whispercpp 1.3.6 → 1.3.8

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 (965) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/README.md +43 -9
  5. data/Rakefile +18 -3
  6. data/ext/dependencies.rb +10 -4
  7. data/ext/dependencies_for_windows.rb +17 -0
  8. data/ext/extconf.rb +20 -8
  9. data/ext/options.rb +54 -14
  10. data/ext/options_for_windows.rb +51 -0
  11. data/ext/ruby_whisper.c +35 -42
  12. data/ext/ruby_whisper.h +141 -0
  13. data/ext/ruby_whisper_context.c +157 -29
  14. data/ext/ruby_whisper_log_queue.c +180 -0
  15. data/ext/ruby_whisper_log_settable.h +46 -0
  16. data/ext/ruby_whisper_parakeet.c +49 -0
  17. data/ext/ruby_whisper_parakeet_context.c +304 -0
  18. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  19. data/ext/ruby_whisper_parakeet_model.c +84 -0
  20. data/ext/ruby_whisper_parakeet_params.c +548 -0
  21. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  22. data/ext/ruby_whisper_parakeet_token.c +188 -0
  23. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  24. data/ext/ruby_whisper_params.c +265 -73
  25. data/ext/ruby_whisper_segment.c +6 -6
  26. data/ext/ruby_whisper_transcribe.cpp +23 -15
  27. data/ext/ruby_whisper_vad_context.c +30 -10
  28. data/ext/ruby_whisper_vad_context_detect.cpp +8 -9
  29. data/ext/ruby_whisper_vad_params.c +4 -4
  30. data/ext/ruby_whisper_vad_segment.c +2 -2
  31. data/ext/sources/CMakeLists.txt +42 -3
  32. data/ext/sources/CMakePresets.json +95 -0
  33. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  34. data/ext/sources/cmake/parakeet.pc.in +10 -0
  35. data/ext/sources/cmake/whisper.pc.in +2 -2
  36. data/ext/sources/examples/CMakeLists.txt +4 -2
  37. data/ext/sources/examples/bench/bench.cpp +1 -1
  38. data/ext/sources/examples/cli/cli.cpp +52 -10
  39. data/ext/sources/examples/common-ggml.cpp +4 -0
  40. data/ext/sources/examples/common-whisper.cpp +139 -67
  41. data/ext/sources/examples/common-whisper.h +11 -0
  42. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  43. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  44. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  45. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  46. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  47. data/ext/sources/examples/server/server.cpp +199 -163
  48. data/ext/sources/examples/vad-speech-segments/speech.cpp +3 -2
  49. data/ext/sources/ggml/CMakeLists.txt +21 -14
  50. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  51. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  52. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  53. data/ext/sources/ggml/include/ggml-backend.h +72 -10
  54. data/ext/sources/ggml/include/ggml-cuda.h +2 -2
  55. data/ext/sources/ggml/include/ggml-rpc.h +3 -3
  56. data/ext/sources/ggml/include/ggml-sycl.h +8 -0
  57. data/ext/sources/ggml/include/ggml.h +103 -9
  58. data/ext/sources/ggml/include/gguf.h +10 -2
  59. data/ext/sources/ggml/src/CMakeLists.txt +30 -6
  60. data/ext/sources/ggml/src/ggml-alloc.c +5 -1
  61. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  62. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2266 -0
  63. data/ext/sources/ggml/src/ggml-backend-reg.cpp +12 -0
  64. data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
  65. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
  66. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
  67. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
  68. data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
  69. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
  70. data/ext/sources/ggml/src/ggml-common.h +24 -2
  71. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +59 -30
  72. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
  73. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +21 -22
  74. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -11
  75. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
  76. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  77. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  78. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +4279 -1292
  79. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
  80. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
  81. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  82. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +319 -31
  83. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
  84. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +12 -2
  85. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  86. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
  87. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +109 -5
  88. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
  89. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
  90. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +107 -82
  91. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +501 -119
  92. data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
  93. data/ext/sources/ggml/src/ggml-cpu/quants.c +106 -0
  94. data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
  95. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
  96. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +91 -1
  97. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +14 -16
  98. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  99. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  100. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  101. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  102. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  103. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  104. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  105. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  106. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  107. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  108. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  109. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  110. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  111. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  112. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  113. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +39 -55
  114. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  115. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
  116. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  117. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  118. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +62 -26
  119. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +134 -64
  120. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  121. data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
  122. data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
  123. data/ext/sources/ggml/src/ggml-cuda/common.cuh +246 -28
  124. data/ext/sources/ggml/src/ggml-cuda/concat.cu +134 -116
  125. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +14 -12
  126. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  127. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  128. data/ext/sources/ggml/src/ggml-cuda/convert.cu +139 -34
  129. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  130. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +88 -29
  131. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  132. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +287 -49
  133. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +335 -130
  134. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  135. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +127 -24
  136. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
  137. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
  138. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +169 -60
  139. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  140. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  141. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  142. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +109 -45
  143. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +10 -0
  144. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +48 -23
  145. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +2034 -2104
  146. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  147. data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
  148. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
  149. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
  150. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +25 -12
  151. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
  152. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
  153. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +562 -97
  154. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
  155. data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
  156. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +66 -7
  157. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
  158. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  159. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
  160. data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
  161. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  162. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +78 -10
  163. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  164. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  165. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  166. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +45 -13
  167. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  168. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  169. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  170. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  171. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +1 -0
  172. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  173. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +1 -0
  174. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  175. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  176. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +1 -0
  177. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  178. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  179. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +1 -0
  180. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  181. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +1 -0
  182. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  183. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  184. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +1 -0
  185. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  186. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  187. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  188. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  189. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  190. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  191. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  192. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  193. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  205. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -4
  206. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +33 -24
  207. data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
  208. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  209. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
  210. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  211. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +23 -4
  212. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -0
  213. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +1 -5
  214. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +2788 -1762
  215. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +13 -4
  216. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
  217. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
  218. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
  219. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +17 -19
  220. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  221. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +170 -127
  222. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  223. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  224. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  225. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +1774 -396
  226. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +303 -0
  227. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  228. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +148 -42
  229. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
  230. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
  231. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +255 -62
  232. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
  233. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +25 -21
  235. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1303 -0
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +167 -0
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +157 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +104 -13
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +222 -57
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +10 -3
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +78 -26
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +48 -8
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +232 -0
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  254. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  255. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +40 -0
