sycl/repeat_back.cpp +76 -0
  996. data/vendor/ggml/src/ggml-sycl/repeat_back.hpp +8 -0
  997. data/vendor/ggml/src/ggml-sycl/roll.cpp +122 -0
  998. data/vendor/ggml/src/ggml-sycl/roll.hpp +20 -0
  999. data/vendor/ggml/src/ggml-sycl/rope.cpp +651 -0
  1000. data/vendor/ggml/src/ggml-sycl/rope.hpp +26 -0
  1001. data/vendor/ggml/src/ggml-sycl/set.cpp +73 -0
  1002. data/vendor/ggml/src/ggml-sycl/set.hpp +5 -0
  1003. data/vendor/ggml/src/ggml-sycl/set_rows.cpp +577 -0
  1004. data/vendor/ggml/src/ggml-sycl/set_rows.hpp +8 -0
  1005. data/vendor/ggml/src/ggml-sycl/softmax.cpp +425 -0
  1006. data/vendor/ggml/src/ggml-sycl/softmax.hpp +24 -0
  1007. data/vendor/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  1008. data/vendor/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  1009. data/vendor/ggml/src/ggml-sycl/ssm_conv.cpp +136 -0
  1010. data/vendor/ggml/src/ggml-sycl/ssm_conv.hpp +5 -0
  1011. data/vendor/ggml/src/ggml-sycl/ssm_scan.cpp +170 -0
  1012. data/vendor/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  1013. data/vendor/ggml/src/ggml-sycl/sycl_hw.cpp +67 -0
  1014. data/vendor/ggml/src/ggml-sycl/sycl_hw.hpp +38 -0
  1015. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  1016. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  1017. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  1018. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  1019. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  1020. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  1021. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  1022. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  1023. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  1024. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  1025. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
  1026. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
  1027. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
  1028. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
  1029. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
  1030. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
  1031. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
  1032. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
  1033. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
  1034. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
  1035. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
  1036. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
  1037. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
  1038. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
  1039. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
  1040. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
  1041. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
  1042. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
  1043. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
  1044. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
  1045. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
  1046. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
  1047. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
  1048. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
  1049. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
  1050. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
  1051. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
  1052. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
  1053. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
  1054. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
  1055. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
  1056. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
  1057. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
  1058. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
  1059. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
  1060. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
  1061. data/vendor/ggml/src/ggml-sycl/topk-moe.cpp +620 -0
  1062. data/vendor/ggml/src/ggml-sycl/topk-moe.hpp +12 -0
  1063. data/vendor/ggml/src/ggml-sycl/tsembd.cpp +73 -0
  1064. data/vendor/ggml/src/ggml-sycl/tsembd.hpp +20 -0
  1065. data/vendor/ggml/src/ggml-sycl/type.hpp +112 -0
  1066. data/vendor/ggml/src/ggml-sycl/upscale.cpp +410 -0
  1067. data/vendor/ggml/src/ggml-sycl/upscale.hpp +9 -0
  1068. data/vendor/ggml/src/ggml-sycl/vecdotq.hpp +1726 -0
  1069. data/vendor/ggml/src/ggml-sycl/wkv.cpp +293 -0
  1070. data/vendor/ggml/src/ggml-sycl/wkv.hpp +10 -0
  1071. data/vendor/ggml/src/ggml-threading.cpp +12 -0
  1072. data/vendor/ggml/src/ggml-threading.h +14 -0
  1073. data/vendor/ggml/src/ggml-version.h.in +4 -0
  1074. data/vendor/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  1075. data/vendor/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  1076. data/vendor/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  1077. data/vendor/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  1078. data/vendor/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  1079. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  1080. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  1081. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  1082. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +149 -0
  1083. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  1084. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  1085. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  1086. data/vendor/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  1087. data/vendor/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  1088. data/vendor/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  1089. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  1090. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  1091. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  1092. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  1093. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  1094. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  1095. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
  1096. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
  1097. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  1098. data/vendor/ggml/src/ggml-virtgpu/ggml-backend.cpp +72 -0
  1099. data/vendor/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  1100. data/vendor/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  1101. data/vendor/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  1102. data/vendor/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
  1103. data/vendor/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  1104. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  1105. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  1106. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  1107. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +194 -0
  1108. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  1109. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +54 -0
  1110. data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
  1111. data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  1112. data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  1113. data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  1114. data/vendor/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
  1115. data/vendor/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
  1116. data/vendor/ggml/src/ggml-vulkan/CMakeLists.txt +254 -0
  1117. data/vendor/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
  1118. data/vendor/ggml/src/ggml-vulkan/ggml-vulkan.cpp +20349 -0
  1119. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +43 -0
  1120. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +37 -0
  1121. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +69 -0
  1122. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp +28 -0
  1123. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
  1124. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp +20 -0
  1125. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +60 -0
  1126. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +86 -0
  1127. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp +114 -0
  1128. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
  1129. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
  1130. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +53 -0
  1131. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
  1132. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +473 -0
  1133. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +424 -0
  1134. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
  1135. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +25 -0
  1136. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
  1137. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +350 -0
  1138. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp +67 -0
  1139. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose_02.comp +61 -0
  1140. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
  1141. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +135 -0
  1142. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp +78 -0
  1143. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp +75 -0
  1144. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +83 -0
  1145. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
  1146. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
  1147. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
  1148. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +741 -0
  1149. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +1466 -0
  1150. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.glsl +13 -0
  1151. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
  1152. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
  1153. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
  1154. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
  1155. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +49 -0
  1156. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +40 -0
  1157. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +51 -0
  1158. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
  1159. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
  1160. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
  1161. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  1162. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  1163. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_0.comp +29 -0
  1164. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
  1165. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
  1166. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
  1167. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
  1168. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
  1169. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
  1170. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
  1171. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
  1172. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
  1173. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +42 -0
  1174. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_tq2_0.comp +31 -0
  1175. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +28 -0
  1176. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
  1177. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
  1178. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  1179. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fa_types.glsl +55 -0
  1180. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/bfloat16.comp +7 -0
  1181. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat.comp +7 -0
  1182. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2.comp +7 -0
  1183. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  1184. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e2m1.comp +7 -0
  1185. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e4m3.comp +7 -0
  1186. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/integer_dot.comp +7 -0
  1187. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp +19 -0
  1188. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +760 -0
  1189. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +223 -0
  1190. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +647 -0
  1191. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +487 -0
  1192. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +146 -0
  1193. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  1194. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  1195. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +121 -0
  1196. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  1197. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
  1198. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
  1199. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
  1200. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
  1201. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +65 -0
  1202. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.glsl +11 -0
  1203. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +75 -0
  1204. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +42 -0
  1205. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
  1206. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +51 -0
  1207. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gla.comp +82 -0
  1208. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +43 -0
  1209. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +35 -0
  1210. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
  1211. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +138 -0
  1212. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +124 -0
  1213. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
  1214. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/lightning_indexer.comp +151 -0
  1215. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +17 -0
  1216. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
  1217. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
  1218. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +264 -0
  1219. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +230 -0
  1220. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +35 -0
  1221. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +132 -0
  1222. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +95 -0
  1223. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
  1224. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +105 -0
  1225. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
  1226. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +93 -0
  1227. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
  1228. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +124 -0
  1229. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +156 -0
  1230. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +128 -0
  1231. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
  1232. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +134 -0
  1233. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +165 -0
  1234. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
  1235. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_tq2_0.comp +102 -0
  1236. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +141 -0
  1237. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +564 -0
  1238. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +475 -0
  1239. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +690 -0
  1240. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +670 -0
  1241. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +89 -0
  1242. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +317 -0
  1243. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +488 -0
  1244. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +99 -0
  1245. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +194 -0
  1246. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
  1247. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
  1248. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
  1249. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/out_prod.comp +59 -0
  1250. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +64 -0
  1251. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp +43 -0
  1252. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool1d.comp +65 -0
  1253. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
  1254. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +127 -0
  1255. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
  1256. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
  1257. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
  1258. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +150 -0
  1259. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
  1260. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
  1261. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
  1262. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +218 -0
  1263. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +19 -0
  1264. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +17 -0
  1265. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +17 -0
  1266. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +17 -0
  1267. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +35 -0
  1268. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +17 -0
