whispercpp 1.3.5 → 1.3.7

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 (1017) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/LICENSE +1 -1
  5. data/README.md +133 -3
  6. data/Rakefile +18 -3
  7. data/ext/dependencies.rb +10 -4
  8. data/ext/dependencies_for_windows.rb +17 -0
  9. data/ext/extconf.rb +20 -7
  10. data/ext/options.rb +54 -14
  11. data/ext/options_for_windows.rb +51 -0
  12. data/ext/ruby_whisper.c +56 -46
  13. data/ext/ruby_whisper.h +165 -2
  14. data/ext/ruby_whisper_context.c +297 -126
  15. data/ext/ruby_whisper_context_params.c +163 -0
  16. data/ext/ruby_whisper_log_queue.c +180 -0
  17. data/ext/ruby_whisper_log_settable.h +47 -0
  18. data/ext/ruby_whisper_model.c +0 -1
  19. data/ext/ruby_whisper_parakeet.c +49 -0
  20. data/ext/ruby_whisper_parakeet_context.c +304 -0
  21. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  22. data/ext/ruby_whisper_parakeet_model.c +84 -0
  23. data/ext/ruby_whisper_parakeet_params.c +548 -0
  24. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  25. data/ext/ruby_whisper_parakeet_token.c +188 -0
  26. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  27. data/ext/ruby_whisper_params.c +256 -66
  28. data/ext/ruby_whisper_segment.c +6 -7
  29. data/ext/ruby_whisper_token.c +29 -9
  30. data/ext/ruby_whisper_transcribe.cpp +46 -16
  31. data/ext/ruby_whisper_vad_context.c +48 -1
  32. data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
  33. data/ext/ruby_whisper_vad_params.c +0 -1
  34. data/ext/ruby_whisper_vad_segment.c +0 -1
  35. data/ext/ruby_whisper_vad_segments.c +0 -1
  36. data/ext/sources/CMakeLists.txt +41 -3
  37. data/ext/sources/CMakePresets.json +95 -0
  38. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  39. data/ext/sources/cmake/parakeet.pc.in +10 -0
  40. data/ext/sources/cmake/whisper-config.cmake.in +5 -40
  41. data/ext/sources/cmake/whisper.pc.in +1 -1
  42. data/ext/sources/examples/CMakeLists.txt +4 -2
  43. data/ext/sources/examples/bench/bench.cpp +24 -19
  44. data/ext/sources/examples/cli/cli.cpp +51 -9
  45. data/ext/sources/examples/common-ggml.cpp +4 -0
  46. data/ext/sources/examples/common-whisper.cpp +139 -67
  47. data/ext/sources/examples/common-whisper.h +11 -0
  48. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  49. data/ext/sources/examples/miniaudio.h +4507 -2131
  50. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  51. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  52. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  53. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  54. data/ext/sources/examples/server/server.cpp +213 -163
  55. data/ext/sources/ggml/CMakeLists.txt +29 -15
  56. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  57. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  58. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  59. data/ext/sources/ggml/include/ggml-backend.h +73 -11
  60. data/ext/sources/ggml/include/ggml-cann.h +1 -1
  61. data/ext/sources/ggml/include/ggml-cpu.h +5 -0
  62. data/ext/sources/ggml/include/ggml-cuda.h +3 -0
  63. data/ext/sources/ggml/include/ggml-openvino.h +37 -0
  64. data/ext/sources/ggml/include/ggml-opt.h +1 -1
  65. data/ext/sources/ggml/include/ggml-rpc.h +8 -3
  66. data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
  67. data/ext/sources/ggml/include/ggml.h +155 -16
  68. data/ext/sources/ggml/include/gguf.h +10 -2
  69. data/ext/sources/ggml/src/CMakeLists.txt +25 -5
  70. data/ext/sources/ggml/src/ggml-alloc.c +9 -10
  71. data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
  72. data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
  73. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  74. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
  75. data/ext/sources/ggml/src/ggml-backend-reg.cpp +40 -86
  76. data/ext/sources/ggml/src/ggml-backend.cpp +114 -10
  77. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  78. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +10 -2
  79. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
  80. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
  81. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +1016 -442
  82. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +111 -85
  83. data/ext/sources/ggml/src/ggml-cann/common.h +23 -14
  84. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +255 -92
  85. data/ext/sources/ggml/src/ggml-common.h +22 -0
  86. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +68 -34
  87. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +44 -19
  88. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
  89. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +101 -101
  90. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -1
  91. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2874 -613
  92. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  93. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  94. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +5480 -840
  95. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1361 -0
  96. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -11
  97. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  98. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +186 -36
  99. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +119 -19
  100. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +112 -26
  101. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
  102. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  103. data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
  104. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -0
  105. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +153 -16
  106. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +17 -0
  107. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
  108. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +976 -251
  109. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +671 -266
  110. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1277 -263
  111. data/ext/sources/ggml/src/ggml-cpu/ops.h +4 -0
  112. data/ext/sources/ggml/src/ggml-cpu/quants.c +95 -0
  113. data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
  114. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2893 -679
  115. data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
  116. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +226 -0
  117. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +114 -19
  118. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  119. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  120. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  121. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  122. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  123. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  124. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  125. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  126. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  127. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  128. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  129. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  130. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  131. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  132. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  133. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
  134. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +54 -53
  135. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  136. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +18 -8
  137. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  138. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  139. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +73 -28
  140. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +69 -41
  141. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  142. data/ext/sources/ggml/src/ggml-cuda/common.cuh +359 -29
  143. data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
  144. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  145. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  146. data/ext/sources/ggml/src/ggml-cuda/convert.cu +94 -27
  147. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  148. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +20 -9
  149. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  150. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +333 -85
  151. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +632 -190
  152. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  153. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +162 -49
  154. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +43 -18
  155. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +44 -14
  156. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
  157. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +241 -23
  158. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  159. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  160. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  161. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +312 -0
  162. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
  163. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
  164. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1454 -599
  165. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  166. data/ext/sources/ggml/src/ggml-cuda/mean.cu +13 -10
  167. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +397 -183
  168. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
  169. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +161 -88
  170. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
  171. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +522 -431
  172. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +139 -72
  173. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
  174. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +608 -88
  175. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -0
  176. data/ext/sources/ggml/src/ggml-cuda/norm.cu +47 -79
  177. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
  178. data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
  179. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +134 -27
  180. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  181. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +7 -17
  182. data/ext/sources/ggml/src/ggml-cuda/rope.cu +244 -137
  183. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  184. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
  185. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  186. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  187. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  188. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
  189. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
  190. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +96 -40
  191. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  192. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  193. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +2 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +2 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +2 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  205. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +2 -0
  206. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  207. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  208. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  209. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  210. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  211. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  212. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  213. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  214. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  215. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  216. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  217. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  218. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  219. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  220. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  221. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  222. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  223. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  224. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  225. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -5
  226. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +202 -135
  227. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
  228. data/ext/sources/ggml/src/ggml-cuda/unary.cu +86 -2
  229. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +4 -0
  230. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +111 -17
  231. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  232. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +30 -2
  233. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +84 -46
  235. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1612 -753
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +51 -11
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +361 -261
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +294 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +753 -241
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +295 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +471 -296
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +159 -53
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +3 -3
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +372 -0
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +137 -0
  254. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
  255. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
  256. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
  257. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
  258. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
  259. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
  260. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
  261. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
  262. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +97 -14
  263. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +163 -67
  264. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
  265. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
  266. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +308 -0
  267. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
  268. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
  269. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  270. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +216 -0
  271. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  272. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  273. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  274. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  275. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  276. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
  277. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  278. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  279. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +142 -0
  280. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  281. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  282. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  283. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -1348
  284. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +547 -635
  285. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3556 -1101
  286. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  287. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  288. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +475 -269
  289. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +94 -72
  290. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +222 -217
  291. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  292. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +432 -0
  293. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  294. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +886 -117
  295. data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
  296. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
  297. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +40 -0
  302. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +28 -9
  303. data/ext/sources/ggml/src/ggml-impl.h +68 -1
  304. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
  305. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
  306. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
  307. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
  308. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +409 -83
  309. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +54 -5
  310. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +254 -52
  311. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +254 -23
  312. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +756 -285
  313. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +7 -4
  314. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +359 -133
  315. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1867 -1123
  316. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  317. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +71 -4
  318. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +14127 -5314
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +97 -88
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +104 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1978 -67
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +195 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  365. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
  366. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  367. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  368. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  369. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  370. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  371. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  372. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  373. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  374. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  375. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  376. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  377. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  378. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  379. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  380. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  381. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  382. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  383. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  384. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  385. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  386. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  387. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  388. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