  256. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  257. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
  258. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +625 -816
  259. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3052 -2166
  260. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.h +650 -0
  261. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  262. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  263. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +337 -106
  264. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
  265. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
  266. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  267. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +245 -151
  268. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
  269. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
  270. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +15 -3
  271. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.h +8 -0
  272. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +390 -0
  273. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -5
  274. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +27 -9
  275. data/ext/sources/ggml/src/ggml-impl.h +6 -1
  276. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +207 -18
  277. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +36 -2
  278. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +186 -29
  279. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +118 -0
  280. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +322 -21
  281. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +4 -0
  282. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
  283. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1226 -467
  284. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  285. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +67 -5
  286. data/ext/sources/ggml/src/ggml-opencl/fa_tune.h +92 -0
  287. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +16290 -6246
  288. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
  289. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
  290. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1997 -92
  291. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +81 -41
  292. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +88 -39
  293. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +1995 -96
  294. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +1615 -0
  295. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1486 -0
  296. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
  297. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +374 -0
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
  300. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
  301. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
  302. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
  303. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
  304. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
  305. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
  306. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  307. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
  308. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  309. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  310. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  311. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  312. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  313. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  314. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
  315. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  316. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  317. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  318. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +121 -0
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +1 -1
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1149 -0
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +48 -64
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +5 -2
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +500 -0
  358. data/ext/sources/ggml/src/ggml-opencl/libdl.h +79 -0
  359. data/ext/sources/ggml/src/ggml-openvino/.clang-format +0 -5
  360. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +2 -4
  361. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +740 -127
  362. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +76 -23
  363. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +75 -14
  364. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +29 -8
  365. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +339 -69
  366. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +330 -192
  367. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +10 -4
  368. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +56 -16
  369. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +1 -1
  370. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +4 -4
  371. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +94 -37
  372. data/ext/sources/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
  373. data/ext/sources/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
  374. data/ext/sources/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
  375. data/ext/sources/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
  376. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +8 -16
  377. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +14 -1
  378. data/ext/sources/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
  379. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +108 -21
  380. data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +282 -0
  381. data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
  382. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +2 -9
  383. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +21 -7
  384. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +41 -8
  385. data/ext/sources/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
  386. data/ext/sources/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +44 -0
  387. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +226 -0
  388. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +19 -9
  389. data/ext/sources/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
  390. data/ext/sources/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
  391. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +58 -13
  392. data/ext/sources/ggml/src/ggml-openvino/openvino/op/repeat.cpp +74 -0
  393. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +13 -6
  394. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +1 -1
  395. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +161 -39
  396. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +3 -3
  397. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +126 -49
  398. data/ext/sources/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +59 -0
  399. data/ext/sources/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
  400. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +32 -1
  401. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +1 -1
  402. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
  403. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +90 -25
  404. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +41 -22
  405. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +18 -4
  406. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +1 -1
  407. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +70 -43
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +5 -4
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +612 -36
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +29 -26
  412. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +460 -114
  413. data/ext/sources/ggml/src/ggml-openvino/utils.h +32 -9
  414. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  415. data/ext/sources/ggml/src/ggml-quants.c +365 -114
  416. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  417. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  418. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  419. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  420. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  421. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +50 -4
  422. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
  423. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +5 -1
  424. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +12 -0
  425. data/ext/sources/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
  426. data/ext/sources/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
  427. data/ext/sources/ggml/src/ggml-sycl/common.cpp +72 -2
  428. data/ext/sources/ggml/src/ggml-sycl/common.hpp +59 -2
  429. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +21 -1
  430. data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.cpp +158 -0
  431. data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
  432. data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
  433. data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
  434. data/ext/sources/ggml/src/ggml-sycl/conv2d.cpp +150 -0
  435. data/ext/sources/ggml/src/ggml-sycl/conv2d.hpp +10 -0
  436. data/ext/sources/ggml/src/ggml-sycl/conv3d.cpp +224 -0
  437. data/ext/sources/ggml/src/ggml-sycl/conv3d.hpp +8 -0
  438. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +121 -13
  439. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
  440. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +706 -0
  441. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +281 -0
  442. data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
  443. data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
  444. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  445. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  446. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +678 -0
  447. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  448. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  449. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +997 -244
  450. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +15 -7
  451. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +215 -204
  452. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
  453. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  454. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  455. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +7 -5
  456. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
  457. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
  458. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
  459. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
  460. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  461. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  462. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +69 -31
  463. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -0
  464. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  465. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  466. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1758 -455
  467. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  468. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  469. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1542 -39
  470. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +33 -0
  471. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +103 -49
  472. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +45 -9
  473. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  474. data/ext/sources/ggml/src/ggml-sycl/pool.cpp +185 -0
  475. data/ext/sources/ggml/src/ggml-sycl/pool.hpp +22 -0
  476. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -1
  477. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
  478. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +17 -3
  479. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +9 -10
  480. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  481. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  482. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  483. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  484. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  485. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  486. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  487. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  488. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
  489. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
  490. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
  491. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
  492. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
  493. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
  494. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
  495. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
  496. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
  497. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
  498. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
  499. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
  500. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
  501. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
  502. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
  503. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
  504. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
  505. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
  506. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