  1269. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
  1270. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
  1271. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  1272. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +195 -0
  1273. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +54 -0
  1274. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp +62 -0
  1275. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp +79 -0
  1276. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp +65 -0
  1277. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large_common.glsl +53 -0
  1278. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp +81 -0
  1279. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +60 -0
  1280. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp +134 -0
  1281. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
  1282. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +47 -0
  1283. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl +19 -0
  1284. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
  1285. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_clamp.comp +12 -0
  1286. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
  1287. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +42 -0
  1288. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp +118 -0
  1289. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +221 -0
  1290. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp +246 -0
  1291. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_radix_select.comp +144 -0
  1292. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +42 -0
  1293. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +1938 -0
  1294. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
  1295. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +178 -0
  1296. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/utils.glsl +37 -0
  1297. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +1367 -0
  1298. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
  1299. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
  1300. data/vendor/ggml/src/ggml-webgpu/CMakeLists.txt +83 -0
  1301. data/vendor/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3457 -0
  1302. data/vendor/ggml/src/ggml-webgpu/ggml-webgpu.cpp +4790 -0
  1303. data/vendor/ggml/src/ggml-webgpu/pre_wgsl.hpp +808 -0
  1304. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  1305. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  1306. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +102 -0
  1307. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  1308. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
  1309. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +967 -0
  1310. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +92 -0
  1311. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +127 -0
  1312. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d_dw.wgsl +100 -0
  1313. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +84 -0
  1314. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  1315. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +89 -0
  1316. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +408 -0
  1317. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_decls.tmpl +134 -0
  1318. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_staging.tmpl +136 -0
  1319. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +240 -0
  1320. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  1321. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  1322. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +503 -0
  1323. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  1324. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +794 -0
  1325. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +178 -0
  1326. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +77 -0
  1327. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
  1328. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +1052 -0
  1329. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  1330. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  1331. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  1332. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +159 -0
  1333. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +210 -0
  1334. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +167 -0
  1335. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1560 -0
  1336. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
  1337. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  1338. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  1339. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
  1340. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +70 -0
  1341. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +151 -0
  1342. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rope.wgsl +228 -0
  1343. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +152 -0
  1344. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +65 -0
  1345. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  1346. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +108 -0
  1347. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  1348. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.wgsl +210 -0
  1349. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +120 -0
  1350. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  1351. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +247 -0
  1352. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  1353. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
  1354. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  1355. data/vendor/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  1356. data/vendor/ggml/src/ggml-zdnn/common.hpp +59 -0
  1357. data/vendor/ggml/src/ggml-zdnn/ggml-zdnn.cpp +638 -0
  1358. data/vendor/ggml/src/ggml-zdnn/mmf.cpp +80 -0
  1359. data/vendor/ggml/src/ggml-zdnn/mmf.hpp +12 -0
  1360. data/vendor/ggml/src/ggml-zdnn/utils.cpp +79 -0
  1361. data/vendor/ggml/src/ggml-zdnn/utils.hpp +19 -0
  1362. data/vendor/ggml/src/ggml-zendnn/CMakeLists.txt +91 -0
  1363. data/vendor/ggml/src/ggml-zendnn/ggml-zendnn.cpp +838 -0
  1364. data/vendor/ggml/src/ggml.c +8088 -0
  1365. data/vendor/ggml/src/ggml.cpp +26 -0
  1366. data/vendor/ggml/src/gguf.cpp +1706 -0
  1367. metadata +1459 -0
@@ -0,0 +1,1703 @@
1
+ #define GGML_COMMON_IMPL_CPP
2
+ #define GGML_COMMON_DECL_CPP
3
+ #include "ggml-common.h"
4
+ #include "ggml-backend-impl.h"
5
+
6
+ #include "ggml-impl.h"
7
+ #include "ggml-cpu.h"
8
+ #include "ggml-cpu-impl.h"
9
+ #include "simd-mappings.h"
10
+ #include "traits.h"
11
+
12
+ #include <cmath>
13
+ #include <cstring>
14
+ #include <cassert>
15
+ #include <cstdlib> // for qsort
16
+ #include <cstdio> // for GGML_ASSERT
17
+
18
+ #define GGML_CPU_CLANG_WORKAROUND
19
+ #include "../../repack.h"
20
+
21
+ #if defined(__GNUC__)
22
+ #pragma GCC diagnostic ignored "-Woverlength-strings"
23
+ #endif
24
+
25
+ #define UNUSED GGML_UNUSED
26
+
27
+ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
28
+ assert(QK8_0 == 32);
29
+ assert(k % QK8_0 == 0);
30
+ const int nb = k / QK8_0;
31
+
32
+ #if defined(__riscv_v_intrinsic)
33
+ block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
34
+ const size_t vl_calc = __riscv_vsetvl_e32m8(QK8_0);
35
+ const size_t vl_save = __riscv_vsetvl_e64m2(4);
36
+ vfloat32m1_t v_scalar_zero = __riscv_vfmv_s_f_f32m1(0.0f, __riscv_vsetvl_e32m1(1));
37
+
38
+ for (int i = 0; i < nb; i++) {
39
+ const float *x_block_base = x + i * QK8_0;
40
+ vint8m2_t q_r0, q_r1, q_r2, q_r3;
41
+ {
42
+ vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 0 * k, vl_calc);
43
+ vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
44
+ vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
45
+ float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
46
+
47
+ float d = amax / 127.0f;
48
+ y[i].d[0] = GGML_CPU_FP32_TO_FP16(d);
49
+
50
+ float id = d ? 1.0f / d : 0.0f;
51
+ vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
52
+ vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
53
+ q_r0 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
54
+ }
55
+ asm volatile ("" ::: "memory");
56
+
57
+ {
58
+ vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 1 * k, vl_calc);
59
+ vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
60
+ vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
61
+ float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
62
+
63
+ float d = amax / 127.0f;
64
+ y[i].d[1] = GGML_CPU_FP32_TO_FP16(d);
65
+ float id = d ? 1.0f / d : 0.0f;
66
+
67
+ vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
68
+ vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
69
+ q_r1 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
70
+ }
71
+ asm volatile ("" ::: "memory");
72
+ {
73
+ vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 2 * k, vl_calc);
74
+ vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
75
+ vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
76
+ float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
77
+
78
+ float d = amax / 127.0f;
79
+ y[i].d[2] = GGML_CPU_FP32_TO_FP16(d);
80
+ float id = d ? 1.0f / d : 0.0f;
81
+
82
+ vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
83
+ vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
84
+ q_r2 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
85
+ }
86
+ asm volatile ("" ::: "memory");
87
+ {
88
+ vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 3 * k, vl_calc);
89
+ vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
90
+ vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
91
+ float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
92
+
93
+ float d = amax / 127.0f;
94
+ y[i].d[3] = GGML_CPU_FP32_TO_FP16(d);
95
+ float id = d ? 1.0f / d : 0.0f;
96
+
97
+ vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
98
+ vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
99
+ q_r3 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
100
+ }
101
+ vint64m2_t v_q64_r0 = __riscv_vreinterpret_v_i8m2_i64m2(q_r0);
102
+ vint64m2_t v_q64_r1 = __riscv_vreinterpret_v_i8m2_i64m2(q_r1);
103
+ vint64m2_t v_q64_r2 = __riscv_vreinterpret_v_i8m2_i64m2(q_r2);
104
+ vint64m2_t v_q64_r3 = __riscv_vreinterpret_v_i8m2_i64m2(q_r3);
105
+ vint64m2x4_t v_quant_tuple = __riscv_vcreate_v_i64m2x4(v_q64_r0, v_q64_r1, v_q64_r2, v_q64_r3);
106
+ __riscv_vsseg4e64_v_i64m2x4((int64_t*)y[i].qs, v_quant_tuple, vl_save);
107
+ }
108
+ #else
109
+ UNUSED(nb);
110
+ ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
111
+ #endif
112
+ }
113
+
114
+ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
115
+ const int qk = QK8_0;
116
+ const int nb = n / qk;
117
+ const int ncols_interleaved = 8;
118
+ const int blocklen = 8;
119
+
120
+ assert (n % qk == 0);
121
+ assert (nc % ncols_interleaved == 0);
122
+
123
+ UNUSED(s);
124
+ UNUSED(bs);
125
+ UNUSED(vx);
126
+ UNUSED(vy);
127
+ UNUSED(nr);
128
+ UNUSED(nc);
129
+ UNUSED(nb);
130
+ UNUSED(ncols_interleaved);
131
+ UNUSED(blocklen);
132
+
133
+ #if defined __riscv_v
134
+ if (__riscv_vlenb() >= QK4_0) {
135
+ const size_t vl = QK4_0;
136
+
137
+ const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
138
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
139
+ const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
140
+
141
+ vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
142
+ for (int l = 0; l < nb; l++) {
143
+ const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
144
+ const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
145
+ const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
146
+ const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
147
+ __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
148
+ const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
149
+ const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
150
+ const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
151
+ const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
152
+
153
+ const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
154
+ const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
155
+ const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
156
+ const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
157
+ const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
158
+ const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
159
+ const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
160
+
161
+ const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
162
+ const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
163
+ const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
164
+ const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
165
+
166
+ const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
167
+ const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
168
+ const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
169
+ const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
170
+ const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
171
+ const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
172
+ const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
173
+ const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
174
+ const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
175
+ const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
176
+ const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
177
+ const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
178
+ const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
179
+
180
+ // vector version needs Zvfhmin extension
181
+ const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
182
+ const float b_scales[8] = {
183
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
184
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
185
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
186
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
187
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
188
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
189
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
190
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
191
+ };
192
+ const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
193
+ const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
194
+ sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
195
+ }
196
+ __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
197
+ }
198
+ return;
199
+ }
200
+
201
+ #endif
202
+ ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
203
+ }
204
+
205
+ #if defined __riscv_zvfh
206
+ void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
207
+ const int qk = QK8_0;
208
+ const int nb = n / qk;
209
+ const int ncols_interleaved = 16;
210
+ const int blocklen = 1;
211
+
212
+ assert (n % qk == 0);
213
+ assert (nc % ncols_interleaved == 0);
214
+
215
+ UNUSED(s);
216
+ UNUSED(bs);
217
+ UNUSED(vx);
218
+ UNUSED(vy);
219
+ UNUSED(nr);
220
+ UNUSED(nc);
221
+ UNUSED(nb);
222
+ UNUSED(ncols_interleaved);
223
+ UNUSED(blocklen);
224
+
225
+ const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
226
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
227
+ const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
228
+
229
+ // 1x16 Accumulator
230
+ vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16);
231
+
232
+ for (int l = 0; l < nb; l++) {
233
+ // 1x16 Integer Accumulator
234
+ vint16m1_t sumi_0_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
235
+ vint16m1_t sumi_0_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
236
+
237
+ // Accumulation loop.
238
+ for (int i = 0; i < QK4_0 / 2; i++) {
239
+ // Load `b_ptr`.
240
+ const vint8mf2_t b_0_packed = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16);
241
+ const vint8mf2_t b_0_lo = __riscv_vsra_vx_i8mf2(__riscv_vsll_vx_i8mf2(b_0_packed, 4, 16), 4, 16);
242
+ const vint8mf2_t b_0_hi = __riscv_vsra_vx_i8mf2(b_0_packed, 4, 16);
243
+
244
+ sumi_0_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_0_lo_16, a_ptr[l].qs[i], b_0_lo, 16);
245
+ sumi_0_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_0_hi_16, a_ptr[l].qs[16 + i], b_0_hi, 16);
246
+ }
247
+
248
+ const vint32m2_t sumi = __riscv_vwadd_vv_i32m2(sumi_0_lo_16, sumi_0_hi_16, 16);
249
+
250
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
251
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16);
252
+
253
+ sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16);
254
+ }
255
+
256
+ __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
257
+ }
258
+ }
259
+
260
+ void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
261
+ const int qk = QK_K;
262
+ const int nb = n / qk;
263
+ const int ncols_interleaved = 16;
264
+ const int blocklen = 1;
265
+
266
+ assert (n % qk == 0);
267
+ assert (nc % ncols_interleaved == 0);
268
+
269
+ UNUSED(s);
270
+ UNUSED(bs);
271
+ UNUSED(vx);
272
+ UNUSED(vy);
273
+ UNUSED(nr);
274
+ UNUSED(nc);
275
+ UNUSED(nb);
276
+ UNUSED(ncols_interleaved);
277
+ UNUSED(blocklen);
278
+
279
+ const block_q8_K * a_ptr = (const block_q8_K *) vy;
280
+
281
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
282
+ const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb);
283
+
284
+ // 1x16 Accumulator
285
+ vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16);
286
+
287
+ for (int l = 0; l < nb; l++) {
288
+ vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0, 16);
289
+
290
+ // Load `dmin`.