  389. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +178 -0
  390. data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  391. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
  392. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
  393. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
  394. data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  395. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
  396. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
  397. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
  398. data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  399. data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
  400. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  401. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +985 -0
  402. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
  403. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +380 -0
  404. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
  405. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1132 -0
  406. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +956 -0
  407. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  412. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  413. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
  414. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
  415. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
  416. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
  417. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
  418. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
  419. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
  420. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
  421. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
  422. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
  423. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
  424. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +149 -0
  425. data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
  426. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
  427. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
  428. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
  429. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
  430. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  431. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
  432. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +47 -0
  433. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +40 -0
  434. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  435. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  436. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  437. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  438. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  439. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  440. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +317 -0
  441. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
  442. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +257 -0
  443. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +86 -0
  444. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +880 -0
  445. data/ext/sources/ggml/src/ggml-openvino/utils.h +143 -0
  446. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  447. data/ext/sources/ggml/src/ggml-quants.c +385 -119
  448. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  449. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  450. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  451. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  452. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  453. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +64 -91
  454. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
  455. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +4 -1
  456. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
  457. data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
  458. data/ext/sources/ggml/src/ggml-sycl/common.hpp +356 -11
  459. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +184 -14
  460. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +31 -1
  461. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
  462. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  463. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  464. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
  465. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  466. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  467. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
  468. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
  469. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +77 -156
  470. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
  471. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  472. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  473. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1181 -0
  474. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
  475. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
  476. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +674 -0
  477. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +227 -0
  478. data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
  479. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  480. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  481. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +347 -0
  482. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +9 -0
  483. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  484. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  485. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1134 -236
  486. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  487. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  488. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
  489. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
  490. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
  491. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
  492. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  493. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
  494. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +72 -1
  495. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
  496. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
  497. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
  498. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
  499. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  500. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  501. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  502. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  503. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  504. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  505. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  506. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  507. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  508. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  509. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  510. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  511. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  512. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  513. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  514. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  515. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  516. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
  517. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
  518. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
  519. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
  520. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
  521. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
  522. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
  523. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
  524. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
  525. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
  526. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
  527. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
  528. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
  529. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
  530. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
  531. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
  532. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
  533. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
  534. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
  535. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
  536. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
  537. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
  538. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
  539. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
  540. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
  541. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
  542. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
  543. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
  544. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
  545. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
  546. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
  547. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
  548. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
  549. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
  550. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
  551. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
  552. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  553. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  554. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  555. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +228 -53
  556. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
  557. data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  558. data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  559. data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  560. data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  561. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  562. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  563. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  564. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  565. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  566. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  567. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  568. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  569. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  570. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  571. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  572. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  573. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  574. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  575. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  576. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  577. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  578. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
  579. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
  580. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  581. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +71 -0
  582. data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  583. data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  584. data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  585. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  586. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  587. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  588. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  589. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  590. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  591. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  592. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
  593. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  594. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  595. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  596. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
  597. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
  598. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +12 -1
  599. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3250 -940
  600. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  601. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
  602. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  603. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  604. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  605. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  606. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  607. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  608. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  609. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  610. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  611. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
  612. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  613. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
  614. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
  615. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  616. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +533 -180
  617. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +113 -68
  618. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +412 -222
  619. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +222 -83
  620. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  621. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  622. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  623. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
  624. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  625. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
  626. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  627. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
  628. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
  629. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  630. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  631. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
  632. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  633. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  634. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
  635. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  636. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  637. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  638. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  639. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
  640. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  641. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +22 -20
  642. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +51 -14
  643. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
  644. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  645. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
  646. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  647. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  648. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  649. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
  650. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +39 -63
  651. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  652. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
  653. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
  654. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
  655. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +13 -7
  656. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
  657. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
  658. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  659. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +27 -11
  660. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
  661. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  662. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -149
  663. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  664. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3221 -97
  665. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3493 -1997
  666. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  667. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  668. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  669. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  670. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  671. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
  672. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +115 -141
  673. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +93 -0
  674. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  675. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -44
  676. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  677. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +198 -230
  678. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  679. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  680. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  681. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  682. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  683. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  684. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +234 -335
  685. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  686. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  687. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +871 -42
  688. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  689. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  690. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  691. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +149 -0
  692. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +36 -138
  693. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +151 -0
  694. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
  695. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
  696. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  697. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  698. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
  699. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