  507. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
  508. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
  509. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
  510. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
  511. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
  512. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
  513. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
  514. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
  515. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
  516. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
  517. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
  518. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
  519. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
  520. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
  521. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
  522. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
  523. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -0
  524. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  525. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  526. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  527. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +242 -45
  528. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
  529. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
  530. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
  531. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
  532. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
  533. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
  534. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +16 -0
  535. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2843 -700
  536. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  537. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
  538. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  539. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  540. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +431 -0
  541. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  542. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  543. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  544. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  545. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  546. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  547. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  548. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +3 -4
  549. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  550. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  551. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +198 -48
  552. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
  553. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +116 -113
  554. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
  555. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  556. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  557. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  558. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +125 -64
  559. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  560. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +21 -19
  561. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
  562. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +29 -1
  563. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +17 -11
  564. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  565. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  566. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +4 -7
  567. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  568. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  569. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  570. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  571. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  572. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  573. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +22 -24
  574. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  575. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +42 -40
  576. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +49 -15
  577. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +222 -171
  578. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  579. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  580. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  581. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +10 -10
  582. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +3 -3
  583. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +3 -3
  584. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +5 -2
  585. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  586. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
  587. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  588. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +11 -1
  589. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +3 -4
  590. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  591. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
  592. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +282 -211
  593. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  594. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2209 -283
  595. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2618 -1416
  596. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  597. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  598. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +8 -7
  599. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +90 -95
  600. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
  601. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  602. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -50
  603. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
  604. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  605. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  606. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  607. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  608. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  609. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  610. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +204 -78
  611. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  612. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  613. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +805 -526
  614. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  615. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  616. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  617. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +8 -6
  618. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
  619. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +90 -413
  620. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1553 -0
  621. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
  622. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  623. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
  624. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  625. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  626. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  627. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +6 -4
  628. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  629. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +2 -3
  630. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  631. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  632. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  633. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  634. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  635. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +68 -48
  636. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  637. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
  638. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
  639. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
  640. data/ext/sources/ggml/src/ggml.c +146 -42
  641. data/ext/sources/ggml/src/gguf.cpp +173 -28
  642. data/ext/sources/include/parakeet.h +342 -0
  643. data/ext/sources/include/whisper.h +31 -0
  644. data/ext/sources/media/matmul.png +0 -0
  645. data/ext/sources/src/CMakeLists.txt +23 -0
  646. data/ext/sources/src/parakeet-arch.h +188 -0
  647. data/ext/sources/src/parakeet.cpp +3838 -0
  648. data/ext/sources/src/whisper.cpp +220 -26
  649. data/extsources.rb +26 -10
  650. data/lib/whisper/log_settable.rb +33 -0
  651. data/lib/whisper/model/uri.rb +13 -8
  652. data/lib/whisper/output.rb +74 -0
  653. data/sig/whisper.rbs +417 -62
  654. data/test/helper.rb +2 -0
  655. data/test/jfk_reader/jfk_reader.c +50 -7
  656. data/test/test_callback.rb +1 -0
  657. data/test/test_package.rb +6 -5
  658. data/test/test_parakeet.rb +28 -0
  659. data/test/test_parakeet_callback.rb +107 -0
  660. data/test/test_parakeet_context.rb +116 -0
  661. data/test/test_parakeet_context_params.rb +24 -0
  662. data/test/test_parakeet_model.rb +21 -0
  663. data/test/test_parakeet_params.rb +78 -0
  664. data/test/test_parakeet_segment.rb +42 -0
  665. data/test/test_parakeet_token.rb +73 -0
  666. data/test/test_params.rb +2 -0
  667. data/test/test_vad.rb +9 -0
  668. data/test/test_vad_context.rb +2 -2
  669. data/test/test_vad_segment.rb +1 -1
  670. data/test/test_whisper.rb +24 -6
  671. data/whispercpp.gemspec +2 -2
  672. metadata +263 -304
  673. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  674. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  675. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  676. data/ext/sources/bindings/javascript/package.json +0 -26
  677. data/ext/sources/bindings/javascript/whisper.js +0 -19
  678. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  679. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  680. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  681. data/ext/sources/examples/addon.node/index.js +0 -59
  682. data/ext/sources/examples/addon.node/package.json +0 -16
  683. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  684. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  685. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  686. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  687. data/ext/sources/examples/coi-serviceworker.js +0 -146
  688. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  689. data/ext/sources/examples/command/command.cpp +0 -802
  690. data/ext/sources/examples/command/commands.txt +0 -9
  691. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  692. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  693. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  694. data/ext/sources/examples/generate-karaoke.sh +0 -57
  695. data/ext/sources/examples/helpers.js +0 -191
  696. data/ext/sources/examples/livestream.sh +0 -112
  697. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  698. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  699. data/ext/sources/examples/lsp/whisper.vim +0 -362
  700. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  701. data/ext/sources/examples/python/whisper_processor.py +0 -54
  702. data/ext/sources/examples/server/bench.js +0 -29
  703. data/ext/sources/examples/server.py +0 -120
  704. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  705. data/ext/sources/examples/stream/stream.cpp +0 -437
  706. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  707. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  708. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  709. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  710. data/ext/sources/examples/sycl/build.sh +0 -22
  711. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  712. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  713. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -48
  714. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  715. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
  716. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
  717. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
  718. data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
  719. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
  720. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  721. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
  722. data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
  723. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
  724. data/ext/sources/examples/talk-llama/llama-context.h +0 -359
  725. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  726. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
  727. data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
  728. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  729. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  730. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2735
  731. data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
  732. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
  733. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
  734. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  735. data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
  736. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  737. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  738. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -330
  739. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  740. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
  741. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