291
+ const vfloat32m2_t dmins_d = __riscv_vfmul_vf_f32m2(
292
+ __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16), a_ptr[l].d, 16);
293
+
294
+ // We process 4 sub-blocks at once.
295
+ for (int j = 0; j < QK_K / 128; j++) {
296
+ // Extract the scales and the mins.
297
+ //
298
+ // Low bits.
299
+ vuint8m2_t scales_mins_lo = __riscv_vle8_v_u8m2(&b_ptr[l].scales[j * 64], 64);
300
+ vuint8m2_t scales_lo = __riscv_vand_vx_u8m2(scales_mins_lo, 0x0F, 64);
301
+ vuint8m2_t mins_lo = __riscv_vsrl_vx_u8m2(scales_mins_lo, 4, 64);
302
+
303
+ // High bits.
304
+ vuint8m2_t scales_mins_hi = __riscv_vle8_v_u8m2(&b_ptr[l].scales[128], 64);
305
+ vuint8m2_t scales_hi;
306
+ vuint8m2_t mins_hi;
307
+ if (!j) {
308
+ scales_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x03, 64), 4, 64);
309
+ mins_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x0C, 64), 2, 64);
310
+ } else {
311
+ scales_hi = __riscv_vand_vx_u8m2(scales_mins_hi, 0x30, 64);
312
+ mins_hi = __riscv_vsrl_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0xC0, 64), 2, 64);
313
+ }
314
+ vuint16m4_t scales = __riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(scales_hi, scales_lo, 64), 64);
315
+ vint16m4_t mins = __riscv_vreinterpret_v_u16m4_i16m4(__riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(mins_hi, mins_lo, 64), 64));
316
+
317
+ // Reduce the mins and multiply with `dmin`.
318
+ //
319
+ // Correct in `sumf`.
320
+ vint32m2_t bsums = __riscv_vmv_v_x_i32m2(0, 16);
321
+ bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8] + a_ptr[l].bsums[j * 8 + 1], __riscv_vget_v_i16m4_i16m1(mins, 0), 16);
322
+ bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 2] + a_ptr[l].bsums[j * 8 + 3], __riscv_vget_v_i16m4_i16m1(mins, 1), 16);
323
+ bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 4] + a_ptr[l].bsums[j * 8 + 5], __riscv_vget_v_i16m4_i16m1(mins, 2), 16);
324
+ bsums = __riscv_vwmacc_vx_i32m2(bsums, a_ptr[l].bsums[j * 8 + 6] + a_ptr[l].bsums[j * 8 + 7], __riscv_vget_v_i16m4_i16m1(mins, 3), 16);
325
+
326
+ sumf = __riscv_vfsub_vv_f32m2(sumf, __riscv_vfmul_vv_f32m2(dmins_d, __riscv_vfcvt_f_x_v_f32m2(bsums, 16), 16), 16);
327
+
328
+ // Accumulation for 2 sub-blocks.
329
+ //
330
+ // This might overflow, so we accumulate in two steps.
331
+ //
332
+ // Recheck.
333
+ for (int k = 0; k < 2; k++) {
334
+ vint16m1_t sumi_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
335
+ vint16m1_t sumi_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
336
+
337
+ for (int i = k * 16; i < k * 16 + QK4_0 / 2; i++) {
338
+ // Load `b_ptr`.
339
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + i * 16], 16);
340
+ const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16));
341
+ const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16));
342
+
343
+ sumi_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_s_0_16, a_ptr[l].qs[j * 128 + i], b_s_0, 16);
344
+ sumi_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_s_1_16, a_ptr[l].qs[j * 128 + 32 + i], b_s_1, 16);
345
+ }
346
+
347
+ sumi = __riscv_vwmacc_vv_i32m2(sumi,
348
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)),
349
+ sumi_s_0_16, 16);
350
+ sumi = __riscv_vwmacc_vv_i32m2(sumi,
351
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)),
352
+ sumi_s_1_16, 16);
353
+ }
354
+ // Accumulation for 2 sub-blocks.
355
+ //
356
+ // This might overflow, so we accumulate in two steps.
357
+ //
358
+ // Recheck.
359
+ for (int k = 0; k < 2; k++) {
360
+ vint16m1_t sumi_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
361
+ vint16m1_t sumi_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
362
+
363
+ for (int i = k * 16; i < k * 16 + QK4_0 / 2; i++) {
364
+ // Load `b_ptr`.
365
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + 512 + i * 16], 16);
366
+ const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16));
367
+ const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16));
368
+
369
+ sumi_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_s_0_16, a_ptr[l].qs[j * 128 + 64 + i], b_s_0, 16);
370
+ sumi_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_s_1_16, a_ptr[l].qs[j * 128 + 96 + i], b_s_1, 16);
371
+ }
372
+
373
+ sumi = __riscv_vwmacc_vv_i32m2(sumi,
374
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)),
375
+ sumi_s_0_16, 16);
376
+ sumi = __riscv_vwmacc_vv_i32m2(sumi,
377
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)),
378
+ sumi_s_1_16, 16);
379
+ }
380
+ }
381
+
382
+ const vfloat32m2_t b_d = __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)&b_ptr[l].d[0], 16), 16);
383
+ const vfloat32m2_t d_0 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d, 16);
384
+
385
+ sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16);
386
+ }
387
+
388
+ __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
389
+ }
390
+ }
391
+
392
+ void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
393
+ const int qk = QK8_0;
394
+ const int nb = n / qk;
395
+ const int ncols_interleaved = 16;
396
+ const int blocklen = 1;
397
+
398
+ assert (n % qk == 0);
399
+ assert (nc % ncols_interleaved == 0);
400
+
401
+ UNUSED(s);
402
+ UNUSED(bs);
403
+ UNUSED(vx);
404
+ UNUSED(vy);
405
+ UNUSED(nr);
406
+ UNUSED(nc);
407
+ UNUSED(nb);
408
+ UNUSED(ncols_interleaved);
409
+ UNUSED(blocklen);
410
+
411
+ const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
412
+ const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
413
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
414
+ const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb);
415
+
416
+ // 1x16 Accumulator1
417
+ vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16);
418
+
419
+ for (int l = 0; l < nb; l++) {
420
+ // 1x16 integer accumulator
421
+ vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0.0f, 16);
422
+
423
+ // Accumulation loop.
424
+ for (int i = 0; i < QK4_NL / 2; i++) {
425
+ // Load `b_ptr`.
426
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2((const uint8_t *)&b_ptr[l].qs[i * 16], 16);
427
+ const vint8mf2_t b_0_lo = __riscv_vrgather_vv_i8mf2(values, __riscv_vand_vx_u8mf2(b_0_packed, 0xf, 16), 16);
428
+ const vint8mf2_t b_0_hi = __riscv_vrgather_vv_i8mf2(values, __riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16), 16);
429
+ // const vint16m1_t b_0_lo_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_lo, 16);
430
+ // const vint16m1_t b_0_hi_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_hi, 16);
431
+
432
+ const vint16m1_t sumi_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i], 16);
433
+ const vint16m1_t sumi_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[16 + i], 16);
434
+ sumi = __riscv_vadd_vv_i32m2(sumi, __riscv_vwadd_vv_i32m2(sumi_lo, sumi_hi, 16), 16);
435
+ }
436
+
437
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
438
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16);
439
+
440
+ sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16);
441
+ }
442
+
443
+ __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
444
+ }
445
+ }
446
+
447
+ void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
448
+ const int qk = QK8_0;
449
+ const int nb = n / qk;
450
+ const int ncols_interleaved = 16;
451
+ const int blocklen = 1;
452
+
453
+ assert (n % qk == 0);
454
+ assert (nc % ncols_interleaved == 0);
455
+
456
+ UNUSED(s);
457
+ UNUSED(bs);
458
+ UNUSED(vx);
459
+ UNUSED(vy);
460
+ UNUSED(nr);
461
+ UNUSED(nc);
462
+ UNUSED(nb);
463
+ UNUSED(ncols_interleaved);
464
+ UNUSED(blocklen);
465
+ UNUSED(bs);
466
+
467
+ const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
468
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
469
+ const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
470
+
471
+ // 1x16 Accumulator
472
+ vfloat32m2_t sumf = __riscv_vfmv_v_f_f32m2(0.0f, 16);
473
+
474
+ for (int l = 0; l < nb; l++) {
475
+ // 1x16 Integer Accumulator
476
+ vint32m2_t sumi = __riscv_vmv_v_x_i32m2(0.0f, 16);
477
+
478
+ // Accumulation loop.
479
+ for (int i = 0; i < QK8_0; i++) {
480
+ // Load `b_ptr`.