  700. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  701. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  702. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  703. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +15 -40
  704. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  705. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +39 -12
  706. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  707. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  708. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  709. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  710. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  711. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  712. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
  713. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  714. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +24 -15
  715. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
  716. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +253 -16
  717. data/ext/sources/ggml/src/ggml.c +268 -52
  718. data/ext/sources/ggml/src/gguf.cpp +377 -47
  719. data/ext/sources/include/parakeet.h +342 -0
  720. data/ext/sources/include/whisper.h +10 -0
  721. data/ext/sources/media/matmul.png +0 -0
  722. data/ext/sources/src/CMakeLists.txt +23 -0
  723. data/ext/sources/src/parakeet-arch.h +188 -0
  724. data/ext/sources/src/parakeet.cpp +3838 -0
  725. data/ext/sources/src/whisper.cpp +62 -40
  726. data/extsources.rb +26 -10
  727. data/lib/whisper/log_settable.rb +36 -0
  728. data/lib/whisper/model/uri.rb +13 -1
  729. data/lib/whisper/output.rb +74 -0
  730. data/sig/whisper.rbs +445 -55
  731. data/test/helper.rb +2 -0
  732. data/test/jfk_reader/jfk_reader.c +50 -7
  733. data/test/test_callback.rb +1 -0
  734. data/test/test_context_params.rb +82 -0
  735. data/test/test_package.rb +6 -5
  736. data/test/test_parakeet.rb +28 -0
  737. data/test/test_parakeet_callback.rb +107 -0
  738. data/test/test_parakeet_context.rb +116 -0
  739. data/test/test_parakeet_context_params.rb +24 -0
  740. data/test/test_parakeet_model.rb +21 -0
  741. data/test/test_parakeet_params.rb +78 -0
  742. data/test/test_parakeet_segment.rb +42 -0
  743. data/test/test_parakeet_token.rb +73 -0
  744. data/test/test_params.rb +2 -0
  745. data/test/test_token.rb +11 -0
  746. data/test/test_vad_context.rb +58 -8
  747. data/test/test_vad_segment.rb +1 -1
  748. data/test/test_whisper.rb +44 -6
  749. data/whispercpp.gemspec +2 -2
  750. metadata +426 -280
  751. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  752. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  753. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  754. data/ext/sources/bindings/javascript/package.json +0 -26
  755. data/ext/sources/bindings/javascript/whisper.js +0 -19
  756. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  757. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  758. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  759. data/ext/sources/examples/addon.node/index.js +0 -59
  760. data/ext/sources/examples/addon.node/package.json +0 -16
  761. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  762. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  763. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  764. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  765. data/ext/sources/examples/coi-serviceworker.js +0 -146
  766. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  767. data/ext/sources/examples/command/command.cpp +0 -802
  768. data/ext/sources/examples/command/commands.txt +0 -9
  769. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  770. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  771. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  772. data/ext/sources/examples/generate-karaoke.sh +0 -57
  773. data/ext/sources/examples/helpers.js +0 -191
  774. data/ext/sources/examples/livestream.sh +0 -112
  775. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  776. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  777. data/ext/sources/examples/lsp/whisper.vim +0 -362
  778. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  779. data/ext/sources/examples/python/whisper_processor.py +0 -54
  780. data/ext/sources/examples/server/bench.js +0 -29
  781. data/ext/sources/examples/server.py +0 -120
  782. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  783. data/ext/sources/examples/stream/stream.cpp +0 -437
  784. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  785. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  786. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  787. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  788. data/ext/sources/examples/sycl/build.sh +0 -22
  789. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  790. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  791. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -47
  792. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  793. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -494
  794. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -88
  795. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2559
  796. data/ext/sources/examples/talk-llama/llama-arch.h +0 -586
  797. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -917
  798. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  799. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -876
  800. data/ext/sources/examples/talk-llama/llama-chat.h +0 -70
  801. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3645
  802. data/ext/sources/examples/talk-llama/llama-context.h +0 -360
  803. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  804. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -42
  805. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  806. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  807. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2282
  808. data/ext/sources/examples/talk-llama/llama-graph.h +0 -910
  809. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -241
  810. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -284
  811. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  812. data/ext/sources/examples/talk-llama/llama-impl.h +0 -63
  813. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  814. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  815. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -328
  816. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  817. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2100
  818. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -390
  819. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  820. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  821. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  822. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1167
  823. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  824. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  825. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  826. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -735
  827. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  828. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1247
  829. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -176
  830. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -285
  831. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -37
  832. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -8338
  833. data/ext/sources/examples/talk-llama/llama-model.h +0 -544
  834. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1072
  835. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  836. data/ext/sources/examples/talk-llama/llama-sampling.cpp +0 -3771
  837. data/ext/sources/examples/talk-llama/llama-sampling.h +0 -44
  838. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3900
  839. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -182
  840. data/ext/sources/examples/talk-llama/llama.cpp +0 -1140
  841. data/ext/sources/examples/talk-llama/llama.h +0 -1540
  842. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -191
  843. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  844. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  845. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -138
  846. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  847. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -122
  848. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -144
  849. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -135
  850. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -178
  851. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -160
  852. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  853. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  854. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  855. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  856. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  857. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  858. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  859. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -123
  860. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  861. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -144
  862. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -259
  863. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -134
  864. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  865. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -150
  866. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  867. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  868. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  869. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -113
  870. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  871. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  872. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  873. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  874. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  875. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  876. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  877. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -150
  878. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  879. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  880. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -196
  881. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -211
  882. data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +0 -283
  883. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  884. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -141
  885. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  886. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -154
  887. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  888. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  889. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  890. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -175
  891. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  892. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  893. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  894. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -168
  895. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  896. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -55
  897. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -123
  898. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -199
  899. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -124
  900. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  901. data/ext/sources/examples/talk-llama/models/models.h +0 -569
  902. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -116
  903. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  904. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -150
  905. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  906. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  907. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  908. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  909. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  910. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  911. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  912. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  913. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  914. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  915. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  916. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  917. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -316
  918. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  919. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -168
  920. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  921. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  922. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  923. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  924. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -117
  925. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -124
  926. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -873
  927. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -149
  928. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -141
  929. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  930. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  931. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -162
  932. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  933. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  934. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -135
  935. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  936. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  937. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  938. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  939. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  940. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  941. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  942. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  943. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  944. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  945. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  946. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  947. data/ext/sources/examples/talk-llama/speak +0 -40
  948. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  949. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  950. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  951. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  952. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  953. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1147
  954. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  955. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  956. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  957. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  958. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  959. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  960. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  961. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  962. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  963. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  964. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  965. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  966. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  967. data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
  968. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
  969. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  970. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +0 -157
  971. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -165
  972. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
  973. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
  974. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
  975. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
  976. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
  977. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
  978. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
  979. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  980. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  981. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  982. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
  983. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
  984. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -147
  985. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  986. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +0 -907
  987. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +0 -247
  988. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
  989. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  990. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
  991. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
  992. data/ext/sources/tests/CMakeLists.txt +0 -112
  993. data/ext/sources/tests/earnings21/eval.mk +0 -58
  994. data/ext/sources/tests/earnings21/eval.py +0 -68
  995. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  996. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  997. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  998. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  999. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  1000. data/ext/sources/tests/en-0-ref.txt +0 -1
  1001. data/ext/sources/tests/en-1-ref.txt +0 -1
  1002. data/ext/sources/tests/en-2-ref.txt +0 -1
  1003. data/ext/sources/tests/es-0-ref.txt +0 -1
  1004. data/ext/sources/tests/librispeech/eval.mk +0 -39
  1005. data/ext/sources/tests/librispeech/eval.py +0 -47
  1006. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  1007. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  1008. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  1009. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  1010. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  1011. data/ext/sources/tests/run-tests.sh +0 -130
  1012. data/ext/sources/tests/test-c.c +0 -3
  1013. data/ext/sources/tests/test-vad-full.cpp +0 -56
  1014. data/ext/sources/tests/test-vad.cpp +0 -83
  1015. data/ext/sources/tests/test-whisper.js +0 -58
  1016. data/lib/whisper/context.rb +0 -15
  1017. data/lib/whisper/segment.rb +0 -58
@@ -45,6 +45,7 @@ std::string target_cpp = "";
45
45
  const std::vector<std::string> type_names = {
46
46
  "f32",
47
47
  "f16",
48
+ "q1_0",
48
49
  "q4_0",
49
50
  "q4_1",
50
51
  "q5_0",
@@ -65,6 +66,7 @@ const std::vector<std::string> type_names = {
65
66
  "iq4_xs",
66
67
  "iq4_nl",
67
68
  "mxfp4",
69
+ "nvfp4",
68
70
  "bf16",
69
71
  };
70
72
 