  742. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  743. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
  744. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
  745. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  746. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  747. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1165
  748. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  749. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  750. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  751. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -752
  752. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  753. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
  754. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
  755. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
  756. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
  757. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
  758. data/ext/sources/examples/talk-llama/llama-model.h +0 -597
  759. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
  760. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  761. data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
  762. data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
  763. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
  764. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
  765. data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
  766. data/ext/sources/examples/talk-llama/llama.h +0 -1573
  767. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
  768. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  769. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  770. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -137
  771. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  772. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
  773. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
  774. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
  775. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
  776. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
  777. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  778. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  779. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  780. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  781. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  782. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  783. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  784. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -122
  785. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  786. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
  787. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
  788. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
  789. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
  790. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  791. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
  792. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  793. data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
  794. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
  795. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  796. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  797. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -111
  798. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  799. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  800. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  801. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  802. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  803. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  804. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  805. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -157
  806. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  807. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  808. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -195
  809. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
  810. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  811. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
  812. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  813. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
  814. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  815. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  816. data/ext/sources/examples/talk-llama/models/jais2.cpp +0 -123
  817. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  818. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
  819. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
  820. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  821. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  822. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  823. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -175
  824. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  825. data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
  826. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
  827. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
  828. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
  829. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
  830. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  831. data/ext/sources/examples/talk-llama/models/models.h +0 -704
  832. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
  833. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  834. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
  835. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  836. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  837. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  838. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  839. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  840. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  841. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  842. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  843. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +0 -122
  844. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  845. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  846. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  847. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  848. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -320
  849. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  850. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
  851. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  852. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  853. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  854. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  855. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -120
  856. data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
  857. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
  858. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
  859. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
  860. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
  861. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
  862. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  863. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  864. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -164
  865. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  866. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  867. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -137
  868. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  869. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  870. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  871. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  872. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  873. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  874. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  875. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +0 -165
  876. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  877. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  878. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  879. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  880. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  881. data/ext/sources/examples/talk-llama/speak +0 -40
  882. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  883. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  884. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  885. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  886. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  887. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1103
  888. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  889. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  890. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  891. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  892. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  893. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  894. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  895. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  896. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  897. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  898. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  899. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  900. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  901. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  902. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
  903. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  904. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  905. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +0 -123
  906. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
  907. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
  908. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp +0 -21
  909. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp +0 -22
  910. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +0 -17
  911. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +0 -17
  912. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +0 -27
  913. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -21
  914. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp +0 -22
  915. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +0 -25
  916. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +0 -39
  917. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +0 -23
  918. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +0 -22
  919. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +0 -22
  920. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +0 -22
  921. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp +0 -20
  922. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +0 -21
  923. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/round.comp +0 -29
  924. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  925. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +0 -21
  926. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +0 -20
  927. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +0 -22
  928. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +0 -17
  929. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp +0 -23
  930. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +0 -17
  931. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +0 -17
  932. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/step.comp +0 -22
  933. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +0 -20
  934. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp +0 -22
  935. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/xielu.comp +0 -35
  936. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
  937. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  938. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -718
  939. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  940. data/ext/sources/tests/CMakeLists.txt +0 -112
  941. data/ext/sources/tests/earnings21/eval.mk +0 -58
  942. data/ext/sources/tests/earnings21/eval.py +0 -68
  943. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  944. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  945. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  946. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  947. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  948. data/ext/sources/tests/en-0-ref.txt +0 -1
  949. data/ext/sources/tests/en-1-ref.txt +0 -1
  950. data/ext/sources/tests/en-2-ref.txt +0 -1
  951. data/ext/sources/tests/es-0-ref.txt +0 -1
  952. data/ext/sources/tests/librispeech/eval.mk +0 -39
  953. data/ext/sources/tests/librispeech/eval.py +0 -47
  954. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  955. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  956. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  957. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  958. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  959. data/ext/sources/tests/run-tests.sh +0 -130
  960. data/ext/sources/tests/test-c.c +0 -3
  961. data/ext/sources/tests/test-vad-full.cpp +0 -56
  962. data/ext/sources/tests/test-vad.cpp +0 -83
  963. data/ext/sources/tests/test-whisper.js +0 -58
  964. data/lib/whisper/context.rb +0 -15
  965. data/lib/whisper/segment.rb +0 -58
@@ -11,6 +11,7 @@
11
11
  #include <future>
12
12
  #include <queue>
13
13
  #include <condition_variable>
14
+ #include <atomic>
14
15
  #include <cstdio>
15
16
  #include <cstring>
16
17
  #include <cstdlib>
@@ -34,6 +35,9 @@
34
35
 
