whispercpp 1.3.5 → 1.3.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (610) hide show
  1. checksums.yaml +4 -4
  2. data/LICENSE +1 -1
  3. data/README.md +99 -2
  4. data/ext/extconf.rb +1 -0
  5. data/ext/ruby_whisper.c +20 -4
  6. data/ext/ruby_whisper.h +30 -2
  7. data/ext/ruby_whisper_context.c +216 -124
  8. data/ext/ruby_whisper_context_params.c +163 -0
  9. data/ext/ruby_whisper_model.c +0 -1
  10. data/ext/ruby_whisper_params.c +0 -1
  11. data/ext/ruby_whisper_segment.c +0 -1
  12. data/ext/ruby_whisper_token.c +29 -9
  13. data/ext/ruby_whisper_transcribe.cpp +4 -1
  14. data/ext/ruby_whisper_vad_context.c +48 -1
  15. data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
  16. data/ext/ruby_whisper_vad_params.c +0 -1
  17. data/ext/ruby_whisper_vad_segment.c +0 -1
  18. data/ext/ruby_whisper_vad_segments.c +0 -1
  19. data/ext/sources/CMakeLists.txt +1 -1
  20. data/ext/sources/bindings/javascript/package.json +1 -1
  21. data/ext/sources/cmake/whisper-config.cmake.in +5 -40
  22. data/ext/sources/examples/bench/bench.cpp +23 -18
  23. data/ext/sources/examples/cli/cli.cpp +8 -0
  24. data/ext/sources/examples/common-ggml.cpp +2 -0
  25. data/ext/sources/examples/miniaudio.h +4507 -2131
  26. data/ext/sources/examples/server/server.cpp +18 -4
  27. data/ext/sources/examples/talk-llama/CMakeLists.txt +3 -2
  28. data/ext/sources/examples/talk-llama/llama-adapter.cpp +7 -13
  29. data/ext/sources/examples/talk-llama/llama-adapter.h +4 -3
  30. data/ext/sources/examples/talk-llama/llama-arch.cpp +335 -17
  31. data/ext/sources/examples/talk-llama/llama-arch.h +42 -0
  32. data/ext/sources/examples/talk-llama/llama-batch.cpp +3 -1
  33. data/ext/sources/examples/talk-llama/llama-chat.cpp +21 -1
  34. data/ext/sources/examples/talk-llama/llama-chat.h +1 -0
  35. data/ext/sources/examples/talk-llama/llama-context.cpp +508 -520
  36. data/ext/sources/examples/talk-llama/llama-context.h +27 -28
  37. data/ext/sources/examples/talk-llama/llama-cparams.h +5 -0
  38. data/ext/sources/examples/talk-llama/llama-ext.h +12 -0
  39. data/ext/sources/examples/talk-llama/llama-grammar.cpp +8 -8
  40. data/ext/sources/examples/talk-llama/llama-graph.cpp +583 -130
  41. data/ext/sources/examples/talk-llama/llama-graph.h +131 -10
  42. data/ext/sources/examples/talk-llama/llama-hparams.cpp +57 -40
  43. data/ext/sources/examples/talk-llama/llama-hparams.h +79 -10
  44. data/ext/sources/examples/talk-llama/llama-impl.cpp +4 -4
  45. data/ext/sources/examples/talk-llama/llama-impl.h +13 -1
  46. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +3 -1
  47. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +274 -89
  48. data/ext/sources/examples/talk-llama/llama-kv-cache.h +2 -3
  49. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +275 -0
  50. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +140 -0
  51. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +11 -13
  52. data/ext/sources/examples/talk-llama/llama-mmap.cpp +28 -11
  53. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +527 -119
  54. data/ext/sources/examples/talk-llama/llama-model-loader.h +35 -5
  55. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +60 -46
  56. data/ext/sources/examples/talk-llama/llama-model-saver.h +5 -2
  57. data/ext/sources/examples/talk-llama/llama-model.cpp +1365 -647
  58. data/ext/sources/examples/talk-llama/llama-model.h +72 -19
  59. data/ext/sources/examples/talk-llama/llama-quant.cpp +578 -346
  60. data/ext/sources/examples/talk-llama/{llama-sampling.cpp → llama-sampler.cpp} +190 -76
  61. data/ext/sources/examples/talk-llama/{llama-sampling.h → llama-sampler.h} +0 -2
  62. data/ext/sources/examples/talk-llama/llama-vocab.cpp +118 -48
  63. data/ext/sources/examples/talk-llama/llama-vocab.h +5 -0
  64. data/ext/sources/examples/talk-llama/llama.cpp +76 -22
  65. data/ext/sources/examples/talk-llama/llama.h +63 -30
  66. data/ext/sources/examples/talk-llama/models/afmoe.cpp +2 -3
  67. data/ext/sources/examples/talk-llama/models/apertus.cpp +3 -3
  68. data/ext/sources/examples/talk-llama/models/arcee.cpp +3 -3
  69. data/ext/sources/examples/talk-llama/models/arctic.cpp +4 -5
  70. data/ext/sources/examples/talk-llama/models/baichuan.cpp +4 -3
  71. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +1 -2
  72. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +3 -5
  73. data/ext/sources/examples/talk-llama/models/bert.cpp +13 -7
  74. data/ext/sources/examples/talk-llama/models/bitnet.cpp +9 -24
  75. data/ext/sources/examples/talk-llama/models/bloom.cpp +2 -2
  76. data/ext/sources/examples/talk-llama/models/chameleon.cpp +3 -3
  77. data/ext/sources/examples/talk-llama/models/chatglm.cpp +2 -2
  78. data/ext/sources/examples/talk-llama/models/codeshell.cpp +3 -3
  79. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +3 -3
  80. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +2 -2
  81. data/ext/sources/examples/talk-llama/models/command-r.cpp +2 -2
  82. data/ext/sources/examples/talk-llama/models/dbrx.cpp +4 -5
  83. data/ext/sources/examples/talk-llama/models/deci.cpp +3 -3
  84. data/ext/sources/examples/talk-llama/models/deepseek.cpp +4 -6
  85. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +24 -21
  86. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +445 -0
  87. data/ext/sources/examples/talk-llama/models/dots1.cpp +4 -6
  88. data/ext/sources/examples/talk-llama/models/dream.cpp +3 -3
  89. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +4 -6
  90. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +3 -3
  91. data/ext/sources/examples/talk-llama/models/eurobert.cpp +97 -0
  92. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +145 -0
  93. data/ext/sources/examples/talk-llama/models/exaone.cpp +3 -3
  94. data/ext/sources/examples/talk-llama/models/exaone4.cpp +3 -3
  95. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +2 -4
  96. data/ext/sources/examples/talk-llama/models/falcon.cpp +3 -3
  97. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +1 -1
  98. data/ext/sources/examples/talk-llama/models/gemma.cpp +1 -1
  99. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +1 -1
  100. data/ext/sources/examples/talk-llama/models/gemma3.cpp +1 -1
  101. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +7 -7
  102. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +3 -3
  103. data/ext/sources/examples/talk-llama/models/glm4.cpp +14 -7
  104. data/ext/sources/examples/talk-llama/models/gpt2.cpp +2 -2
  105. data/ext/sources/examples/talk-llama/models/gptneox.cpp +2 -2
  106. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +4 -5
  107. data/ext/sources/examples/talk-llama/models/granite.cpp +4 -5
  108. data/ext/sources/examples/talk-llama/models/grok.cpp +4 -4
  109. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +5 -7
  110. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +3 -3
  111. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +4 -5
  112. data/ext/sources/examples/talk-llama/models/internlm2.cpp +3 -3
  113. data/ext/sources/examples/talk-llama/models/jais.cpp +2 -2
  114. data/ext/sources/examples/talk-llama/models/jais2.cpp +123 -0
  115. data/ext/sources/examples/talk-llama/models/jamba.cpp +3 -3
  116. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +381 -0
  117. data/ext/sources/examples/talk-llama/models/lfm2.cpp +145 -124
  118. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +4 -4
  119. data/ext/sources/examples/talk-llama/models/llada.cpp +3 -3
  120. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +4 -4
  121. data/ext/sources/examples/talk-llama/models/llama.cpp +18 -11
  122. data/ext/sources/examples/talk-llama/models/maincoder.cpp +3 -3
  123. data/ext/sources/examples/talk-llama/models/{graph-context-mamba.cpp → mamba-base.cpp} +9 -3
  124. data/ext/sources/examples/talk-llama/models/mamba.cpp +1 -2