@@ -137,6 +139,7 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
137
139
 
138
140
  pid_t pid = fork();
139
141
  if (pid < 0) {
142
+ std::cerr << strerror(errno) << "\n";
140
143
  throw std::runtime_error("Failed to fork process");
141
144
  }
142
145
 
@@ -330,10 +333,11 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
330
333
  std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, in_path, "-o", out_path};
331
334
  #endif
332
335
 
333
- // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
336
+ // disable spirv-opt for coopmat shaders for https://github.com/ggml-org/llama.cpp/issues/10734
334
337
  // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
335
338
  // disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
336
- if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos) {
339
+ // disable spirv-opt for dot2 shaders (spirv-opt doesn't recognize SPV_VALVE_mixed_float_dot_product capability)
340
+ if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos && name.find("_dot2") == std::string::npos) {
337
341
  cmd.push_back("-O");
338
342
  }
339
343
 
@@ -404,8 +408,8 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
404
408
  }
405
409
 
406
410
  static std::vector<std::future<void>> compiles;
407
- void string_to_spv(std::string name, const std::string& source, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
408
- name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
411
+ void string_to_spv(std::string name, const std::string& source, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false, const std::string& suffix = "") {
412
+ name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")) + suffix;
409
413
  std::string out_path = join_paths(output_dir, name + ".spv");
410
414
 
411
415
  if (input_filepath == "") {
@@ -424,10 +428,11 @@ void string_to_spv(std::string name, const std::string& source, const std::map<s
424
428
  generate_dep_file = false;
425
429
  }
426
430
 
427
- void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc) {
431
+ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc, bool dot2 = false) {
428
432
  std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
429
433
  std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
430
434
  std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
435
+ std::string dot2_sfx = dot2 ? "_dot2" : "";
431
436
 
432
437
  std::map<std::string, std::string> base_dict;
433
438
  std::string shader_name = "matmul";
@@ -445,8 +450,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
445
450
  base_dict["FLOAT16"] = "1";
446
451
  }
447
452
 
448
- base_dict["ACC_TYPE" ] = f16acc ? "float16_t" : "float";
449
- base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2" : "vec2";
453
+ base_dict["ACC_TYPE" ] = f16acc ? "float16_t" : "float";
454
+ base_dict["ACC_TYPEV2"] = f16acc ? "f16vec2" : "vec2";
450
455
  if (f16acc) {
451
456
  base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
452
457
  }
@@ -454,6 +459,15 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
454
459
  if (coopmat) {
455
460
  base_dict["COOPMAT"] = "1";
456
461
  }
462
+ #if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
463
+ if (coopmat2) {
464
+ base_dict["GGML_VULKAN_COOPMAT2_DECODE_VECTOR"] = "1";
465
+ }
466
+ #endif
467
+
468
+ if (dot2) {
469
+ base_dict["DOT2_F16"] = "1";
470
+ }
457
471
 
458
472
  const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
459
473
 
@@ -513,18 +527,18 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
513
527
  };
514
528
 