35
36
  std::mutex lock;
36
37
  std::vector<std::pair<std::string, std::string>> shader_fnames;
38
+ // Set when any shader subprocess fails (non-zero exit / stderr / launch failure) so the
39
+ // build is stopped instead of silently producing a broken libggml-vulkan. (issue #24393)
40
+ static std::atomic<bool> compile_failed{false};
37
41
  std::locale c_locale("C");
38
42
 
39
43
  std::string GLSLC = "glslc";
@@ -45,6 +49,7 @@ std::string target_cpp = "";
45
49
  const std::vector<std::string> type_names = {
46
50
  "f32",
47
51
  "f16",
52
+ "q1_0",
48
53
  "q4_0",
49
54
  "q4_1",
50
55
  "q5_0",
@@ -65,6 +70,7 @@ const std::vector<std::string> type_names = {
65
70
  "iq4_xs",
66
71
  "iq4_nl",
67
72
  "mxfp4",
73
+ "nvfp4",
68
74
  "bf16",
69
75
  };
70
76
 
@@ -76,7 +82,7 @@ enum MatMulIdType {
76
82
 
77
83
  namespace {
78
84
 
79
- void execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
85
+ int execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
80
86
  #ifdef _WIN32
81
87
  HANDLE stdout_read, stdout_write;
82
88
  HANDLE stderr_read, stderr_write;
@@ -125,8 +131,11 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
125
131
  CloseHandle(stdout_read);
126
132
  CloseHandle(stderr_read);
127
133
  WaitForSingleObject(pi.hProcess, INFINITE);
134
+ DWORD exit_code = 1;
135
+ GetExitCodeProcess(pi.hProcess, &exit_code);
128
136
  CloseHandle(pi.hProcess);
129
137
  CloseHandle(pi.hThread);
138
+ return (int)exit_code;
130
139
  #else
131
140
  int stdout_pipe[2];
132
141
  int stderr_pipe[2];
@@ -137,6 +146,7 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
137
146
 
138
147
  pid_t pid = fork();
139
148
  if (pid < 0) {
149
+ std::cerr << strerror(errno) << "\n";
140
150
  throw std::runtime_error("Failed to fork process");
141
151
  }
142
152
 