  125. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +11 -5
  126. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +14 -13
  127. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +4 -5
  128. data/ext/sources/examples/talk-llama/models/mistral3.cpp +4 -4
  129. data/ext/sources/examples/talk-llama/models/models.h +181 -46
  130. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +2 -9
  131. data/ext/sources/examples/talk-llama/models/mpt.cpp +2 -2
  132. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +26 -14
  133. data/ext/sources/examples/talk-llama/models/nemotron.cpp +3 -3
  134. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +2 -2
  135. data/ext/sources/examples/talk-llama/models/olmo.cpp +3 -3
  136. data/ext/sources/examples/talk-llama/models/olmo2.cpp +3 -3
  137. data/ext/sources/examples/talk-llama/models/olmoe.cpp +4 -4
  138. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +1 -1
  139. data/ext/sources/examples/talk-llama/models/openelm.cpp +3 -3
  140. data/ext/sources/examples/talk-llama/models/orion.cpp +3 -3
  141. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +122 -0
  142. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +3 -3
  143. data/ext/sources/examples/talk-llama/models/phi2.cpp +2 -2
  144. data/ext/sources/examples/talk-llama/models/phi3.cpp +3 -3
  145. data/ext/sources/examples/talk-llama/models/plamo.cpp +3 -3
  146. data/ext/sources/examples/talk-llama/models/plamo2.cpp +9 -5
  147. data/ext/sources/examples/talk-llama/models/plamo3.cpp +2 -2
  148. data/ext/sources/examples/talk-llama/models/plm.cpp +15 -14
  149. data/ext/sources/examples/talk-llama/models/qwen.cpp +2 -2
  150. data/ext/sources/examples/talk-llama/models/qwen2.cpp +3 -3
  151. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +4 -4
  152. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +3 -3
  153. data/ext/sources/examples/talk-llama/models/qwen3.cpp +12 -9
  154. data/ext/sources/examples/talk-llama/models/qwen35.cpp +381 -0
  155. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +422 -0
  156. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +15 -8
  157. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +84 -432
  158. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +9 -18
  159. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +8 -17
  160. data/ext/sources/examples/talk-llama/models/refact.cpp +2 -2
  161. data/ext/sources/examples/talk-llama/models/rnd1.cpp +4 -4
  162. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +2 -0
  163. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +2 -0
  164. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +3 -3
  165. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +4 -4
  166. data/ext/sources/examples/talk-llama/models/smollm3.cpp +3 -3
  167. data/ext/sources/examples/talk-llama/models/stablelm.cpp +2 -2
  168. data/ext/sources/examples/talk-llama/models/starcoder.cpp +2 -2
  169. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +3 -3
  170. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +165 -0
  171. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +2 -2
  172. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +2 -2
  173. data/ext/sources/examples/talk-llama/models/xverse.cpp +3 -3
  174. data/ext/sources/examples/talk-llama/unicode.cpp +21 -65
  175. data/ext/sources/ggml/CMakeLists.txt +9 -3
  176. data/ext/sources/ggml/include/ggml-backend.h +1 -1
  177. data/ext/sources/ggml/include/ggml-cann.h +1 -1
  178. data/ext/sources/ggml/include/ggml-cpu.h +5 -0
  179. data/ext/sources/ggml/include/ggml-openvino.h +37 -0
  180. data/ext/sources/ggml/include/ggml-opt.h +1 -1
  181. data/ext/sources/ggml/include/ggml-rpc.h +6 -1
  182. data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
  183. data/ext/sources/ggml/include/ggml.h +56 -9
  184. data/ext/sources/ggml/src/CMakeLists.txt +3 -0
  185. data/ext/sources/ggml/src/ggml-alloc.c +4 -9
  186. data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
  187. data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
  188. data/ext/sources/ggml/src/ggml-backend-reg.cpp +28 -86
  189. data/ext/sources/ggml/src/ggml-backend.cpp +5 -2
  190. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  191. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +6 -2
  192. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
  193. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
  194. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +348 -189
  195. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +40 -85
  196. data/ext/sources/ggml/src/ggml-cann/common.h +3 -4
  197. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +44 -62
  198. data/ext/sources/ggml/src/ggml-common.h +11 -0
  199. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +16 -11
  200. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +42 -19
  201. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
  202. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +85 -85
  203. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +85 -1
  204. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2744 -548
  205. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +1653 -0
  206. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1391 -0
  207. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -10
  208. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +9 -9
  209. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +118 -18
  210. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +107 -26
  211. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
  212. data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
  213. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +3 -0
  214. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +59 -12
  215. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +15 -0
  216. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
  217. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +965 -252
  218. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +584 -197
  219. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +903 -188
  220. data/ext/sources/ggml/src/ggml-cpu/ops.h +1 -0
  221. data/ext/sources/ggml/src/ggml-cpu/quants.c +40 -0
  222. data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
  223. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2890 -679
  224. data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
  225. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +136 -0
  226. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +111 -3
  227. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
  228. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +17 -0
  229. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +1 -1
  230. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +19 -10
  231. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +32 -30
  232. data/ext/sources/ggml/src/ggml-cuda/common.cuh +134 -18
  233. data/ext/sources/ggml/src/ggml-cuda/convert.cu +41 -27
  234. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +6 -3
  235. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +78 -64
  236. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +384 -143
  237. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +36 -22
  238. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +3 -3
  239. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +26 -5
  240. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
  241. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +127 -12
  242. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +263 -0
  243. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
  244. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +595 -200
  245. data/ext/sources/ggml/src/ggml-cuda/mean.cu +9 -8