481
+ const vint8mf2_t b_0 = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16);
482
+ // const vint16m1_t b_0_16 = __riscv_vwcvt_x_x_v_i16m1(b_0, 16);
483
+
484
+ sumi = __riscv_vwadd_wv_i32m2(sumi, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i], 16), 16);
485
+ }
486
+
487
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
488
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d, 16);
489
+
490
+ sumf = __riscv_vfmacc_vv_f32m2(sumf, __riscv_vfcvt_f_x_v_f32m2(sumi, 16), d_0, 16);
491
+ }
492
+
493
+ __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
494
+ }
495
+ }
496
+
497
+ void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
498
+ assert(n % QK_K == 0);
499
+ assert(nr == 1);
500
+ assert(nc % 16 == 0);
501
+
502
+ UNUSED(bs);
503
+
504
+ const int N_COLS_TILE = 16;
505
+ const int num_k_blocks = n / QK_K;
506
+
507
+ const size_t vl = __riscv_vsetvl_e32m2(N_COLS_TILE);
508
+ for (int col_tile = 0; col_tile < nc; col_tile += N_COLS_TILE) {
509
+
510
+ const block_q8_K* lhs_base_ptr = (const block_q8_K*)vy;
511
+ const block_q2_Kx16* rhs_base_ptr = (const block_q2_Kx16*)vx + (col_tile / N_COLS_TILE) * num_k_blocks;
512
+
513
+ vfloat32m2_t v_sumf = __riscv_vfmv_v_f_f32m2(0.0f, vl);
514
+
515
+ for (int k_block = 0; k_block < num_k_blocks; ++k_block) {
516
+ const block_q8_K* lhs_current = &lhs_base_ptr[k_block];
517
+ const block_q2_Kx16* rhs_current = &rhs_base_ptr[k_block];
518
+
519
+ // 1. Prepare Global Min Scales
520
+ vfloat16m1_t v_g_min_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->dmin, vl);
521
+ vfloat32m2_t v_g_min_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_min_f16, vl);
522
+
523
+ vfloat32m2_t v_g_min_final = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d, vl);
524
+
525
+ vint32m2_t v_isum = __riscv_vmv_v_x_i32m2(0, vl);
526
+
527
+ const uint8_t* rhs_qs_ptr = rhs_current->qs;
528
+ const uint8_t* rhs_sc_ptr = rhs_current->scales;
529
+ const int8_t* lhs_qs_ptr = lhs_current->qs;
530
+
531
+ // --- Phase Loop (4 phases x 64 elements) ---
532
+ for (int phase = 0; phase < 4; ++phase) {
533
+
534
+ // A. Load Scales/Mins
535
+ vuint16m1_t v_d_sb_0, v_d_sb_1, v_d_sb_2, v_d_sb_3;
536
+ vuint16m1_t v_m_sb_0, v_m_sb_1, v_m_sb_2, v_m_sb_3;
537
+
538
+ {
539
+ vuint8mf2_t v_raw;
540
+ // Sub-block 0
541
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 0, vl);
542
+ v_d_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
543
+ v_m_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
544
+
545
+ // Sub-block 1
546
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 16, vl);
547
+ v_d_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
548
+ v_m_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
549
+
550
+ // Sub-block 2
551
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 32, vl);
552
+ v_d_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
553
+ v_m_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
554
+
555
+ // Sub-block 3
556
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 48, vl);
557
+ v_d_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
558
+ v_m_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
559
+
560
+ rhs_sc_ptr += 64;
561
+ }
562
+
563
+ int base_k_phase = (phase < 2) ? (phase * 16) : (128 + (phase-2)*16);
564
+ int k_offsets[4] = {0, 32, 64, 96};
565
+
566
+ // B. Inner Dot Product Loop
567
+ for (int l = 0; l < 16; ++l) {
568
+ vuint8mf2_t v_rhs_data = __riscv_vle8_v_u8mf2(rhs_qs_ptr, vl);
569
+ rhs_qs_ptr += 16;
570
+
571
+ // Sub-block 0
572
+ {
573
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(v_rhs_data, 3, vl);
574
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
575
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
576
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_0), vl);
577
+
578
+ int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[0] + l];
579
+ v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl);
580
+ }
581
+ // Sub-block 1
582
+ {
583
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 2, vl), 3, vl);
584
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
585
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
586
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_1), vl);
587
+
588
+ int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[1] + l];
589
+ v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl);
590
+ }
591
+ // Sub-block 2
592
+ {
593
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 4, vl), 3, vl);
594
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
595
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
596
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_2), vl);
597
+
598
+ int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[2] + l];
599
+ v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl);
600
+ }
601
+ // Sub-block 3
602
+ {
603
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 6, vl), 3, vl);
604
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
605
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
606
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_3), vl);
607
+
608
+ int8_t q8 = lhs_qs_ptr[base_k_phase + k_offsets[3] + l];
609
+ v_isum = __riscv_vwmacc_vx_i32m2(v_isum, (int16_t)q8, v_w, vl);
610
+ }
611
+ }
612
+
613
+ // correction
614
+ int sb_base_abs = base_k_phase / 16;
615
+
616
+ // Sub-block 0
617
+ {
618
+ int sb_idx = sb_base_abs + (k_offsets[0] / 16);
619
+ int16_t bsum = lhs_current->bsums[sb_idx];
620
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_0);
621
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl);
622
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl);
623
+ v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl);
624
+ }
625
+ // Sub-block 1
626
+ {
627
+ int sb_idx = sb_base_abs + (k_offsets[1] / 16);
628
+ int16_t bsum = lhs_current->bsums[sb_idx];
629
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_1);
630
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl);
631
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl);
632
+ v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl);
633
+ }
634
+ // Sub-block 2
635
+ {
636
+ int sb_idx = sb_base_abs + (k_offsets[2] / 16);
637
+ int16_t bsum = lhs_current->bsums[sb_idx];
638
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_2);
639
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl);
640
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl);
641
+ v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl);
642
+ }
643
+ // Sub-block 3
644
+ {
645
+ int sb_idx = sb_base_abs + (k_offsets[3] / 16);
646
+ int16_t bsum = lhs_current->bsums[sb_idx];
647
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_3);
648
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, bsum, vl);
649
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min_final, vl);
650
+ v_sumf = __riscv_vfsub_vv_f32m2(v_sumf, vf_c, vl);
651
+ }
652
+
653
+ } // End Phase Loop
654
+
655
+ // Apply global Scales
656
+ vfloat16m1_t v_g_all_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->d, vl);
657
+ vfloat32m2_t v_g_all_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_all_f16, vl);
658
+
659
+ vfloat32m2_t v_g_all_final = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d, vl);
660
+ vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum, vl);
661
+ v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all_final, vl);
662
+ v_sumf = __riscv_vfadd_vv_f32m2(v_sumf, v_sum, vl);
663
+
664
+ } // End K-Block
665
+ __riscv_vse32_v_f32m2(s + col_tile, v_sumf, vl);
666
+ }
667
+ }
668
+ #endif
669
+
670
+ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
671
+ const int qk = QK8_0;
672
+ const int nb = n / qk;
673
+ const int ncols_interleaved = 8;
674
+ const int blocklen = 8;
675
+
676
+ assert (n % qk == 0);
677
+ assert (nr % 4 == 0);
678
+ assert (nc % ncols_interleaved == 0);
679
+
680
+ UNUSED(s);
681
+ UNUSED(bs);
682
+ UNUSED(vx);
683
+ UNUSED(vy);
684
+ UNUSED(nr);
685
+ UNUSED(nc);
686
+ UNUSED(nb);
687
+ UNUSED(ncols_interleaved);
688
+ UNUSED(blocklen);
689
+
690
+ #if defined __riscv_v
691
+ if (__riscv_vlenb() >= QK4_0) {
692
+ const size_t vl = QK4_0;
693
+
694
+ for (int y = 0; y < nr / 4; y++) {
695
+ const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
696
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
697
+ const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
698
+ vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
699
+ vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
700
+ vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
701
+ vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
702
+ for (int l = 0; l < nb; l++) {
703
+ const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
704
+ const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
705
+ const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
706
+ const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
707
+ const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
708
+ const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
709
+ const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
710
+
711
+ // vector version needs Zvfhmin extension
712
+ const float a_scales[4] = {
713
+ GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
714
+ GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
715
+ GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
716
+ GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
717
+ };
718
+ const float b_scales[8] = {
719
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
720
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
721
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
722
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
723
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
724
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
725
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
726
+ GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
727
+ };
728
+ const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
729
+
730
+ const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
731
+ const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
732
+ const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
733
+ const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
734
+ __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
735
+ vint16m4_t sumi_l0;
736
+ {
737
+ const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
738
+ const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
739
+ const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
740
+ const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
741
+ const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
742
+ const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
743
+ const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
744
+ const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
745
+
746
+ sumi_l0 = sumi_hi_m;
747
+ }
748
+
749
+ {
750
+ const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
751
+ const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
752
+ const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
753
+ const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
754
+ const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
755
+ const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
756
+ const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
757
+ const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
758
+ const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
759
+ const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
760
+ const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
761
+ const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
762
+ const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
763
+
764
+ const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
765
+ sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
766
+ }
767
+
768
+ const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
769
+ const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
770
+ const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
771
+ const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
772
+ __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
773
+ vint16m4_t sumi_l1;
774
+ {
775
+ const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
776
+ const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
777
+ const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
778
+ const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
779
+ const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
780
+ const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
781
+ const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
782
+ const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
783
+
784
+ sumi_l1 = sumi_hi_m;
785
+ }
786
+
787
+ {
788
+ const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
789
+ const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
790
+ const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
791
+ const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
792
+ const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
793
+ const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
794
+ const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
795
+ const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
796
+ const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
797
+ const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
798
+ const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
799
+ const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
800
+ const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
801
+
802
+ const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
803
+ sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
804
+ }
805
+
806
+ const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
807
+ const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
808
+ const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
809
+ const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
810
+ __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
811
+ vint16m4_t sumi_l2;
812
+ {
813
+ const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
814
+ const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
815
+ const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
816
+ const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
817
+ const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
818
+ const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
819
+ const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
820
+ const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
821
+
822
+ sumi_l2 = sumi_hi_m;
823
+ }
824
+
825
+ {
826
+ const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
827
+ const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
828
+ const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
829
+ const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
830
+ const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
831
+ const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
832
+ const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
833
+ const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
834
+ const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
835
+ const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
836
+ const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
837
+ const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
838
+ const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
839
+
840
+ const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
841
+ sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
842
+ }
843
+
844
+ const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
845
+ const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
846
+ const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
847
+ const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
848
+ __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
849
+ vint16m4_t sumi_l3;
850
+ {
851
+ const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
852
+ const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
853
+ const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
854
+ const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
855
+ const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
856
+ const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
857
+ const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
858
+ const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
859
+
860
+ sumi_l3 = sumi_hi_m;
861
+ }
862
+
863
+ {
864
+ const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
865
+ const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
866
+ const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
867
+ const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
868
+ const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
869
+ const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
870
+ const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
871
+ const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
872
+ const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
873
+ const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
874
+ const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
875
+ const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
876
+ const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
877
+
878
+ const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
879
+ sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
880
+ }
881
+ }
882
+ __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
883
+ __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
884
+ __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
885
+ __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
886
+ }
887
+ }
888
+
889
+ return;
890
+ }
891
+
892
+ #endif
893
+ ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
894
+ }
895
+
896
+ #if defined __riscv_zvfh
897
+ void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
898
+ const int qk = QK8_0;
899
+ const int nb = n / qk;
900
+ const int ncols_interleaved = 16;
901
+ const int blocklen = 1;
902
+
903
+ assert (n % qk == 0);
904
+ assert (nr % 4 == 0);
905
+ assert (nc % ncols_interleaved == 0);
906
+
907
+ UNUSED(s);
908
+ UNUSED(bs);
909
+ UNUSED(vx);
910
+ UNUSED(vy);
911
+ UNUSED(nr);
912
+ UNUSED(nc);
913
+ UNUSED(nb);
914
+ UNUSED(ncols_interleaved);
915
+ UNUSED(blocklen);
916
+
917
+ for (int y = 0; y < nr / 4; y++) {
918
+ const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
919
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
920
+ const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
921
+
922
+ // 4x16 Accumulators
923
+ vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
924
+ vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
925
+ vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
926
+ vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
927
+
928
+ for (int l = 0; l < nb; l++) {
929
+ // 4x16 integer accumulators
930
+ vint16m1_t sumi_0_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
931
+ vint16m1_t sumi_1_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
932
+ vint16m1_t sumi_2_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
933
+ vint16m1_t sumi_3_lo_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
934
+ vint16m1_t sumi_0_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
935
+ vint16m1_t sumi_1_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
936
+ vint16m1_t sumi_2_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
937
+ vint16m1_t sumi_3_hi_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
938
+
939
+ // Accumulation loop.