515
529
  const std::map<std::string, std::string> float_type_dict_f16 = {
516
- {"FLOAT_TYPE", FLOAT_TYPE(1, "f16")},
517
- {"FLOAT_TYPE_VEC2", FLOAT_TYPE(2, "f16")},
518
- {"FLOAT_TYPE_VEC4", FLOAT_TYPE(4, "f16")},
519
- {"FLOAT_TYPE_VEC8", FLOAT_TYPE(8, "f16")},
530
+ {"FLOAT_TYPE", FLOAT_TYPE(1, "f16")},
531
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, "f16")},
532
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, "f16")},
533
+ {"FLOAT_TYPEV8", FLOAT_TYPE(8, "f16")},
520
534
  };
521
535
 
522
536
  // Shaders with f16 B_TYPE
523
- string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
524
- string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
537
+ string_to_spv(shader_name + "_f32_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
538
+ string_to_spv(shader_name + "_f32_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
525
539
 
526
- string_to_spv(shader_name + "_f16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
527
- string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
540
+ string_to_spv(shader_name + "_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
541
+ string_to_spv(shader_name + "_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
528
542
 
529
543
  // bf16
530
544
  {
@@ -535,9 +549,9 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
535
549
  std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
536
550
 
537
551
  const std::map<std::string, std::string> float_type_dict_bf16 = {
538
- {"FLOAT_TYPE", FLOAT_TYPE(1, "bf16")},
539
- {"FLOAT_TYPE_VEC2", FLOAT_TYPE(2, "bf16")},
540
- {"FLOAT_TYPE_VEC4", FLOAT_TYPE(4, "bf16")},
552
+ {"FLOAT_TYPE", FLOAT_TYPE(1, "bf16")},
553
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, "bf16")},
554
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, "bf16")},
541
555
  };
542
556
 
543
557
  // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
@@ -545,16 +559,18 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
545
559
  if (!(coopmat || coopmat2))
546
560
  #endif
547
561
  {
548
- string_to_spv(shader_name + "_bf16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}), fp16, coopmat, coopmat2, f16acc);
549
- string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
562
+ if (!dot2) {
563
+ string_to_spv(shader_name + "_bf16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}), fp16, coopmat, coopmat2, f16acc);
564
+ string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
565
+ }
550
566
  }
551
567
  }
552
568
 
553
569
  for (const auto& tname : type_names) {
554
570
  std::string load_vec_quant = "2";
555
- if ((tname == "q4_0") || (tname == "q4_1") || (tname == "q5_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
571
+ if ((tname == "q1_0") || (tname == "q4_0") || (tname == "q4_1") || (tname == "q5_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
556
572
  load_vec_quant = "8";
557
- else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
573
+ else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_xs") || (tname == "iq4_nl") || (tname == "mxfp4") || (tname == "nvfp4"))
558
574
  load_vec_quant = "4";
559
575
 
560
576
  if (tname == "bf16") {
@@ -568,26 +584,26 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
568
584
  std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
569
585
 
570
586
  const std::map<std::string, std::string> float_type_dict = {
571
- {"FLOAT_TYPE", FLOAT_TYPE(1, tname)},
572
- {"FLOAT_TYPE_VEC2", FLOAT_TYPE(2, tname)},
573
- {"FLOAT_TYPE_VEC4", FLOAT_TYPE(4, tname)},
574
- {"FLOAT_TYPE_VEC8", FLOAT_TYPE(8, tname)},
587
+ {"FLOAT_TYPE", FLOAT_TYPE(1, tname)},
588
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, tname)},
589
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, tname)},
590
+ {"FLOAT_TYPEV8", FLOAT_TYPE(8, tname)},
575
591
  };
576
592
 
577
593
  // don't generate f32 variants for coopmat2
578
594
  if (!coopmat2) {
579
- string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
580
- string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
595
+ string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
596
+ string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
581
597
  }
582
598
 
583
599
  if (tname != "f16" && tname != "f32") {
584
- string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
585
- string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
600
+ string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
601
+ string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
586
602
  }
587
603
 
588
604
  #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
589
- // Integer dot mmq performs better with f32 accumulators
590
- if (!f16acc && !coopmat && !coopmat2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
605
+ // Integer dot mmq performs better with f32 accumulators (different shader, skip for dot2)
606
+ if (!f16acc && !coopmat && !coopmat2 && !dot2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
591
607
  string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
592
608
  }
593
609
  #endif
@@ -595,8 +611,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
595
611
  }
596
612
 
597
613
  void process_shaders() {
598
- std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
599
-
600
614
  // matmul
601
615
  for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
602
616
  // No coopmats
@@ -607,6 +621,10 @@ void process_shaders() {
607
621
  matmul_shaders(true, matmul_id_type, false, false, false);
608
622
  matmul_shaders(true, matmul_id_type, false, false, true);
609
623
 
624
+ // dot2 variants (scalar fp16 only)
625
+ matmul_shaders(true, matmul_id_type, false, false, false, true);
626
+ matmul_shaders(true, matmul_id_type, false, false, true, true);
627
+
610
628
  if (matmul_id_type != MatMulIdType::DEFAULT) {
611
629
  #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
612
630
  // Coopmat, fp32acc and fp16acc
@@ -622,77 +640,103 @@ void process_shaders() {
622
640
  }
623
641
  }
624
642
 
625
- // flash attention
626
- for (const auto& f16acc : {false, true}) {
627
- std::map<std::string, std::string> fa_base_dict = base_dict;
628
- fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
629
- fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
630
- if (f16acc) {
631
- fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
643
+ for (const bool& fp16 : {false, true}) {
644
+ std::map<std::string, std::string> base_dict;
645
+ if (fp16) {
646
+ base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV2", "f16vec2"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
647
+ } else {
648
+ base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"FLOAT_TYPEV4", "vec4"}};
632
649
  }
633
650
 
634
- for (const auto& tname : type_names) {
635
- if (tname == "bf16") continue;
651
+ // flash attention
652
+ for (const bool& f16acc : {false, true}) {
653
+ std::map<std::string, std::string> fa_base_dict = base_dict;
654
+ fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
655
+ fa_base_dict["ACC_TYPEV2"] = fp16 && f16acc ? "f16vec2" : "vec2";
656
+ fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
657
+ if (fp16 && f16acc) {
658
+ fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
659
+ }
636
660
 