  246. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +173 -6
  247. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
  248. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +158 -85
  249. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +34 -22
  250. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +127 -67
  251. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
  252. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +157 -65
  253. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +1 -0
  254. data/ext/sources/ggml/src/ggml-cuda/norm.cu +18 -76
  255. data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
  256. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +1 -1
  257. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +2 -16
  258. data/ext/sources/ggml/src/ggml-cuda/rope.cu +233 -133
  259. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
  260. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
  261. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +56 -32
  262. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  263. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +5 -0
  264. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  265. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +5 -0
  266. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  267. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  268. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  269. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +3 -3
  270. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +0 -1
  271. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +199 -135
  272. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
  273. data/ext/sources/ggml/src/ggml-cuda/unary.cu +55 -0
  274. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  275. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +31 -17
  276. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +10 -0
  277. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +82 -45
  278. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +334 -160
  279. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +7 -5
  280. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +328 -197
  281. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
  282. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +765 -234
  283. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +252 -0
  284. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +412 -265
  285. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +23 -23
  286. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +1 -1
  287. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.h → hex-dma.h} +28 -3
  288. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
  289. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  290. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
  291. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +1 -1
  292. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +27 -37
  293. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +6 -35
  294. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
  295. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +240 -0
  296. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
  297. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +251 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  302. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
  303. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  304. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
  305. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  306. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  307. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +20 -1347
  308. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +211 -13
  309. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +1119 -952
  310. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +254 -244
  311. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +36 -36
  312. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +155 -138
  313. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +339 -0
  314. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  315. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +209 -114
  316. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
  317. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  318. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  319. data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
  320. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
  321. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +6 -0
  322. data/ext/sources/ggml/src/ggml-impl.h +62 -0
  323. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
  324. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
  325. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
  326. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
  327. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +274 -73
  328. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +22 -4
  329. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +102 -36
  330. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +174 -23
  331. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +580 -280
  332. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +5 -4
  333. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +320 -107
  334. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1068 -825
  335. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +19 -1
  336. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +3108 -636
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +41 -99
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +45 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +204 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
  365. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
  366. data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  367. data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
  368. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  369. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +975 -0
  370. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
  371. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +373 -0
  372. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
  373. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1110 -0
  374. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +884 -0
  375. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
  376. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
  377. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  378. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  379. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  380. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  381. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
  382. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
  383. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
  384. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
  385. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
  386. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
  387. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
  388. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
  389. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
  390. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
  391. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
  392. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +123 -0
  393. data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
  394. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
  395. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
  396. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