@@ -172,7 +182,9 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
172
182
 
173
183
  close(stdout_pipe[0]);
174
184
  close(stderr_pipe[0]);
175
- waitpid(pid, nullptr, 0);
185
+ int status = 0;
186
+ waitpid(pid, &status, 0);
187
+ return WIFEXITED(status) ? WEXITSTATUS(status) : -1;
176
188
  }
177
189
  #endif
178
190
  }
@@ -333,7 +345,8 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
333
345
  // disable spirv-opt for coopmat shaders for https://github.com/ggml-org/llama.cpp/issues/10734
334
346
  // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
335
347
  // 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) {
348
+ // disable spirv-opt for dot2 shaders (spirv-opt doesn't recognize SPV_VALVE_mixed_float_dot_product capability)
349
+ if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos && name.find("_dot2") == std::string::npos) {
337
350
  cmd.push_back("-O");
338
351
  }
339
352
 
@@ -368,13 +381,14 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
368
381
  // }
369
382
  // std::cout << std::endl;
370
383
 
371
- execute_command(cmd, stdout_str, stderr_str);
372
- if (!stderr_str.empty()) {
373
- std::cerr << "cannot compile " << name << "\n\n";
384
+ int exit_code = execute_command(cmd, stdout_str, stderr_str);
385
+ if (exit_code != 0 || !stderr_str.empty()) {
386
+ std::cerr << "cannot compile " << name << " (exit code " << exit_code << ")\n\n";
374
387
  for (const auto& part : cmd) {
375
388
  std::cerr << part << " ";
376
389
  }
377
390
  std::cerr << "\n\n" << stderr_str << std::endl;
391
+ compile_failed = true;
378
392
  return;
379
393
  }
380
394
 
@@ -394,6 +408,7 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
394
408
  shader_fnames.push_back(std::make_pair(name, out_path));
395
409
  } catch (const std::exception& e) {
396
410
  std::cerr << "Error executing command for " << name << ": " << e.what() << std::endl;
411
+ compile_failed = true;
397
412
  }
398
413
  }
399
414
 
@@ -403,9 +418,9 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
403
418
  return result;
404
419
  }
405
420
 
406
- 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"));
421
+ static std::deque<std::future<void>> compiles;
422
+ 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 = "") {
423
+ name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")) + suffix;
409
424
  std::string out_path = join_paths(output_dir, name + ".spv");
410
425
 
411
426
  if (input_filepath == "") {
@@ -422,12 +437,18 @@ void string_to_spv(std::string name, const std::string& source, const std::map<s
422
437
  string_to_spv_func, name, input_filepath, out_path, defines, coopmat, generate_dep_file, std::move(slot)));
423
438
  // Don't write the same dep file from multiple processes
424
439
  generate_dep_file = false;
440
+
441
+ // Clean up completed futures - don't accumulate virtual memory for completed threads' stacks.
442
+ while (!compiles.empty() && compiles.front().wait_for(std::chrono::seconds(0)) == std::future_status::ready) {
443
+ compiles.pop_front();
444
+ }
425
445
  }
426
446
 
427
- void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc) {
447
+ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc, bool dot2 = false) {
428
448
  std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
429
449
  std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
430
450
  std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
451
+ std::string dot2_sfx = dot2 ? "_dot2" : "";
431
452
 
432
453
  std::map<std::string, std::string> base_dict;
433
454
  std::string shader_name = "matmul";
@@ -445,8 +466,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
445
466
  base_dict["FLOAT16"] = "1";
446
467
  }
447
468
 
448
- base_dict["ACC_TYPE" ] = f16acc ? "float16_t" : "float";
449
- base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2" : "vec2";
469
+ base_dict["ACC_TYPE" ] = f16acc ? "float16_t" : "float";
470
+ base_dict["ACC_TYPEV2"] = f16acc ? "f16vec2" : "vec2";
450
471
  if (f16acc) {
451
472
  base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
452
473
  }
@@ -454,6 +475,15 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
454
475
  if (coopmat) {
455
476
  base_dict["COOPMAT"] = "1";
456
477
  }
478
+ #if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
479
+ if (coopmat2) {
480
+ base_dict["GGML_VULKAN_COOPMAT2_DECODE_VECTOR"] = "1";
481
+ }
482
+ #endif
483
+
484
+ if (dot2) {
485
+ base_dict["DOT2_F16"] = "1";
486
+ }
457
487
 
458
488
  const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
459
489
 
@@ -513,18 +543,16 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
513
543
  };
514
544
 
515
545
  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")},
546
+ {"FLOAT_TYPE", FLOAT_TYPE(1, "f16")},
547
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, "f16")},
548
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, "f16")},
549
+ {"FLOAT_TYPEV8", FLOAT_TYPE(8, "f16")},
520
550
  };
521
551
 
522
552
  // 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);
553
+ 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"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
525
554
 