940
+ for (int i = 0; i < QK4_0 / 2; i++) {
941
+ // Load `b_ptr`.
942
+ const vint8mf2_t b_0_packed = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16);
943
+ const vint8mf2_t b_0_lo = __riscv_vsra_vx_i8mf2(__riscv_vsll_vx_i8mf2(b_0_packed, 4, 16), 4, 16);
944
+ const vint8mf2_t b_0_hi = __riscv_vsra_vx_i8mf2(b_0_packed, 4, 16);
945
+
946
+ sumi_0_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_0_lo_16, a_ptr[l].qs[i * 4], b_0_lo, 16);
947
+ sumi_1_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_1_lo_16, a_ptr[l].qs[i * 4 + 1], b_0_lo, 16);
948
+ sumi_2_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_2_lo_16, a_ptr[l].qs[i * 4 + 2], b_0_lo, 16);
949
+ sumi_3_lo_16 = __riscv_vwmacc_vx_i16m1(sumi_3_lo_16, a_ptr[l].qs[i * 4 + 3], b_0_lo, 16);
950
+
951
+ sumi_0_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_0_hi_16, a_ptr[l].qs[64 + i * 4], b_0_hi, 16);
952
+ sumi_1_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_1_hi_16, a_ptr[l].qs[64 + i * 4 + 1], b_0_hi, 16);
953
+ sumi_2_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_2_hi_16, a_ptr[l].qs[64 + i * 4 + 2], b_0_hi, 16);
954
+ sumi_3_hi_16 = __riscv_vwmacc_vx_i16m1(sumi_3_hi_16, a_ptr[l].qs[64 + i * 4 + 3], b_0_hi, 16);
955
+ }
956
+
957
+ // Do the final accumulation in i32 to prevent overflow.
958
+ const vint32m2_t sumi_0 = __riscv_vwadd_vv_i32m2(sumi_0_lo_16, sumi_0_hi_16, 16);
959
+ const vint32m2_t sumi_1 = __riscv_vwadd_vv_i32m2(sumi_1_lo_16, sumi_1_hi_16, 16);
960
+ const vint32m2_t sumi_2 = __riscv_vwadd_vv_i32m2(sumi_2_lo_16, sumi_2_hi_16, 16);
961
+ const vint32m2_t sumi_3 = __riscv_vwadd_vv_i32m2(sumi_3_lo_16, sumi_3_hi_16, 16);
962
+
963
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
964
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16);
965
+ const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16);
966
+ const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16);
967
+ const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16);
968
+
969
+ sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16);
970
+ sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16);
971
+ sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16);
972
+ sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16);
973
+ }
974
+
975
+ __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16);
976
+ __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16);
977
+ __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16);
978
+ __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
979
+ }
980
+ }
981
+ }
982
+
983
+ void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
984
+ const int qk = QK_K;
985
+ const int nb = n / qk;
986
+ const int ncols_interleaved = 16;
987
+ const int blocklen = 1;
988
+
989
+ assert (n % qk == 0);
990
+ assert (nr % 4 == 0);
991
+ assert (nc % ncols_interleaved == 0);
992
+
993
+ UNUSED(s);
994
+ UNUSED(bs);
995
+ UNUSED(vx);
996
+ UNUSED(vy);
997
+ UNUSED(nr);
998
+ UNUSED(nc);
999
+ UNUSED(nb);
1000
+ UNUSED(ncols_interleaved);
1001
+ UNUSED(blocklen);
1002
+
1003
+ for (int y = 0; y < nr / 4; y++) {
1004
+ const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
1005
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
1006
+ const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb);
1007
+
1008
+ // 4x16 Accumulators
1009
+ vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1010
+ vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1011
+ vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1012
+ vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1013
+
1014
+ for (int l = 0; l < nb; l++) {
1015
+ vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0, 16);
1016
+ vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0, 16);
1017
+ vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0, 16);
1018
+ vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0, 16);
1019
+
1020
+ // Load `dmin`.
1021
+ const vfloat32m2_t dmins = __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].dmin, 16), 16);
1022
+
1023
+ // We process 4 sub-blocks at once.
1024
+ for (int j = 0; j < QK_K / 128; j++) {
1025
+ // Extract the scales and the mins.
1026
+ //
1027
+ // Low bits.
1028
+ vuint8m2_t scales_mins_lo = __riscv_vle8_v_u8m2(&b_ptr[l].scales[j * 64], 64);
1029
+ vuint8m2_t scales_lo = __riscv_vand_vx_u8m2(scales_mins_lo, 0x0F, 64);
1030
+ vuint8m2_t mins_lo = __riscv_vsrl_vx_u8m2(scales_mins_lo, 4, 64);
1031
+
1032
+ // High bits.
1033
+ vuint8m2_t scales_mins_hi = __riscv_vle8_v_u8m2(&b_ptr[l].scales[128], 64);
1034
+ vuint8m2_t scales_hi;
1035
+ vuint8m2_t mins_hi;
1036
+ if (!j) {
1037
+ scales_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x03, 64), 4, 64);
1038
+ mins_hi = __riscv_vsll_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0x0C, 64), 2, 64);
1039
+ } else {
1040
+ scales_hi = __riscv_vand_vx_u8m2(scales_mins_hi, 0x30, 64);
1041
+ mins_hi = __riscv_vsrl_vx_u8m2(__riscv_vand_vx_u8m2(scales_mins_hi, 0xC0, 64), 2, 64);
1042
+ }
1043
+ vuint16m4_t scales = __riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(scales_hi, scales_lo, 64), 64);
1044
+ vint16m4_t mins = __riscv_vreinterpret_v_u16m4_i16m4(__riscv_vzext_vf2_u16m4(__riscv_vor_vv_u8m2(mins_hi, mins_lo, 64), 64));
1045
+
1046
+ // Reduce the mins and multiply with `dmin`.
1047
+ //
1048
+ // Correct in `sumf`.
1049
+ vint32m2_t bsums_0 = __riscv_vmv_v_x_i32m2(0, 16);
1050
+ vint32m2_t bsums_1 = __riscv_vmv_v_x_i32m2(0, 16);
1051
+ vint32m2_t bsums_2 = __riscv_vmv_v_x_i32m2(0, 16);
1052
+ vint32m2_t bsums_3 = __riscv_vmv_v_x_i32m2(0, 16);
1053
+
1054
+ bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0,
1055
+ a_ptr[l].bsums[j * 32] + a_ptr[l].bsums[j * 32 + 4],
1056
+ __riscv_vget_v_i16m4_i16m1(mins, 0), 16);
1057
+ bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1,
1058
+ a_ptr[l].bsums[j * 32 + 1] + a_ptr[l].bsums[j * 32 + 5],
1059
+ __riscv_vget_v_i16m4_i16m1(mins, 0), 16);
1060
+ bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2,
1061
+ a_ptr[l].bsums[j * 32 + 2] + a_ptr[l].bsums[j * 32 + 6],
1062
+ __riscv_vget_v_i16m4_i16m1(mins, 0), 16);
1063
+ bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3,
1064
+ a_ptr[l].bsums[j * 32 + 3] + a_ptr[l].bsums[j * 32 + 7],
1065
+ __riscv_vget_v_i16m4_i16m1(mins, 0), 16);
1066
+ bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0,
1067
+ a_ptr[l].bsums[j * 32 + 8] + a_ptr[l].bsums[j * 32 + 8 + 4],
1068
+ __riscv_vget_v_i16m4_i16m1(mins, 1), 16);
1069
+ bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1,
1070
+ a_ptr[l].bsums[j * 32 + 8 + 1] + a_ptr[l].bsums[j * 32 + 8 + 5],
1071
+ __riscv_vget_v_i16m4_i16m1(mins, 1), 16);
1072
+ bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2,
1073
+ a_ptr[l].bsums[j * 32 + 8 + 2] + a_ptr[l].bsums[j * 32 + 8 + 6],
1074
+ __riscv_vget_v_i16m4_i16m1(mins, 1), 16);
1075
+ bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3,
1076
+ a_ptr[l].bsums[j * 32 + 8 + 3] + a_ptr[l].bsums[j * 32 + 8 + 7],
1077
+ __riscv_vget_v_i16m4_i16m1(mins, 1), 16);
1078
+ bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0,
1079
+ a_ptr[l].bsums[j * 32 + 16] + a_ptr[l].bsums[j * 32 + 16 + 4],
1080
+ __riscv_vget_v_i16m4_i16m1(mins, 2), 16);
1081
+ bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1,
1082
+ a_ptr[l].bsums[j * 32 + 16 + 1] + a_ptr[l].bsums[j * 32 + 16 + 5],
1083
+ __riscv_vget_v_i16m4_i16m1(mins, 2), 16);
1084
+ bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2,
1085
+ a_ptr[l].bsums[j * 32 + 16 + 2] + a_ptr[l].bsums[j * 32 + 16 + 6],
1086
+ __riscv_vget_v_i16m4_i16m1(mins, 2), 16);
1087
+ bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3,
1088
+ a_ptr[l].bsums[j * 32 + 16 + 3] + a_ptr[l].bsums[j * 32 + 16 + 7],
1089
+ __riscv_vget_v_i16m4_i16m1(mins, 2), 16);
1090
+ bsums_0 = __riscv_vwmacc_vx_i32m2(bsums_0,
1091
+ a_ptr[l].bsums[j * 32 + 24 + 0] + a_ptr[l].bsums[j * 32 + 24 + 4],
1092
+ __riscv_vget_v_i16m4_i16m1(mins, 3), 16);
1093
+ bsums_1 = __riscv_vwmacc_vx_i32m2(bsums_1,
1094
+ a_ptr[l].bsums[j * 32 + 24 + 1] + a_ptr[l].bsums[j * 32 + 24 + 5],
1095
+ __riscv_vget_v_i16m4_i16m1(mins, 3), 16);
1096
+ bsums_2 = __riscv_vwmacc_vx_i32m2(bsums_2,
1097
+ a_ptr[l].bsums[j * 32 + 24 + 2] + a_ptr[l].bsums[j * 32 + 24 + 6],
1098
+ __riscv_vget_v_i16m4_i16m1(mins, 3), 16);
1099
+ bsums_3 = __riscv_vwmacc_vx_i32m2(bsums_3,
1100
+ a_ptr[l].bsums[j * 32 + 24 + 3] + a_ptr[l].bsums[j * 32 + 24 + 7],
1101
+ __riscv_vget_v_i16m4_i16m1(mins, 3), 16);
1102
+
1103
+ const vfloat32m2_t dmins_d_0 = __riscv_vfmul_vf_f32m2(dmins, a_ptr[l].d[0], 16);
1104
+ const vfloat32m2_t dmins_d_1 = __riscv_vfmul_vf_f32m2(dmins, a_ptr[l].d[1], 16);
1105
+ const vfloat32m2_t dmins_d_2 = __riscv_vfmul_vf_f32m2(dmins, a_ptr[l].d[2], 16);
1106
+ const vfloat32m2_t dmins_d_3 = __riscv_vfmul_vf_f32m2(dmins, a_ptr[l].d[3], 16);
1107
+
1108
+ sumf_0 = __riscv_vfsub_vv_f32m2(sumf_0, __riscv_vfmul_vv_f32m2(dmins_d_0, __riscv_vfcvt_f_x_v_f32m2(bsums_0, 16), 16), 16);
1109
+ sumf_1 = __riscv_vfsub_vv_f32m2(sumf_1, __riscv_vfmul_vv_f32m2(dmins_d_1, __riscv_vfcvt_f_x_v_f32m2(bsums_1, 16), 16), 16);
1110
+ sumf_2 = __riscv_vfsub_vv_f32m2(sumf_2, __riscv_vfmul_vv_f32m2(dmins_d_2, __riscv_vfcvt_f_x_v_f32m2(bsums_2, 16), 16), 16);
1111
+ sumf_3 = __riscv_vfsub_vv_f32m2(sumf_3, __riscv_vfmul_vv_f32m2(dmins_d_3, __riscv_vfcvt_f_x_v_f32m2(bsums_3, 16), 16), 16);
1112
+
1113
+
1114
+ // Accumulation for 2 sub-blocks.