  397. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  398. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
  399. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +46 -0
  400. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +39 -0
  401. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +123 -0
  402. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +17 -0
  403. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  404. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  405. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  406. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  407. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +293 -0
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +226 -0
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +85 -0
  412. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +823 -0
  413. data/ext/sources/ggml/src/ggml-openvino/utils.h +123 -0
  414. data/ext/sources/ggml/src/ggml-quants.c +96 -5
  415. data/ext/sources/ggml/src/ggml-quants.h +3 -0
  416. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +15 -88
  417. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
  418. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +1 -0
  419. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
  420. data/ext/sources/ggml/src/ggml-sycl/common.hpp +315 -10
  421. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +69 -1
  422. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +22 -1
  423. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
  424. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
  425. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +78 -68
  426. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
  427. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1179 -0
  428. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +55 -0
  429. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1338 -0
  430. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +667 -0
  431. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +225 -0
  432. data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
  433. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +309 -0
  434. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +8 -0
  435. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +316 -51
  436. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
  437. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
  438. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
  439. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +1 -1
  440. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
  441. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
  442. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
  443. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  444. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  445. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  446. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  447. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  448. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  449. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  450. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  451. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  452. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +7 -0
  453. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +7 -0
  454. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +7 -0
  455. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +7 -0
  456. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +7 -0
  457. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +7 -0
  458. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +7 -0
  459. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +7 -0
  460. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +7 -0
  461. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +7 -0
  462. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +7 -0
  463. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +7 -0
  464. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +7 -0
  465. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +7 -0
  466. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +7 -0
  467. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +7 -0
  468. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +7 -0
  469. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +7 -0
  470. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +7 -0
  471. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +7 -0
  472. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +7 -0
  473. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +7 -0
  474. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +7 -0
  475. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +7 -0
  476. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +7 -0
  477. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +7 -0
  478. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +7 -0
  479. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +7 -0
  480. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +7 -0
  481. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +7 -0
  482. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +7 -0
  483. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +7 -0
  484. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +7 -0
  485. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +7 -0
  486. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +7 -0
  487. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +7 -0
  488. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +13 -0
  489. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
  490. data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  491. data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  492. data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  493. data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  494. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  495. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  496. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  497. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  498. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  499. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  500. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  501. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  502. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  503. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  504. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  505. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  506. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  507. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  508. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  509. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  510. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  511. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
  512. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
  513. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  514. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
  515. data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  516. data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  517. data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  518. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
  519. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  520. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  521. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  522. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  523. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  524. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  525. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  526. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