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);
555
+ string_to_spv(shader_name + "_f16" + dot2_sfx, 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_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
528
556
 
529
557
  // bf16
530
558
  {
@@ -535,9 +563,9 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
535
563
  std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
536
564
 
537
565
  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")},
566
+ {"FLOAT_TYPE", FLOAT_TYPE(1, "bf16")},
567
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, "bf16")},
568
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, "bf16")},
541
569
  };
542
570
 
543
571
  // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
@@ -545,16 +573,17 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
545
573
  if (!(coopmat || coopmat2))
546
574
  #endif
547
575
  {
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);
576
+ if (!dot2) {
577
+ 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"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"B_TYPE_SCALAR", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}), fp16, coopmat, coopmat2, f16acc);
578
+ }
550
579
  }
551
580
  }
552
581
 
553
582
  for (const auto& tname : type_names) {
554
583
  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"))
584
+ 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
585
  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"))
586
+ 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
587
  load_vec_quant = "4";
559
588
 
560
589
  if (tname == "bf16") {
@@ -562,32 +591,28 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
562
591
  }
563
592
 
564
593
  std::string data_a_key = "DATA_A_" + to_uppercase(tname);
565
- // For unaligned, load one at a time for f32/f16, or two at a time for quants
566
- std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
567
594
  // For aligned matmul loads
568
595
  std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
569
596
 
570
597
  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)},
598
+ {"FLOAT_TYPE", FLOAT_TYPE(1, tname)},
599
+ {"FLOAT_TYPEV2", FLOAT_TYPE(2, tname)},
600
+ {"FLOAT_TYPEV4", FLOAT_TYPE(4, tname)},
601
+ {"FLOAT_TYPEV8", FLOAT_TYPE(8, tname)},
575
602
  };
576
603
 
577
604
  // don't generate f32 variants for coopmat2
578
605
  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);
606
+ 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}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPE_SCALAR", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
581
607
  }
582
608
 
583
609
  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);
610
+ 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}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
586
611
  }
587
612
 
588
613
  #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")) {
614
+ // Integer dot mmq performs better with f32 accumulators (different shader, skip for dot2)
615
+ if (!f16acc && !coopmat && !coopmat2 && !dot2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
591
616
  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
617
  }
593
618
  #endif
@@ -605,6 +630,10 @@ void process_shaders() {
605
630
  matmul_shaders(true, matmul_id_type, false, false, false);
606
631
  matmul_shaders(true, matmul_id_type, false, false, true);
607
632
 
633
+ // dot2 variants (scalar fp16 only)
634
+ matmul_shaders(true, matmul_id_type, false, false, false, true);
635
+ matmul_shaders(true, matmul_id_type, false, false, true, true);
636
+
608
637
  if (matmul_id_type != MatMulIdType::DEFAULT) {
609
638
  #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
610
639
  // Coopmat, fp32acc and fp16acc
@@ -623,88 +652,100 @@ void process_shaders() {
623
652
  for (const bool& fp16 : {false, true}) {
624
653
  std::map<std::string, std::string> base_dict;
625
654
  if (fp16) {
626
- base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
655
+ base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV2", "f16vec2"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
627
656
  } else {
628
- base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV4", "vec4"}};
657
+ base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}, {"FLOAT_TYPEV4", "vec4"}};
629
658
  }
630
659
 
631
660
  // flash attention
632
661
  for (const bool& f16acc : {false, true}) {
633
662
  std::map<std::string, std::string> fa_base_dict = base_dict;
634
663
  fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
664
+ fa_base_dict["ACC_TYPEV2"] = fp16 && f16acc ? "f16vec2" : "vec2";
635
665
  fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
636
666
  if (fp16 && f16acc) {
637
667
  fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
638
668
  }
639
669
 
640
- for (const auto& tname : type_names) {
641
- if (tname == "bf16") continue;
642
-
643
- if (fp16) {
670
+ if (fp16) {
644
671
  #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
645
- if (tname == "f16") {
646
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
647
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
648
- } else {
649
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
650
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
651
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, true, f16acc);
652
- }
672
+ string_to_spv("flash_attn_f32_f16", "flash_attn_cm2.comp",
673
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
653
674
  #endif
675
+
654
676
  #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
655
- if (tname == "f16") {
656
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
657
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
658
- } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
659
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
660
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
661
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
662
- }
677
+ string_to_spv("flash_attn_f32_f16", "flash_attn_cm1.comp",
678
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
663
679
  #endif
664
- }
680
+ }
665
681
 
666
- if (tname == "f16") {
667
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
668
- merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
669
- } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
670
- std::string data_a_key = "DATA_A_" + to_uppercase(tname);
671
- string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
672
- merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, false, f16acc);
673
- }
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"}}), fp16, false, false, f16acc);
684
+
685
+ if (fp16) {
686
+ string_to_spv("flash_attn_f32_f16_dot2", "flash_attn.comp",
687
+ merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DOT2_F16", "1"}}), fp16, false, false, f16acc);
674
688
  }
689
+
690
+ #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
691
+ string_to_spv("flash_attn_f32_f16", "flash_attn.comp",
692
+ 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");
693
+ #endif
675
694
  }
676
695
  }
677
696
 