1115
+ //
1116
+ // This might overflow, so we accumulate in two steps.
1117
+ //
1118
+ // Recheck.
1119
+ for (int k = 0; k < 2; k++) {
1120
+ // 4x16 integer accumulators
1121
+ vint16m1_t sumi_0_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1122
+ vint16m1_t sumi_1_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1123
+ vint16m1_t sumi_2_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1124
+ vint16m1_t sumi_3_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1125
+ vint16m1_t sumi_0_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1126
+ vint16m1_t sumi_1_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1127
+ vint16m1_t sumi_2_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1128
+ vint16m1_t sumi_3_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1129
+
1130
+ for (int i = k * 16; i < k * 16 + QK4_0 / 2; i++) {
1131
+ // Load `b_ptr`.
1132
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + i * 16], 16);
1133
+ const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16));
1134
+ const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16));
1135
+
1136
+ sumi_0_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_0_16, a_ptr[l].qs[j * 512 + i * 4], b_s_0, 16);
1137
+ sumi_1_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 1], b_s_0, 16);
1138
+ sumi_2_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 2], b_s_0, 16);
1139
+ sumi_3_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_0_16, a_ptr[l].qs[j * 512 + i * 4 + 3], b_s_0, 16);
1140
+
1141
+ sumi_0_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4], b_s_1, 16);
1142
+ sumi_1_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 1], b_s_1, 16);
1143
+ sumi_2_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 2], b_s_1, 16);
1144
+ sumi_3_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_1_16, a_ptr[l].qs[j * 512 + 128 + i * 4 + 3], b_s_1, 16);
1145
+ }
1146
+
1147
+ sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0,
1148
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)),
1149
+ sumi_0_s_0_16, 16);
1150
+ sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0,
1151
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)),
1152
+ sumi_0_s_1_16, 16);
1153
+ sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1,
1154
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)),
1155
+ sumi_1_s_0_16, 16);
1156
+ sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1,
1157
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)),
1158
+ sumi_1_s_1_16, 16);
1159
+ sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2,
1160
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)),
1161
+ sumi_2_s_0_16, 16);
1162
+ sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2,
1163
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)),
1164
+ sumi_2_s_1_16, 16);
1165
+ sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3,
1166
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 0)),
1167
+ sumi_3_s_0_16, 16);
1168
+ sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3,
1169
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 1)),
1170
+ sumi_3_s_1_16, 16);
1171
+ }
1172
+ // Accumulation for 2 sub-blocks.
1173
+ //
1174
+ // This might overflow, so we accumulate in two steps.
1175
+ //
1176
+ // Recheck.
1177
+ for (int k = 0; k < 2; k++) {
1178
+ // 4x16 integer accumulators
1179
+ vint16m1_t sumi_0_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1180
+ vint16m1_t sumi_1_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1181
+ vint16m1_t sumi_2_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1182
+ vint16m1_t sumi_3_s_0_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1183
+ vint16m1_t sumi_0_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1184
+ vint16m1_t sumi_1_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1185
+ vint16m1_t sumi_2_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1186
+ vint16m1_t sumi_3_s_1_16 = __riscv_vmv_v_x_i16m1(0.0f, 16);
1187
+
1188
+ for (int i = k * 16; i < k * 16 + QK4_0 / 2; i++) {
1189
+ // Load `b_ptr`.
1190
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2(&b_ptr[l].qs[j * 1024 + 512 + i * 16], 16);
1191
+ const vint8mf2_t b_s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(b_0_packed, 0xF, 16));
1192
+ const vint8mf2_t b_s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16));
1193
+
1194
+ sumi_0_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4], b_s_0, 16);
1195
+ sumi_1_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 1], b_s_0, 16);
1196
+ sumi_2_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 2], b_s_0, 16);
1197
+ sumi_3_s_0_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_0_16, a_ptr[l].qs[j * 512 + 256 + i * 4 + 3], b_s_0, 16);
1198
+
1199
+ sumi_0_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_0_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4], b_s_1, 16);
1200
+ sumi_1_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_1_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 1], b_s_1, 16);
1201
+ sumi_2_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_2_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 2], b_s_1, 16);
1202
+ sumi_3_s_1_16 = __riscv_vwmacc_vx_i16m1(sumi_3_s_1_16, a_ptr[l].qs[j * 512 + 384 + i * 4 + 3], b_s_1, 16);
1203
+ }
1204
+
1205
+ sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0,
1206
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)),
1207
+ sumi_0_s_0_16, 16);
1208
+ sumi_0 = __riscv_vwmacc_vv_i32m2(sumi_0,
1209
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)),
1210
+ sumi_0_s_1_16, 16);
1211
+ sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1,
1212
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)),
1213
+ sumi_1_s_0_16, 16);
1214
+ sumi_1 = __riscv_vwmacc_vv_i32m2(sumi_1,
1215
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)),
1216
+ sumi_1_s_1_16, 16);
1217
+ sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2,
1218
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)),
1219
+ sumi_2_s_0_16, 16);
1220
+ sumi_2 = __riscv_vwmacc_vv_i32m2(sumi_2,
1221
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)),
1222
+ sumi_2_s_1_16, 16);
1223
+ sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3,
1224
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 2)),
1225
+ sumi_3_s_0_16, 16);
1226
+ sumi_3 = __riscv_vwmacc_vv_i32m2(sumi_3,
1227
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vget_v_u16m4_u16m1(scales, 3)),
1228
+ sumi_3_s_1_16, 16);
1229
+ }
1230
+ }
1231
+
1232
+ const vfloat32m2_t b_d = __riscv_vfwcvt_f_f_v_f32m2(__riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16), 16);
1233
+ const vfloat32m2_t d_0 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[0], 16);
1234
+ const vfloat32m2_t d_1 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[1], 16);
1235
+ const vfloat32m2_t d_2 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[2], 16);
1236
+ const vfloat32m2_t d_3 = __riscv_vfmul_vf_f32m2(b_d, a_ptr[l].d[3], 16);
1237
+
1238
+ sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16);
1239
+ sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16);
1240
+ sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16);
1241
+ sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16);
1242
+ }
1243
+
1244
+ __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16);
1245
+ __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16);
1246
+ __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16);
1247
+ __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
1248
+ }
1249
+ }
1250
+ }
1251
+
1252
+ void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
1253
+ const int qk = QK8_0;
1254
+ const int nb = n / qk;
1255
+ const int ncols_interleaved = 16;
1256
+ const int blocklen = 1;
1257
+
1258
+ assert (n % qk == 0);
1259
+ assert (nr % 4 == 0);
1260
+ assert (nc % ncols_interleaved == 0);
1261
+
1262
+ UNUSED(s);
1263
+ UNUSED(bs);
1264
+ UNUSED(vx);
1265
+ UNUSED(vy);
1266
+ UNUSED(nr);
1267
+ UNUSED(nc);
1268
+ UNUSED(nb);
1269
+ UNUSED(ncols_interleaved);
1270
+ UNUSED(blocklen);
1271
+
1272
+ const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
1273
+ for (int y = 0; y < nr / 4; y++) {
1274
+ const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
1275
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
1276
+ const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb);
1277
+
1278
+ // 4x16 Accumulators
1279
+ vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1280
+ vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1281
+ vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1282
+ vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1283
+
1284
+ for (int l = 0; l < nb; l++) {
1285
+ // 4x16 integer accumulators
1286
+ vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1287
+ vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1288
+ vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1289
+ vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1290
+
1291
+ // Accumulation loop.