  527. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  528. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  529. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  530. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
  531. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
  532. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +1 -1
  533. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1250 -465
  534. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
  535. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
  536. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +374 -170
  537. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +66 -22
  538. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +389 -201
  539. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +106 -58
  540. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  541. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
  542. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +128 -0
  543. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
  544. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
  545. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -3
  546. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +8 -4
  547. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  548. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
  549. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +3 -3
  550. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
  551. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +36 -63
  552. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
  553. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
  554. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
  555. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +10 -5
  556. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
  557. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
  558. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +16 -10
  559. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +55 -35
  560. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1314 -109
  561. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1660 -1371
  562. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  563. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  564. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  565. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +141 -0
  566. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +65 -72
  567. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +75 -0
  568. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +6 -0
  569. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  570. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +40 -5
  571. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +105 -60
  572. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +53 -259
  573. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat.tmpl.wgsl → mul_mat.wgsl} +68 -257
  574. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +692 -23
  575. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_reg_tile.tmpl.wgsl → mul_mat_reg_tile.wgsl} +28 -128
  576. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +31 -137
  577. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +480 -0
  578. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  579. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
  580. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +9 -36
  581. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +40 -12
  582. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  583. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
  584. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +6 -1
  585. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
  586. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +9 -6
  587. data/ext/sources/ggml/src/ggml.c +167 -33
  588. data/ext/sources/ggml/src/gguf.cpp +229 -44
  589. data/ext/sources/src/whisper.cpp +6 -28
  590. data/sig/whisper.rbs +43 -2
  591. data/test/test_context_params.rb +82 -0
  592. data/test/test_token.rb +11 -0
  593. data/test/test_vad_context.rb +58 -8
  594. data/test/test_whisper.rb +20 -0
  595. data/whispercpp.gemspec +1 -1
  596. metadata +240 -28
  597. data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
  598. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
  599. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
  600. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
  601. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
  602. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
  603. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
  604. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
  605. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
  606. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
  607. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
  608. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
  609. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
  610. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl ADDED
@@ -0,0 +1,283 @@
1
+ #pragma OPENCL EXTENSION cl_khr_fp16 : enable
2
+ #pragma OPENCL EXTENSION cl_khr_subgroups : enable
3
+
4
+ #ifdef cl_qcom_reqd_sub_group_size
5
+ #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
6
+ #define ADRENO_GPU 1
7
+ #define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
8
+ #endif
9
+
10
+ #define QK4_0 32
11
+ #define NSUBGROUPS 4
12
+ #define SUBGROUP_SIZE 64
13
+
14
+ #define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, minv, y) \
15
+ float shared_y; \
16
+ shared_y = sub_group_broadcast(y.s0, 0); \
17
+ total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
18
+ total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
19
+ shared_y = sub_group_broadcast(y.s1, 0); \
20
+ total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
21
+ total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
22
+ shared_y = sub_group_broadcast(y.s2, 0); \
23
+ total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
24
+ total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
25
+ shared_y = sub_group_broadcast(y.s3, 0); \
26
+ total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
27
+ total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
28
+ shared_y = sub_group_broadcast(y.s4, 0); \
29
+ total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
30
+ total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
31
+ shared_y = sub_group_broadcast(y.s5, 0); \
32
+ total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
33
+ total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
34
+ shared_y = sub_group_broadcast(y.s6, 0); \
35
+ total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
36
+ total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
37
+ shared_y = sub_group_broadcast(y.s7, 0); \
38
+ total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
39
+ total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
40
+ shared_y = sub_group_broadcast(y.s0, 1); \
41
+ total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
42
+ total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
43
+ shared_y = sub_group_broadcast(y.s1, 1); \
44
+ total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
45
+ total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
46
+ shared_y = sub_group_broadcast(y.s2, 1); \
47
+ total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
48
+ total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
49
+ shared_y = sub_group_broadcast(y.s3, 1); \
50
+ total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
51
+ total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
52