661
+ if (fp16) {
637
662
  #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
638
- if (tname == "f16") {
639
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
640
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, true, f16acc);
641
- } else {
642
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
643
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
644
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
645
- }
663
+ string_to_spv("flash_attn_f32_f16", "flash_attn_cm2.comp",
664
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
646
665
  #endif
666
+
647
667
  #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
648
- if (tname == "f16") {
649
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
650
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"COOPMAT", "1"}}), true, true, false, f16acc);
651
- } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
652
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
653
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
654
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
655
- }
668
+ string_to_spv("flash_attn_f32_f16", "flash_attn_cm1.comp",
669
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
656
670
  #endif
657
- if (tname == "f16") {
658
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
659
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, false, f16acc);
660
- } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
661
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
662
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
663
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
664
671
  }
672
+
673
+ string_to_spv("flash_attn_f32_f16", "flash_attn.comp",
674
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
675
+
676
+ if (fp16) {
677
+ string_to_spv("flash_attn_f32_f16_dot2", "flash_attn.comp",
678
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DOT2_F16", "1"}}), fp16, false, false, f16acc);
679
+ }
680
+
681
+ #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
682
+ string_to_spv("flash_attn_f32_f16", "flash_attn.comp",
683
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"MMQ", "1"}, {"FA_MMQ_MIXED", "1"}}), fp16, false, false, f16acc, "_int8");
684
+ #endif
665
685
  }
666
686
  }
667
687
 
688
+ const std::map<std::string, std::string> fa_bf16_dict = {
689
+ {"FLOAT_TYPE", "bfloat16_t"},
690
+ {"FLOAT_TYPEV2", "bf16vec2"},
691
+ {"FLOAT_TYPEV4", "bf16vec4"},
692
+ {"ACC_TYPE", "float"},
693
+ {"ACC_TYPEV2", "vec2"},
694
+ {"ACC_TYPEV4", "vec4"},
695
+ {"BFLOAT16", "1"},
696
+ };
697
+
698
+ #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
699
+ string_to_spv("flash_attn_f32_f16_bf16", "flash_attn_cm1.comp",
700
+ merge_maps(fa_bf16_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}),
701
+ true, true, false, false);
702
+ #endif
703
+
704
+ #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
705
+ string_to_spv("flash_attn_f32_f16_bf16", "flash_attn_cm2.comp",
706
+ merge_maps(fa_bf16_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}),
707
+ true, false, true, false);
708
+ #endif
709
+
710
+ std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}};
711
+
668
712
  for (const auto& tname : type_names) {
669
713
  // mul mat vec
670
714
  std::string data_a_key = "DATA_A_" + to_uppercase(tname);
671
715
  std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
672
716
 
673
- string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
674
- string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
717
+ string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
718
+ string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}));
675
719
 
676
- string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
677
- string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
720
+ string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
721
+ string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
678
722
 
679
- string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
680
- string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
723
+ string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
724
+ string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
681
725
 
682
- string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
683
- string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
684
- string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
726
+ string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
727
+ string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
728
+ string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
685
729
 
686
730
  // mul mat vec with integer dot product
687
731
  #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
688
732
  if (is_legacy_quant(tname) || tname == "mxfp4" || is_k_quant(tname) || tname == "iq1_s" || tname == "iq1_m") {
689
- string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
690
- string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
691
- string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
733
+ string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}}));
734
+ string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
735
+ string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
692
736
 
693
- string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
694
- string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
695
- string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
737
+ string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}}));
738
+ string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
739
+ string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
696
740
  }
697
741
  #endif
698
742
 
@@ -713,9 +757,9 @@ void process_shaders() {
713
757
 
714
758
  string_to_spv("get_rows_i32", "get_rows.comp", {{"TEMP_TYPE", "uint"}, {"A_TYPE", "uint"}, {"B_TYPE", "int"}, {"D_TYPE", "uint"}});
715
759
 
716
- string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
717
- string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});
718
- string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});
760
+ string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPEV4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
761
+ string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPEV4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}});
762
+ string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPEV4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}});
719
763
 
720
764
  // Norms
721
765
  string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -723,7 +767,7 @@ void process_shaders() {
723
767
  string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
724
768
  string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
725
769
  string_to_spv("rms_norm_mul_rope_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
726
- string_to_spv("rms_norm_mul_rope_f32_f16_rte", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}, {"RTE16", "1"}}));
770
+ string_to_spv("rms_norm_mul_rope_f32_f16", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
727
771
  string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
728
772
  string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
729
773
 
@@ -732,6 +776,7 @@ void process_shaders() {
732
776
  string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
733
777
  string_to_spv("cpy_f16_f32", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
734
778
  string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
779
+ string_to_spv("cpy_bf16_f32","copy.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
735
780
  string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
736
781
  string_to_spv("contig_cpy_f32_i32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
737
782
  string_to_spv("contig_cpy_i32_f32", "contig_copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
@@ -739,23 +784,21 @@ void process_shaders() {
739
784
  string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
740
785
  string_to_spv("contig_cpy_f16_f32", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
741
786
  string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
787
+ string_to_spv("contig_cpy_bf16_f32","contig_copy.comp",{{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
742
788
  string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
743
789
  string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
744
790
 
745
791
  string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
746
792
  string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
747
793
 
748
- for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
794
+ for (std::string t : {"q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
749
795
  string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
750
- string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
751
796
  string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
752
797
  }
753
798
 
754
- for (std::string t : {"f32", "f16", "bf16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
755
- string_to_spv("set_rows_" + t + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
756
- string_to_spv("set_rows_" + t + "_i32_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
757
- string_to_spv("set_rows_" + t + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
758
- string_to_spv("set_rows_" + t + "_i64_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
799
+ for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
800
+ string_to_spv("set_rows_" + t + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
801
+ string_to_spv("set_rows_" + t + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
759
802
  }
760
803
 
761
804
  auto get_type_str = [](bool f16) {
@@ -772,12 +815,10 @@ void process_shaders() {
772
815
  for (auto src0_f16 : {false, true}) {
773
816
  for (auto src1_f16 : {false, true}) {
774
817
  for (auto dst_f16 : {false, true}) {
775
- for (auto rte : {false, true}) {
776
818
  auto source = op == "add_rms" ? std::string("add") : op;
777
- auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
819
+ auto name = op + get_suffix(src0_f16, src1_f16, dst_f16);
778
820
  auto add_rms = op == "add_rms" ? "1" : "0";
779
- string_to_spv(name.c_str(), source + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}, {"ADD_RMS" , add_rms}});
780
- }
821
+ string_to_spv(name.c_str(), source + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"ADD_RMS" , add_rms}});
781
822
  }
782
823
  }
783
824
  }
@@ -790,6 +831,8 @@ void process_shaders() {
790
831
  string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
791
832
  string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {});
792
833
 