1292
+ for (int i = 0; i < QK4_NL / 2; i++) {
1293
+ // Load `b_ptr`.
1294
+ const vuint8mf2_t b_0_packed = __riscv_vle8_v_u8mf2((const uint8_t *)&b_ptr[l].qs[i * 16], 16);
1295
+ const vint8mf2_t b_0_lo = __riscv_vrgather_vv_i8mf2(values, __riscv_vand_vx_u8mf2(b_0_packed, 0xf, 16), 16);
1296
+ const vint8mf2_t b_0_hi = __riscv_vrgather_vv_i8mf2(values, __riscv_vsrl_vx_u8mf2(b_0_packed, 4, 16), 16);
1297
+ // const vint16m1_t b_0_lo_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_lo, 16);
1298
+ // const vint16m1_t b_0_hi_16 = __riscv_vwcvt_x_x_v_i16m1(b_0_hi, 16);
1299
+
1300
+ const vint16m1_t sumi_0_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4], 16);
1301
+ const vint16m1_t sumi_1_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 1], 16);
1302
+ const vint16m1_t sumi_2_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 2], 16);
1303
+ const vint16m1_t sumi_3_lo = __riscv_vwmul_vx_i16m1(b_0_lo, a_ptr[l].qs[i * 4 + 3], 16);
1304
+
1305
+ const vint16m1_t sumi_0_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4], 16);
1306
+ const vint16m1_t sumi_1_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 1], 16);
1307
+ const vint16m1_t sumi_2_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 2], 16);
1308
+ const vint16m1_t sumi_3_hi = __riscv_vwmul_vx_i16m1(b_0_hi, a_ptr[l].qs[64 + i * 4 + 3], 16);
1309
+
1310
+ sumi_0 = __riscv_vadd_vv_i32m2(sumi_0, __riscv_vwadd_vv_i32m2(sumi_0_lo, sumi_0_hi, 16), 16);
1311
+ sumi_1 = __riscv_vadd_vv_i32m2(sumi_1, __riscv_vwadd_vv_i32m2(sumi_1_lo, sumi_1_hi, 16), 16);
1312
+ sumi_2 = __riscv_vadd_vv_i32m2(sumi_2, __riscv_vwadd_vv_i32m2(sumi_2_lo, sumi_2_hi, 16), 16);
1313
+ sumi_3 = __riscv_vadd_vv_i32m2(sumi_3, __riscv_vwadd_vv_i32m2(sumi_3_lo, sumi_3_hi, 16), 16);
1314
+ }
1315
+
1316
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
1317
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16);
1318
+ const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16);
1319
+ const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16);
1320
+ const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16);
1321
+
1322
+ sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16);
1323
+ sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16);
1324
+ sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16);
1325
+ sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16);
1326
+ }
1327
+
1328
+ __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16);
1329
+ __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16);
1330
+ __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16);
1331
+ __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
1332
+ }
1333
+ }
1334
+ }
1335
+
1336
+ void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
1337
+ const int qk = QK8_0;
1338
+ const int nb = n / qk;
1339
+ const int ncols_interleaved = 16;
1340
+ const int blocklen = 1;
1341
+
1342
+ assert (n % qk == 0);
1343
+ assert (nr % 4 == 0);
1344
+ assert (nc % ncols_interleaved == 0);
1345
+
1346
+ UNUSED(s);
1347
+ UNUSED(bs);
1348
+ UNUSED(vx);
1349
+ UNUSED(vy);
1350
+ UNUSED(nr);
1351
+ UNUSED(nc);
1352
+ UNUSED(nb);
1353
+ UNUSED(ncols_interleaved);
1354
+ UNUSED(blocklen);
1355
+
1356
+ for (int y = 0; y < nr / 4; y++) {
1357
+ const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
1358
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
1359
+ const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
1360
+
1361
+ // 4x16 Accumulators
1362
+ vfloat32m2_t sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1363
+ vfloat32m2_t sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1364
+ vfloat32m2_t sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1365
+ vfloat32m2_t sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, 16);
1366
+
1367
+ for (int l = 0; l < nb; l++) {
1368
+ // 4x16 Integer Accumulators
1369
+ vint32m2_t sumi_0 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1370
+ vint32m2_t sumi_1 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1371
+ vint32m2_t sumi_2 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1372
+ vint32m2_t sumi_3 = __riscv_vmv_v_x_i32m2(0.0f, 16);
1373
+
1374
+ // Accumulation loop.
1375
+ for (int i = 0; i < QK8_0; i++) {
1376
+ // Load `b_ptr`.
1377
+ const vint8mf2_t b_0 = __riscv_vle8_v_i8mf2((const int8_t *)&b_ptr[l].qs[i * 16], 16);
1378
+ // const vint16m1_t b_0_16 = __riscv_vwcvt_x_x_v_i16m1(b_0, 16);
1379
+
1380
+ sumi_0 = __riscv_vwadd_wv_i32m2(sumi_0, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 0], 16), 16);
1381
+ sumi_1 = __riscv_vwadd_wv_i32m2(sumi_1, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 1], 16), 16);
1382
+ sumi_2 = __riscv_vwadd_wv_i32m2(sumi_2, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 2], 16), 16);
1383
+ sumi_3 = __riscv_vwadd_wv_i32m2(sumi_3, __riscv_vwmul_vx_i16m1(b_0, a_ptr[l].qs[i * 4 + 3], 16), 16);
1384
+ }
1385
+
1386
+ const vfloat16m1_t b_d = __riscv_vle16_v_f16m1((const _Float16 *)b_ptr[l].d, 16);
1387
+ const vfloat32m2_t d_0 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[0], 16);
1388
+ const vfloat32m2_t d_1 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[1], 16);
1389
+ const vfloat32m2_t d_2 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[2], 16);
1390
+ const vfloat32m2_t d_3 = __riscv_vfwmul_vf_f32m2(b_d, *(const _Float16 *)&a_ptr[l].d[3], 16);
1391
+
1392
+ sumf_0 = __riscv_vfmacc_vv_f32m2(sumf_0, __riscv_vfcvt_f_x_v_f32m2(sumi_0, 16), d_0, 16);
1393
+ sumf_1 = __riscv_vfmacc_vv_f32m2(sumf_1, __riscv_vfcvt_f_x_v_f32m2(sumi_1, 16), d_1, 16);
1394
+ sumf_2 = __riscv_vfmacc_vv_f32m2(sumf_2, __riscv_vfcvt_f_x_v_f32m2(sumi_2, 16), d_2, 16);
1395
+ sumf_3 = __riscv_vfmacc_vv_f32m2(sumf_3, __riscv_vfcvt_f_x_v_f32m2(sumi_3, 16), d_3, 16);
1396
+ }
1397
+
1398
+ __riscv_vse32_v_f32m2(s + (y * 4 + 0) * bs + x * 16, sumf_0, 16);
1399
+ __riscv_vse32_v_f32m2(s + (y * 4 + 1) * bs + x * 16, sumf_1, 16);
1400
+ __riscv_vse32_v_f32m2(s + (y * 4 + 2) * bs + x * 16, sumf_2, 16);
1401
+ __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
1402
+ }
1403
+ }
1404
+ }
1405
+
1406
+ void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
1407
+ assert(n % QK_K == 0);
1408
+ const int num_k_blocks = n / QK_K;
1409
+ const int N_ROWS_TILE = 4;
1410
+ const int N_COLS_TILE = 16;
1411
+ assert(nr % N_ROWS_TILE == 0);
1412
+ assert(nc % N_COLS_TILE == 0);
1413
+
1414
+ const size_t vl = __riscv_vsetvl_e32m2(N_COLS_TILE);
1415
+ // --- Tiling Loops ---
1416
+ #pragma GCC unroll 1
1417
+ for (int row_tile = 0; row_tile < nr; row_tile += N_ROWS_TILE) {
1418
+ #pragma GCC unroll 1
1419
+ for (int col_tile = 0; col_tile < nc; col_tile += N_COLS_TILE) {
1420
+ // Base Pointers
1421
+ const block_q8_Kx4* lhs_base_ptr = (const block_q8_Kx4*)vy + (row_tile / N_ROWS_TILE) * num_k_blocks;
1422
+ const block_q2_Kx16* rhs_base_ptr = (const block_q2_Kx16*)vx + (col_tile / N_COLS_TILE) * num_k_blocks;
1423
+
1424
+ // Persistent Float Accumulators
1425
+ vfloat32m2_t v_sumf_0 = __riscv_vfmv_v_f_f32m2(0.0f, vl);
1426
+ vfloat32m2_t v_sumf_1 = __riscv_vfmv_v_f_f32m2(0.0f, vl);
1427
+ vfloat32m2_t v_sumf_2 = __riscv_vfmv_v_f_f32m2(0.0f, vl);
1428
+ vfloat32m2_t v_sumf_3 = __riscv_vfmv_v_f_f32m2(0.0f, vl);
1429
+
1430
+ // --- Super-Block Loop (K=0..255) ---
1431
+ #pragma GCC unroll 1
1432
+ for (int k_block = 0; k_block < num_k_blocks; ++k_block) {
1433
+ const block_q8_Kx4* lhs_current = &lhs_base_ptr[k_block];
1434
+ const block_q2_Kx16* rhs_current = &rhs_base_ptr[k_block];
1435
+
1436
+ // 1. Load Global Min Scales (Keep as F16/LMUL=1 to save registers)
1437
+ vfloat16m1_t v_g_min_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->dmin, vl);
1438
+ vfloat32m2_t v_g_min_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_min_f16, vl);
1439
+
1440
+ // 2. Initialize Integer Accumulators
1441
+ vint32m2_t v_isum_0 = __riscv_vmv_v_x_i32m2(0, vl);
1442
+ vint32m2_t v_isum_1 = __riscv_vmv_v_x_i32m2(0, vl);
1443
+ vint32m2_t v_isum_2 = __riscv_vmv_v_x_i32m2(0, vl);
1444
+ vint32m2_t v_isum_3 = __riscv_vmv_v_x_i32m2(0, vl);
1445
+
1446
+ const uint8_t* rhs_qs_ptr = rhs_current->qs;
1447
+ const uint8_t* rhs_sc_ptr = rhs_current->scales;
1448
+ const int8_t* lhs_qs_ptr = lhs_current->qs;
1449
+
1450
+ // --- Phase Loop (4 phases x 64 elements) ---
1451
+ #pragma GCC unroll 1
1452
+ for (int phase = 0; phase < 4; ++phase) {
1453
+
1454
+ // A. Load Scales/Mins for the 4 interleaved sub-blocks
1455
+ vuint16m1_t v_d_sb_0, v_d_sb_1, v_d_sb_2, v_d_sb_3;
1456
+ vuint16m1_t v_m_sb_0, v_m_sb_1, v_m_sb_2, v_m_sb_3;
1457
+
1458
+ // Unrolled Load Logic
1459
+ {
1460
+ vuint8mf2_t v_raw;
1461
+ // Sub-block 0
1462
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 0, vl);
1463
+ v_d_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
1464
+ v_m_sb_0 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
1465
+
1466
+ // Sub-block 1
1467
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 16, vl);
1468
+ v_d_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
1469
+ v_m_sb_1 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
1470
+
1471
+ // Sub-block 2
1472
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 32, vl);
1473
+ v_d_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
1474
+ v_m_sb_2 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
1475
+
1476
+ // Sub-block 3
1477
+ v_raw = __riscv_vle8_v_u8mf2(rhs_sc_ptr + 48, vl);
1478
+ v_d_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vand_vx_u8mf2(v_raw, 0xF, vl), vl);
1479
+ v_m_sb_3 = __riscv_vzext_vf2_u16m1(__riscv_vsrl_vx_u8mf2(v_raw, 4, vl), vl);
1480
+
1481
+ rhs_sc_ptr += 64;
1482
+ }
1483
+
1484
+ int base_k_phase = (phase < 2) ? (phase * 16) : (128 + (phase-2)*16);
1485
+ int k_offsets[4] = {0, 32, 64, 96};
1486
+
1487
+ // B. Inner Dot Product Loop
1488
+ #pragma GCC unroll 1
1489
+ for (int l = 0; l < 16; ++l) {