+ shared_y = sub_group_broadcast(y.s4, 1); \
53
+ total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
54
+ total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
55
+ shared_y = sub_group_broadcast(y.s5, 1); \
56
+ total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
57
+ total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
58
+ shared_y = sub_group_broadcast(y.s6, 1); \
59
+ total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
60
+ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
61
+ shared_y = sub_group_broadcast(y.s7, 1); \
62
+ total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
63
+ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
64
+
65
+
66
+ #define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, minv, y) \
67
+ shared_y = sub_group_broadcast(y.s0, 2); \
68
+ total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
69
+ total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
70
+ shared_y = sub_group_broadcast(y.s1, 2); \
71
+ total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
72
+ total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
73
+ shared_y = sub_group_broadcast(y.s2, 2); \
74
+ total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
75
+ total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
76
+ shared_y = sub_group_broadcast(y.s3, 2); \
77
+ total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
78
+ total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
79
+ shared_y = sub_group_broadcast(y.s4, 2); \
80
+ total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
81
+ total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
82
+ shared_y = sub_group_broadcast(y.s5, 2); \
83
+ total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
84
+ total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
85
+ shared_y = sub_group_broadcast(y.s6, 2); \
86
+ total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
87
+ total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
88
+ shared_y = sub_group_broadcast(y.s7, 2); \
89
+ total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
90
+ total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
91
+ shared_y = sub_group_broadcast(y.s0, 3); \
92
+ total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
93
+ total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
94
+ shared_y = sub_group_broadcast(y.s1, 3); \
95
+ total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
96
+ total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
97
+ shared_y = sub_group_broadcast(y.s2, 3); \
98
+ total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
99
+ total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
100
+ shared_y = sub_group_broadcast(y.s3, 3); \
101
+ total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
102
+ total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
103
+ shared_y = sub_group_broadcast(y.s4, 3); \
104
+ total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y; \
105
+ total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y; \
106
+ shared_y = sub_group_broadcast(y.s5, 3); \
107
+ total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y; \
108
+ total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y; \
109
+ shared_y = sub_group_broadcast(y.s6, 3); \
110
+ total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y; \
111
+ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y; \
112
+ shared_y = sub_group_broadcast(y.s7, 3); \
113
+ total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y; \
114
+ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y; \
115
+
116
+
117
+ #define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, minv, y) \
118
+ float8 shared_y; \
119
+ shared_y = sub_group_broadcast(y, 0); \
120
+ total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y.s0; \
121
+ total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s1; \
122
+ total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s2; \
123
+ total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
124
+ total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y.s4; \
125
+ total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s5; \
126
+ total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s6; \
127
+ total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
128
+ total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y.s0; \
129
+ total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s1; \
130
+ total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s2; \
131
+ total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
132
+ total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y.s4; \
133
+ total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s5; \
134
+ total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s6; \
135
+ total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
136
+ shared_y = sub_group_broadcast(y, 1); \
137
+ total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y.s0; \
138
+ total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s1; \
139
+ total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s2; \
140
+ total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
141
+ total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y.s4; \
142
+ total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s5; \
143
+ total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s6; \
144
+ total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
145
+ total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y.s0; \
146
+ total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s1; \
147
+ total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s2; \
148
+ total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
149
+ total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y.s4; \
150
+ total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s5; \
151
+ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s6; \
152
+ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
153
+
154
+
155
+ #define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, minv, y) \
156
+ shared_y = sub_group_broadcast(y, 2); \
157
+ total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 + minv.s0) * shared_y.s0; \
158
+ total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s1; \
159
+ total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s2; \
160
+ total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
161
+ total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 + minv.s0) * shared_y.s4; \
162
+ total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s5; \
163
+ total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s6; \
164
+ total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
165
+ total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 + minv.s1) * shared_y.s0; \
166
+ total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s1; \
167
+ total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s2; \
168
+ total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
169
+ total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 + minv.s1) * shared_y.s4; \
170
+ total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s5; \
171
+ total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s6; \
172
+ total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
173
+ shared_y = sub_group_broadcast(y, 3); \
174
+ total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 + minv.s0) * shared_y.s0; \
175
+ total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s1; \
176
+ total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s2; \
177
+ total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s3; \
178
+ total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 + minv.s0) * shared_y.s4; \
179