834
+ string_to_spv("fa_mask_opt", "flash_attn_mask_opt.comp", {});
835
+
793
836
  string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {});
794
837
  string_to_spv("quantize_q8_1_subgroup", "quantize_q8_1.comp", {{"USE_SUBGROUPS", "1"}});
795
838
 
@@ -800,9 +843,11 @@ void process_shaders() {
800
843
 
801
844
  string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
802
845
 
803
- string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
846
+ string_to_spv("repeat_i32", "repeat.comp", {{"A_TYPE", "int32_t"}, {"D_TYPE", "int32_t"}});
804
847
  string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
805
848
 
849
+ string_to_spv("repeat_i16", "repeat.comp", {{"A_TYPE", "int16_t"}, {"D_TYPE", "int16_t"}});
850
+
806
851
  string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
807
852
 
808
853
  string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -823,14 +868,11 @@ void process_shaders() {
823
868
 
824
869
  string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
825
870
 
826
- for (auto rte : {false, true}) {
827
- std::string suffix = rte ? "_rte" : "";
828
- string_to_spv("exp_f16" + suffix, "exp.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
829
- string_to_spv("exp_f32" + suffix, "exp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"} , {"RTE16", rte ? "1" : "0"}});
871
+ string_to_spv("exp_f16", "exp.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
872
+ string_to_spv("exp_f32", "exp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
830
873
 
831
- string_to_spv("log_f16" + suffix, "log.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
832
- string_to_spv("log_f32" + suffix, "log.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
833
- }
874
+ string_to_spv("log_f16", "log.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
875
+ string_to_spv("log_f32", "log.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
834
876
  string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
835
877
  string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
836
878
  string_to_spv("gelu_erf_f16", "gelu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
@@ -853,8 +895,12 @@ void process_shaders() {
853
895
  string_to_spv("hardswish_f32", "hardswish.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
854
896
  string_to_spv("abs_f16", "abs.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
855
897
  string_to_spv("abs_f32", "abs.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
898
+ string_to_spv("elu_f16", "elu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
899
+ string_to_spv("elu_f32", "elu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
856
900
  string_to_spv("xielu_f16", "xielu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
857
901
  string_to_spv("xielu_f32", "xielu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
902
+ string_to_spv("sgn_f16", "sgn.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
903
+ string_to_spv("sgn_f32", "sgn.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
858
904
 
859
905
  string_to_spv("tri_f16", "tri.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
860
906
  string_to_spv("tri_f32", "tri.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -869,6 +915,7 @@ void process_shaders() {
869
915
  string_to_spv("add1_f32_f32", "add1.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
870
916
  string_to_spv("arange_f32", "arange.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
871
917
  string_to_spv("fill_f32", "fill.comp", {{"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
918
+ string_to_spv("fill_f16", "fill.comp", {{"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
872
919
  string_to_spv("step_f16", "step.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
873
920
  string_to_spv("step_f32", "step.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
874
921
  string_to_spv("round_f16", "round.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
@@ -880,21 +927,18 @@ void process_shaders() {
880
927
  string_to_spv("trunc_f16", "trunc.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
881
928
  string_to_spv("trunc_f32", "trunc.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
882
929
 
883
- for (auto rte : {false, true}) {
884
- std::string suffix = rte ? "_rte" : "";
885
- string_to_spv("geglu_f16" + suffix, "geglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
886
- string_to_spv("geglu_f32" + suffix, "geglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
887
- string_to_spv("reglu_f16" + suffix, "reglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
888
- string_to_spv("reglu_f32" + suffix, "reglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
889
- string_to_spv("swiglu_f16" + suffix, "swiglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
890
- string_to_spv("swiglu_f32" + suffix, "swiglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
891
- string_to_spv("swiglu_oai_f16" + suffix, "swiglu_oai.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
892
- string_to_spv("swiglu_oai_f32" + suffix, "swiglu_oai.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
893
- string_to_spv("geglu_erf_f16" + suffix, "geglu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
894
- string_to_spv("geglu_erf_f32" + suffix, "geglu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
895
- string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
896
- string_to_spv("geglu_quick_f32" + suffix,"geglu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
897
- }
930
+ string_to_spv("geglu_f16", "geglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
931
+ string_to_spv("geglu_f32", "geglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
932
+ string_to_spv("reglu_f16", "reglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
933
+ string_to_spv("reglu_f32", "reglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
934
+ string_to_spv("swiglu_f16", "swiglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
935
+ string_to_spv("swiglu_f32", "swiglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
936
+ string_to_spv("swiglu_oai_f16", "swiglu_oai.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
937
+ string_to_spv("swiglu_oai_f32", "swiglu_oai.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
938
+ string_to_spv("geglu_erf_f16", "geglu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
939
+ string_to_spv("geglu_erf_f32", "geglu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
940
+ string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
941
+ string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
898
942
 
899
943
  string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
900
944
  string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -914,25 +958,18 @@ void process_shaders() {
914
958
 
915
959
  string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
916
960
  string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
917
- string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
918
961
  string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
919
- string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
920
962
 
921
963
  string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
922
964
  string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
923
- string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
924
965
  string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
925
- string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
926
966
 
927
967
  string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
928
968
  string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
929
- string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
930
969
  string_to_spv("rope_multi_f32_f16", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
931
- string_to_spv("rope_multi_f32_f16_rte", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
932
970
 
933
971
  string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
934
972
  string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
935
- string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
936
973
 
937
974
  string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
938
975
  string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
@@ -942,6 +979,8 @@ void process_shaders() {
942
979
 