1490
+ vuint8mf2_t v_rhs_data = __riscv_vle8_v_u8mf2(rhs_qs_ptr, vl);
1491
+ rhs_qs_ptr += 16;
1492
+
1493
+ // Unroll over 4 sub-blocks (0, 1, 2, 3 relative to phase)
1494
+
1495
+ // --- Sub-block 0 ---
1496
+ {
1497
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(v_rhs_data, 3, vl);
1498
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
1499
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
1500
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_0), vl);
1501
+
1502
+ const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[0] + l) * 4];
1503
+ v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl);
1504
+ v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl);
1505
+ v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl);
1506
+ v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl);
1507
+ }
1508
+ // --- Sub-block 1 ---
1509
+ {
1510
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 2, vl), 3, vl);
1511
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
1512
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
1513
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_1), vl);
1514
+
1515
+ const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[1] + l) * 4];
1516
+ v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl);
1517
+ v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl);
1518
+ v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl);
1519
+ v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl);
1520
+ }
1521
+ // --- Sub-block 2 ---
1522
+ {
1523
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 4, vl), 3, vl);
1524
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
1525
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
1526
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_2), vl);
1527
+
1528
+ const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[2] + l) * 4];
1529
+ v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl);
1530
+ v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl);
1531
+ v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl);
1532
+ v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl);
1533
+ }
1534
+ // --- Sub-block 3 ---
1535
+ {
1536
+ vuint8mf2_t v_q2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(v_rhs_data, 6, vl), 3, vl);
1537
+ vint16m1_t v_w = __riscv_vmul_vv_i16m1(
1538
+ __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(v_q2, vl)),
1539
+ __riscv_vreinterpret_v_u16m1_i16m1(v_d_sb_3), vl);
1540
+
1541
+ const int8_t* q8 = &lhs_qs_ptr[(base_k_phase + k_offsets[3] + l) * 4];
1542
+ v_isum_0 = __riscv_vwmacc_vx_i32m2(v_isum_0, (int16_t)q8[0], v_w, vl);
1543
+ v_isum_1 = __riscv_vwmacc_vx_i32m2(v_isum_1, (int16_t)q8[1], v_w, vl);
1544
+ v_isum_2 = __riscv_vwmacc_vx_i32m2(v_isum_2, (int16_t)q8[2], v_w, vl);
1545
+ v_isum_3 = __riscv_vwmacc_vx_i32m2(v_isum_3, (int16_t)q8[3], v_w, vl);
1546
+ }
1547
+ }
1548
+
1549
+ // C CORRECTION
1550
+ int sb_base_abs = base_k_phase / 16;
1551
+
1552
+ // --- Correction Sub-block 0 ---
1553
+ {
1554
+ int sb_abs = sb_base_abs + (k_offsets[0] / 16);
1555
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_0);
1556
+
1557
+ // Row 0
1558
+ vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl);
1559
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl);
1560
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1561
+ v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl);
1562
+
1563
+ // Row 1
1564
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl);
1565
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl);
1566
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1567
+ v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl);
1568
+
1569
+ // Row 2
1570
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl);
1571
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl);
1572
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1573
+ v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl);
1574
+
1575
+ // Row 3
1576
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl);
1577
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl);
1578
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1579
+ v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl);
1580
+ }
1581
+
1582
+ // --- Correction Sub-block 1 ---
1583
+ {
1584
+ int sb_abs = sb_base_abs + (k_offsets[1] / 16);
1585
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_1);
1586
+
1587
+ vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl);
1588
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl);
1589
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1590
+ v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl);
1591
+
1592
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl);
1593
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl);
1594
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1595
+ v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl);
1596
+
1597
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl);
1598
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl);
1599
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1600
+ v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl);
1601
+
1602
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl);
1603
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl);
1604
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1605
+ v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl);
1606
+ }
1607
+
1608
+ // --- Correction Sub-block 2 ---
1609
+ {
1610
+ int sb_abs = sb_base_abs + (k_offsets[2] / 16);
1611
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_2);
1612
+
1613
+ vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl);
1614
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl);
1615
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1616
+ v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl);
1617
+
1618
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl);
1619
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl);
1620
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1621
+ v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl);
1622
+
1623
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl);
1624
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl);
1625
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1626
+ v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl);
1627
+
1628
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl);
1629
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl);
1630
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1631
+ v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl);
1632
+ }
1633
+
1634
+ // --- Correction Sub-block 3 ---
1635
+ {
1636
+ int sb_abs = sb_base_abs + (k_offsets[3] / 16);
1637
+ vint16m1_t v_min = __riscv_vreinterpret_v_u16m1_i16m1(v_m_sb_3);
1638
+
1639
+ vfloat32m2_t v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[0], vl);
1640
+ vint32m2_t v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 0], vl);
1641
+ vfloat32m2_t vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1642
+ v_sumf_0 = __riscv_vfsub_vv_f32m2(v_sumf_0, vf_c, vl);
1643
+
1644
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[1], vl);
1645
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 1], vl);
1646
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1647
+ v_sumf_1 = __riscv_vfsub_vv_f32m2(v_sumf_1, vf_c, vl);
1648
+
1649
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[2], vl);
1650
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 2], vl);
1651
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1652
+ v_sumf_2 = __riscv_vfsub_vv_f32m2(v_sumf_2, vf_c, vl);
1653
+
1654
+ v_g_min = __riscv_vfmul_vf_f32m2(v_g_min_base, lhs_current->d[3], vl);
1655
+ v_c = __riscv_vwmul_vx_i32m2(v_min, lhs_current->bsums[sb_abs * 4 + 3], vl);
1656
+ vf_c = __riscv_vfmul_vv_f32m2(__riscv_vfcvt_f_x_v_f32m2(v_c, vl), v_g_min, vl);
1657
+ v_sumf_3 = __riscv_vfsub_vv_f32m2(v_sumf_3, vf_c, vl);
1658
+ }
1659
+
1660
+ } // End Phase Loop
1661
+
1662
+ // --- Apply Main Scales ---
1663
+ vfloat16m1_t v_g_all_f16 = __riscv_vle16_v_f16m1((const _Float16*)rhs_current->d, vl);
1664
+ vfloat32m2_t v_g_all_base = __riscv_vfwcvt_f_f_v_f32m2(v_g_all_f16, vl);
1665
+
1666
+ {
1667
+ vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[0], vl);
1668
+ vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_0, vl);
1669
+ v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl);
1670
+ v_sumf_0 = __riscv_vfadd_vv_f32m2(v_sumf_0, v_sum, vl);
1671
+ }
1672
+ // Row 1
1673
+ {
1674
+ vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[1], vl);
1675
+ vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_1, vl);
1676
+ v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl);
1677
+ v_sumf_1 = __riscv_vfadd_vv_f32m2(v_sumf_1, v_sum, vl);
1678
+ }
1679
+ // Row 2
1680
+ {
1681
+ vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[2], vl);
1682
+ vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_2, vl);
1683
+ v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl);
1684
+ v_sumf_2 = __riscv_vfadd_vv_f32m2(v_sumf_2, v_sum, vl);
1685
+ }
1686
+ // Row 3
1687
+ {
1688
+ vfloat32m2_t v_g_all = __riscv_vfmul_vf_f32m2(v_g_all_base, lhs_current->d[3], vl);
1689
+ vfloat32m2_t v_sum = __riscv_vfcvt_f_x_v_f32m2(v_isum_3, vl);
1690
+ v_sum = __riscv_vfmul_vv_f32m2(v_sum, v_g_all, vl);
1691
+ v_sumf_3 = __riscv_vfadd_vv_f32m2(v_sumf_3, v_sum, vl);
1692
+ }
1693
+
1694
+ } // End K-Block
1695
+
1696
+ __riscv_vse32_v_f32m2(s + (row_tile + 0) * bs + col_tile, v_sumf_0, vl);
1697
+ __riscv_vse32_v_f32m2(s + (row_tile + 1) * bs + col_tile, v_sumf_1, vl);
1698
+ __riscv_vse32_v_f32m2(s + (row_tile + 2) * bs + col_tile, v_sumf_2, vl);
1699
+ __riscv_vse32_v_f32m2(s + (row_tile + 3) * bs + col_tile, v_sumf_3, vl);
1700
+ }
1701
+ }
1702
+ }
1703
+ #endif