+ total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 + minv.s0) * shared_y.s5; \
180
+ total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 + minv.s0) * shared_y.s6; \
181
+ total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 + minv.s0) * shared_y.s7; \
182
+ total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 + minv.s1) * shared_y.s0; \
183
+ total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s1; \
184
+ total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s2; \
185
+ total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s3; \
186
+ total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 + minv.s1) * shared_y.s4; \
187
+ total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 + minv.s1) * shared_y.s5; \
188
+ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 + minv.s1) * shared_y.s6; \
189
+ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 + minv.s1) * shared_y.s7; \
190
+
191
+ #ifdef ADRENO_GPU
192
+ REQD_SUBGROUP_SIZE_64
193
+ #endif
194
+ kernel void kernel_gemv_noshuffle_q4_1_f32(
195
+ read_only image1d_buffer_t src0_q,
196
+ global half2 * src0_d,
197
+ global half2 * src0_m,
198
+ read_only image1d_buffer_t src1,
199
+ global float * dst,
200
+ ulong offsetd,
201
+ int ne00,
202
+ int ne01)
203
+ {
204
+ uint groupId = get_local_id(1);
205
+ uint gid = get_global_id(0);
206
+ ushort slid = get_sub_group_local_id();
207
+
208
+ uint K = ne00;
209
+ uint M = ne01;
210
+
211
+ uint LINE_STRIDE_A = M / 2;
212
+ uint BLOCK_STRIDE_A = NSUBGROUPS * M;
213
+
214
+ private uint4 regA;
215
+ private half2 regS;
216
+ private half2 regM;
217
+ private float8 regB;
218
+
219
+ private float2 totalSum = (float2)(0.0f);
220
+
221
+ // loop along K in block granularity, skip 4 blocks every iter
222
+ for (uint k = groupId; k < (K / QK4_0); k += NSUBGROUPS) {
223
+ regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
224
+ regM = src0_m[gid + k * LINE_STRIDE_A]; // each fiber loads min of two rows
225
+ // first 4 fibers in each wave load 8 B values to its private scope
226
+ if (slid < 4) {
227
+ regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
228
+ regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
229
+ }
230
+
231
+ // load half weights for two blocks in consecutive rows
232
+ regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
233
+ regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
234
+ regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
235
+ regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
236
+ #ifdef VECTOR_SUB_GROUP_BROADCAT
237
+ dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
238
+ #else
239
+ dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
240
+ #endif // VECTOR_SUB_GROUP_BROADCAT
241
+
242
+ regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
243
+ regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
244
+ regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
245
+ regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
246
+ #ifdef VECTOR_SUB_GROUP_BROADCAT
247
+ dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
248
+ #else
249
+ dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
250
+ #endif // VECTOR_SUB_GROUP_BROADCAT
251
+ }
252
+
253
+ // reduction in local memory, assumes #wave=4
254
+ local float2 reduceLM[SUBGROUP_SIZE * 3];
255
+ if (groupId == 1) {
256
+ reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
257
+ }
258
+ if (groupId == 2) {
259
+ reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
260
+ }
261
+ if (groupId == 3) {
262
+ reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
263
+ }
264
+
265
+ barrier(CLK_LOCAL_MEM_FENCE);
266
+
267
+ if (groupId == 0) {
268
+ totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
269
+ }
270
+ if (groupId == 0) {
271
+ totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
272
+ }
273
+ if (groupId == 0) {
274
+ totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
275
+ }
276
+
277
+ // 2 outputs per fiber in wave 0
278
+ if (groupId == 0) {
279
+ dst = (global float*)((global char*)dst + offsetd);
280
+ vstore2(totalSum, 0, &(dst[gid * 2]));
281
+ }
282
+
283
+ }
data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl ADDED
@@ -0,0 +1,71 @@
1
+ #ifdef cl_intel_required_subgroup_size
2
+ #pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
3
+ #define INTEL_GPU 1
4
+ #define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
5
+ #define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
6
+ #elif defined(cl_qcom_reqd_sub_group_size)
7
+ #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
8
+ #define ADRENO_GPU 1
9
+ #define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
10
+ #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
11
+ #endif
12
+
13
+ #ifdef INTEL_GPU
14
+ REQD_SUBGROUP_SIZE_32
15
+ #elif defined (ADRENO_GPU)
16
+ REQD_SUBGROUP_SIZE_64
17
+ #endif
18
+ kernel void kernel_l2_norm_f32(
19
+ global void * src0,
20
+ ulong offset0,
21
+ global float * dst,
22
+ ulong offsetd,
23
+ int ne00,
24
+ int ne01,
25
+ int ne02,
26
+ int ne03,
27
+ ulong nb01,
28
+ ulong nb02,
29
+ ulong nb03,
30
+ float eps,
31
+ local float * sum
32
+ ) {
33
+ src0 = (global void*)((global char*)src0 + offset0);
34
+ dst = (global float*)((global char*)dst + offsetd);
35
+
36
+ int i03 = get_group_id(2);
37
+ int i02 = get_group_id(1);
38
+ int i01 = get_group_id(0);
39
+
40
+ global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
41
+ global float * y = (global float *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
42
+
43
+ float sumf = 0;
44
+
45
+ // parallel sum
46
+ for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
47
+ sumf += x[i00] * x[i00];
48
+ }
49
+ sumf = sub_group_reduce_add(sumf);
50
+
51
+ if (get_sub_group_local_id() == 0) {
52
+ sum[get_sub_group_id()] = sumf;
53
+ }
54
+
55
+ barrier(CLK_LOCAL_MEM_FENCE);
56
+
57
+ // broadcast
58
+ for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
59
+ if (get_local_id(0) < i) {
60
+ sum[get_local_id(0)] += sum[get_local_id(0) + i];
61
+ }
62
+ }
63
+
64
+ barrier(CLK_LOCAL_MEM_FENCE);
65
+
66
+ const float scale = 1.0f/max(sqrt(sum[0]), eps);
67
+
68
+ for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
69
+ y[i00] = x[i00] * scale;
70
+ }
71
+ }
data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl CHANGED
@@ -1,8 +1,13 @@
1
+ #pragma OPENCL EXTENSION cl_khr_fp16 : enable
2
+ #pragma OPENCL EXTENSION cl_khr_subgroups : enable
1
3
 
4
+ // Most devices have max workgroup size of 1024, so this is enough for subgroup
5
+ // sizes of 16, 32, 64 and 128. Increase this value for smaller subgroups sizes
6
+ #define MAX_SUBGROUPS 64
2
7
  kernel void kernel_mean_f32(
3
- global float * src0,
8
+ global char * src0,
4
9
  ulong offset0,
5
- global float * dst,
10
+ global char * dst,
6
11
  ulong offsetd,
7
12
  int ne00,
8
13
  int ne01,
@@ -15,25 +20,121 @@ kernel void kernel_mean_f32(
15
20
  ulong nb2,
16
21
  ulong nb3
17
22
  ) {
18
- src0 = (global float *)((global char *)src0 + offset0);
19
- dst = (global float *)((global char *)dst + offsetd);
23
+ src0 = src0 + offset0;
24
+ dst = dst + offsetd;
20
25
 
21
- int i3 = get_global_id(2);
22
- int i2 = get_global_id(1);
23
- int i1 = get_global_id(0);
26
+ const int i3 = get_group_id(2);
27
+ const int i2 = get_group_id(1);
28
+ const int i1 = get_group_id(0);
29
+
30
+ const int lid = get_local_id(0);
31
+ const int lsize = get_local_size(0);
32
+
33
+ const uint sg_size = get_sub_group_size();
34
+ const uint sg_id = get_sub_group_id();
35
+ const uint sg_lid = get_sub_group_local_id();
36
+
37
+ __local float lmem[MAX_SUBGROUPS];
24
38
 
25
39
  if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
26
40
  return;
27
41
  }
28
42
 
29
- global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
30
- global float * dst_row = (global float *) ((global char *) dst + i1*nb1 + i2*nb2 + i3*nb3);