943
980
  string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
944
981
  string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
982
+ string_to_spv("fwht_f32", "fwht.comp", {});
983
+ string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
945
984
  string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
946
985
  string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
947
986
  string_to_spv("cumsum_multipass1_f32", "cumsum_multipass1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -955,7 +994,6 @@ void process_shaders() {
955
994
  std::string bda_def = bda ? "1" : "0";
956
995
  string_to_spv("im2col" + dim_str + "_f32" + bda_str, "im2col" + dim_str + ".comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"D_SIZE", "4"}, {"BDA", bda_def}}));
957
996
  string_to_spv("im2col" + dim_str + "_f32_f16" + bda_str, "im2col" + dim_str + ".comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"D_SIZE", "2"}, {"BDA", bda_def}}));
958
- string_to_spv("im2col" + dim_str + "_f32_f16_rte" + bda_str, "im2col" + dim_str + ".comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"D_SIZE", "2"}, {"RTE16", "1"}, {"BDA", bda_def}}));
959
997
  }
960
998
  }
961
999
 
@@ -963,12 +1001,20 @@ void process_shaders() {
963
1001
 
964
1002
  string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
965
1003
 
1004
+ string_to_spv("snake_f32", "snake.comp", {{"DATA_A_F32", "1"}, {"A_TYPE", "float"}, {"D_TYPE", "float"}});
1005
+ string_to_spv("snake_f16", "snake.comp", {{"DATA_A_F16", "1"}, {"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
1006
+ string_to_spv("snake_bf16", "snake.comp", {{"DATA_A_BF16", "1"}, {"DATA_D_BF16", "1"}, {"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
1007
+
966
1008
  string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
967
1009
 
968
1010
  string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
969
1011
 
970
1012
  string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
971
1013
 
1014
+ string_to_spv("gated_delta_net_f32", "gated_delta_net.comp", merge_maps(base_dict, {{"FLOAT_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}, {"USE_SUBGROUP_CLUSTERED", "1"}}));
1015
+ string_to_spv("gated_delta_net_f32_nocluster", "gated_delta_net.comp", merge_maps(base_dict, {{"FLOAT_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}, {"USE_SUBGROUP_CLUSTERED", "0"}}));
1016
+ string_to_spv("gated_delta_net_f32_shmem", "gated_delta_net.comp", merge_maps(base_dict, {{"FLOAT_TYPE", "float"}, {"USE_SUBGROUP_ADD", "0"}, {"USE_SUBGROUP_CLUSTERED", "0"}}));
1017
+
972
1018
  string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
973
1019
  string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
974
1020
 
@@ -987,8 +1033,16 @@ void process_shaders() {
987
1033
  string_to_spv(name + (unroll ? "_unroll" : ""), "conv2d_mm.comp", defines);
988
1034
  #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
989
1035
  if (unroll) {
990
- defines["COOPMAT2"] = "1";
991
- string_to_spv(name, "conv2d_mm.comp", defines, true, false, true);
1036
+ auto cm2_defines = defines;
1037
+ cm2_defines["COOPMAT2"] = "1";
1038
+ string_to_spv(name, "conv2d_mm.comp", cm2_defines, true, false, true);
1039
+ }
1040
+ #endif
1041
+ #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
1042
+ if (unroll) {
1043
+ auto cm1_defines = defines;
1044
+ cm1_defines["COOPMAT"] = "1";
1045
+ string_to_spv(name, "conv2d_mm.comp", cm1_defines, true, true, false);
992
1046
  }
993
1047
  #endif
994
1048
  }
@@ -1004,8 +1058,8 @@ void process_shaders() {
1004
1058
 
1005
1059
  string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
1006
1060
 
1007
- string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "0"}});
1008
- string_to_spv("multi_add_rms_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "1"}});
1061
+ string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"ADD_RMS" , "0"}});
1062
+ string_to_spv("multi_add_rms_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"ADD_RMS" , "1"}});
1009
1063
 
1010
1064
  string_to_spv("ssm_scan_f32", "ssm_scan.comp", {{"A_TYPE", "float"}});
1011
1065
  string_to_spv("ssm_scan_subgroup_f32", "ssm_scan.comp", {{"A_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
@@ -1058,8 +1112,8 @@ void write_output_files() {
1058
1112
 
1059
1113
  std::string suffixes[2] = {"_f32", "_f16"};
1060
1114
  for (std::string op : {"add", "sub", "mul", "div", "add_rms"}) {
1061
- hdr << "extern const void * " << op << "_data[2][2][2][2];\n";
1062
- hdr << "extern const uint64_t " << op << "_len[2][2][2][2];\n";
1115
+ hdr << "extern const void * " << op << "_data[2][2][2];\n";
1116
+ hdr << "extern const uint64_t " << op << "_len[2][2][2];\n";
1063
1117
 
1064
1118
  std::string op_file = op == "add_rms" ? "add.comp" : std::string(op) + ".comp";
1065
1119
  if (basename(input_filepath) != op_file) {
@@ -1067,8 +1121,8 @@ void write_output_files() {
1067
1121
  }
1068
1122
  std::stringstream data = make_generic_stringstream();
1069
1123
  std::stringstream len = make_generic_stringstream();
1070
- data << "const void * " << op << "_data[2][2][2][2] = ";
1071
- len << "const uint64_t " << op << "_len[2][2][2][2] = ";
1124
+ data << "const void * " << op << "_data[2][2][2] = ";
1125
+ len << "const uint64_t " << op << "_len[2][2][2] = ";
1072
1126
  for (uint32_t t0 = 0; t0 < 2; ++t0) {
1073
1127
  if (t0 == 0) {
1074
1128
  data << "{";
@@ -1084,20 +1138,10 @@ void write_output_files() {
1084
1138
  data << "{";
1085
1139
  len << "{";
1086
1140
  }
1087
- for (uint32_t rte = 0; rte < 2; ++rte) {
1088
- if (rte == 0) {
1089
- data << "{";
1090
- len << "{";
1091
- }
1092
- data << op << suffixes[t0] << suffixes[t1] << suffixes[t2] << ((rte != 0) ? "_rte" : "");
1093
- len << op << suffixes[t0] << suffixes[t1] << suffixes[t2] << ((rte != 0) ? "_rte" : "");
1094
- data << "_data,";
1095
- len << "_len,";
1096
- if (rte == 1) {
1097
- data << "}, ";
1098
- len << "}, ";
1099
- }
1100
- }
1141
+ data << op << suffixes[t0] << suffixes[t1] << suffixes[t2];
1142
+ len << op << suffixes[t0] << suffixes[t1] << suffixes[t2];
1143
+ data << "_data,";
1144
+ len << "_len,";
1101
1145
  if (t2 == 1) {
1102
1146
  data << "}, ";
1103
1147
  len << "}, ";