43
+ if(sg_id == 0){
44
+ lmem[sg_lid] = 0.0f;
45
+ }
31
46
 
32
- float row_sum = 0;
47
+ global float * src_row = (global float *) (src0 + i1*nb01 + i2*nb02 + i3*nb03);
48
+ global float * dst_row = (global float *) (dst + i1*nb1 + i2*nb2 + i3*nb3);
33
49
 
34
- for (int i0 = 0; i0 < ne00; i0++) {
35
- row_sum += src_row[i0];
50
+ float sumf = 0.0f;
51
+
52
+ for (int i0 = lid; i0 < ne00; i0 += lsize) {
53
+ sumf += src_row[i0];
36
54
  }
37
55
 
38
- dst_row[0] = row_sum / ne00;
56
+ sumf = sub_group_reduce_add(sumf);
57
+
58
+ barrier(CLK_LOCAL_MEM_FENCE);
59
+
60
+ if(sg_lid == 0){
61
+ lmem[sg_id] = sumf;
62
+ }
63
+
64
+ barrier(CLK_LOCAL_MEM_FENCE);
65
+
66
+ sumf = lmem[sg_lid];
67
+ sumf = sub_group_reduce_add(sumf);
68
+
69
+ if (lid == 0) {
70
+ dst_row[0] = sumf / ne00;
71
+ }
72
+ }
73
+
74
+ kernel void kernel_mean_f32_4(
75
+ global char * src0,
76
+ ulong offset0,
77
+ global char * dst,
78
+ ulong offsetd,
79
+ int ne00,
80
+ int ne01,
81
+ int ne02,
82
+ int ne03,
83
+ ulong nb01,
84
+ ulong nb02,
85
+ ulong nb03,
86
+ ulong nb1,
87
+ ulong nb2,
88
+ ulong nb3
89
+ ) {
90
+ src0 = src0 + offset0;
91
+ dst = dst + offsetd;
92
+
93
+ const int i3 = get_group_id(2);
94
+ const int i2 = get_group_id(1);
95
+ const int i1 = get_group_id(0);
96
+
97
+ const int lid = get_local_id(0);
98
+ const int lsize = get_local_size(0);
99
+
100
+ const uint sg_size = get_sub_group_size();
101
+ const uint sg_id = get_sub_group_id();
102
+ const uint sg_lid = get_sub_group_local_id();
103
+
104
+ __local float lmem[MAX_SUBGROUPS];
105
+
106
+ if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
107
+ return;
108
+ }
109
+
110
+ if(sg_id == 0){
111
+ lmem[sg_lid] = 0.0f;
112
+ }
113
+
114
+ global float4 * src_row = (global float4 *) (src0 + i1*nb01 + i2*nb02 + i3*nb03);
115
+ global float * dst_row = (global float *) (dst + i1*nb1 + i2*nb2 + i3*nb3);
116
+
117
+ float4 sum_vec = (float4)0.0f;
118
+
119
+ for (int i0 = lid; i0 < ne00 / 4; i0 += lsize) {
120
+ sum_vec += src_row[i0];
121
+ }
122
+
123
+ float sumf = dot(sum_vec, (float4)(1.0f));
124
+ sumf = sub_group_reduce_add(sumf);
125
+
126
+ barrier(CLK_LOCAL_MEM_FENCE);
127
+
128
+ if(sg_lid == 0){
129
+ lmem[sg_id] = sumf;
130
+ }
131
+
132
+ barrier(CLK_LOCAL_MEM_FENCE);
133
+
134
+ sumf = lmem[sg_lid];
135
+ sumf = sub_group_reduce_add(sumf);
136
+
137
+ if (lid == 0) {
138
+ dst_row[0] = sumf / ne00;
139
+ }
39
140
  }
data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl ADDED
@@ -0,0 +1,163 @@
1
+ #pragma OPENCL EXTENSION cl_khr_fp16 : enable
2
+
3
+ #define LOAD_VEC_A 8
4
+ #define LOAD_VEC_B 4
5
+
6
+ #define BM 64
7
+ #define BN 64
8
+ #define BK 32
9
+ #define TM 4
10
+ #define TN 8
11
+
12
+ kernel void kernel_mul_mm_q4_0_f32_l4_lm(
13
+ global uchar4 * src0_q,
14
+ global half * src0_d,
15
+ global float4 * src1,
16
+ ulong offset1,
17
+ global float * dst,
18
+ ulong offsetd,
19
+
20
+ int ne00,
21
+ int ne01,
22
+ int ne02,
23
+ int ne11,
24
+ int ne12,
25
+
26
+ int stride_a,
27
+ int stride_b,
28
+ int stride_d,
29
+
30
+ int batch_stride_a,
31
+ int batch_stride_b,
32
+ int batch_stride_d,
33
+
34
+ int r2,
35
+ int r3
36
+ ) {
37
+ src1 = (global float4*)((global char*)src1 + offset1);
38
+ dst = (global float *)((global char*)dst + offsetd);
39
+
40
+ local float buf_a[BM * BK];
41
+ local float buf_b[BN * BK];
42
+
43
+ const int batch_idx = get_global_id(2);
44
+
45
+ const int i13 = batch_idx / ne12;
46
+ const int i12 = batch_idx % ne12;
47
+
48
+ const int i03 = i13 / r3;
49
+ const int i02 = i12 / r2;
50
+
51
+ const int batch_idx_a = i03 * ne02 + i02;
52
+
53
+ const int ir = get_group_id(0);
54
+ const int ic = get_group_id(1);
55
+
56
+ const int tid = get_local_id(0);
57
+ const int th_r = tid % (BM / TM);
58
+ const int th_c = tid / (BM / TM);
59
+
60
+ const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
61
+ const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
62
+ const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
63
+ const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
64
+
65
+ const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
66
+ const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
67
+
68
+ int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
69
+ int pos_b = (batch_idx * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
70
+
71
+ float sums[TM * TN];
72
+ float cache_a[TM];
73
+ float cache_b[TN];
74
+
75
+ for (int i = 0; i < TM * TN; i++) {
76
+ sums[i] = 0.0f;
77
+ }
78
+
79
+ for (int block = 0; block < ne00; block += BK) {
80
+ for (int l = 0; l < BM; l += loadstride_a) {
81
+ if (ir*BM + loadc_a + l < ne01) {
82
+ int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
83
+ int ib = idx / 4;
84
+ int iqs = idx % 4;
85
+
86
+ float d = (float)src0_d[ib];
87
+ global uchar4 * qs = src0_q + ib*4 + iqs;
88
+ uchar4 q = *qs;
89
+ float4 v1 = (convert_float4((uchar4)((q.s0 )&0x0F, (q.s1 )&0x0F, (q.s2 )&0x0F, (q.s3 )&0x0F)) - 8.0f)*d;
90
+ float4 v2 = (convert_float4((uchar4)((q.s0>>4)&0x0F, (q.s1>>4)&0x0F, (q.s2>>4)&0x0F, (q.s3>>4)&0x0F)) - 8.0f)*d;
91
+
92
+ buf_a[(loadr_a * 4 + 0) * BM + loadc_a + l] = v1.s0;
93
+ buf_a[(loadr_a * 4 + 1) * BM + loadc_a + l] = v1.s1;
94
+ buf_a[(loadr_a * 4 + 2) * BM + loadc_a + l] = v1.s2;
95
+ buf_a[(loadr_a * 4 + 3) * BM + loadc_a + l] = v1.s3;
96
+ buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = v2.s0;
97
+ buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = v2.s1;
98
+ buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = v2.s2;
99
+ buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = v2.s3;
100
+ } else {
101
+ buf_a[(loadr_a * 4 + 0) * BM + loadc_a + l] = 0.0f;
102
+ buf_a[(loadr_a * 4 + 1) * BM + loadc_a + l] = 0.0f;
103
+ buf_a[(loadr_a * 4 + 2) * BM + loadc_a + l] = 0.0f;
104
+ buf_a[(loadr_a * 4 + 3) * BM + loadc_a + l] = 0.0f;
105
+ buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = 0.0f;
106
+ buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = 0.0f;
107
+ buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = 0.0f;
108
+ buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = 0.0f;
109
+ }
110
+ }
111
+
112
+ for (int l = 0; l < BN; l += loadstride_b) {
113
+ if (ic*BN + loadc_b + l < ne11) {
114
+ int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
115
+ buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
116
+ buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
117
+ buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
118
+ buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
119
+ } else {
120
+ buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
121
+ buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
122
+ buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
123
+ buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
124
+ }
125
+ }
126
+
127
+ barrier(CLK_LOCAL_MEM_FENCE);
128
+
129
+ pos_a += BK / LOAD_VEC_A;
130
+ pos_b += BK / LOAD_VEC_B;
131
+
132
+ for (int i = 0; i < BK; i++) {
133
+ for (int j = 0; j < TM; j++) {
134
+ cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
135
+ }
136
+
137
+ for (int j = 0; j < TN; j++) {
138
+ cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
139
+ }
140
+
141
+ for (int cc = 0; cc < TN; cc++) {
142
+ for (int cr = 0; cr < TM; cr++) {
143
+ const int sums_idx = cc*TM + cr;
144
+ sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
145
+ }
146
+ }
147
+ }
148
+ barrier(CLK_LOCAL_MEM_FENCE);
149
+ }
150
+
151
+ const int dr = ir * BM + th_r * TM;
152
+ const int dc = ic * BN + th_c * TN;
153
+
154
+ const int offsets = batch_idx * batch_stride_d;
155
+
156
+ for (int cc = 0; cc < TN; cc++) {
157
+ for (int cr = 0; cr < TM; cr++) {
158
+ if (dr + cr < ne01 && dc + cc < ne11) {
159
+ dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
160
+ }
161
+ }
162
+ }
163
+ }