678
- std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
697
+ const std::map<std::string, std::string> fa_bf16_dict = {
698
+ {"FLOAT_TYPE", "bfloat16_t"},
699
+ {"FLOAT_TYPEV2", "bf16vec2"},
700
+ {"FLOAT_TYPEV4", "bf16vec4"},
701
+ {"ACC_TYPE", "float"},
702
+ {"ACC_TYPEV2", "vec2"},
703
+ {"ACC_TYPEV4", "vec4"},
704
+ {"BFLOAT16", "1"},
705
+ };
706
+
707
+ #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
708
+ string_to_spv("flash_attn_f32_f16_bf16", "flash_attn_cm1.comp",
709
+ merge_maps(fa_bf16_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}),
710
+ true, true, false, false);
711
+ #endif
712
+
713
+ #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
714
+ string_to_spv("flash_attn_f32_f16_bf16", "flash_attn_cm2.comp",
715
+ merge_maps(fa_bf16_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}),
716
+ true, false, true, false);
717
+ #endif
718
+
719
+ std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV2", "vec2"}};
679
720
 
680
721
  for (const auto& tname : type_names) {
681
722
  // mul mat vec
682
723
  std::string data_a_key = "DATA_A_" + to_uppercase(tname);
683
724
  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";
684
725
 
685
- 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"}}));
686
- 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"}}));
726
+ 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"}}));
727
+ 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"}}));
687
728
 
688
- 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"}}));
689
- 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"}}));
729
+ 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"}}));
730
+ 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"}}));
690
731
 
691
- 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"}}));
692
- 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"}}));
732
+ 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"}}));
733
+ 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"}}));
693
734
 
694
- 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"}}));
695
- 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"}}));
696
- 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"}}));
735
+ 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"}}));
736
+ 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"}}));
737
+ 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"}}));
697
738
 
698
739
  // mul mat vec with integer dot product
699
740
  #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
700
741
  if (is_legacy_quant(tname) || tname == "mxfp4" || is_k_quant(tname) || tname == "iq1_s" || tname == "iq1_m") {
701
- 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"}}));
702
- 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"}}));
703
- 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"}}));
742
+ 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"}}));
743
+ 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"}}));
744
+ 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"}}));
704
745
 
705
- 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"}}));
706
- 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"}}));
707
- 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"}}));
746
+ 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"}}));
747
+ 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"}}));
748
+ 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"}}));
708
749
  }
709
750
  #endif
710
751
 
@@ -725,9 +766,9 @@ void process_shaders() {
725
766
 
726
767
  string_to_spv("get_rows_i32", "get_rows.comp", {{"TEMP_TYPE", "uint"}, {"A_TYPE", "uint"}, {"B_TYPE", "int"}, {"D_TYPE", "uint"}});
727
768
 
728
- 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"}});
729
- 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"}});
730
- 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"}});
769
+ 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"}});
770
+ 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"}});
771
+ 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"}});
731
772
 
732
773
  // Norms
733
774
  string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -735,7 +776,7 @@ void process_shaders() {
735
776
  string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
736
777
  string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
737
778
  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"}}));
738
- 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"}}));
779
+ 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"}}));
739
780
  string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
740
781
  string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
741
782
 
@@ -744,6 +785,7 @@ void process_shaders() {
744
785
  string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
745
786
  string_to_spv("cpy_f16_f32", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
746
787
  string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
788
+ string_to_spv("cpy_bf16_f32","copy.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
747
789
  string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
748
790
  string_to_spv("contig_cpy_f32_i32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
749
791
  string_to_spv("contig_cpy_i32_f32", "contig_copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
@@ -751,23 +793,21 @@ void process_shaders() {
751
793
  string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
752
794
  string_to_spv("contig_cpy_f16_f32", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
753
795
  string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
796
+ string_to_spv("contig_cpy_bf16_f32","contig_copy.comp",{{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
754
797
  string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
755
798
  string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
756
799
 
757
800
  string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
758
801
  string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
759
802
 
760
- for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
803
+ for (std::string t : {"q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
761
804
  string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
762
- 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"}});
763
805
  string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
764
806
  }
765
807
 
766
- for (std::string t : {"f32", "f16", "bf16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
767
- 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"}});
768
- 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"}});
769
- 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"}});
770
- 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"}});
808
+ for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
809
+ 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"}});
810
+ 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"}});
771
811
  }
772
812
 
773
813
  auto get_type_str = [](bool f16) {
@@ -784,12 +824,10 @@ void process_shaders() {
784
824
  for (auto src0_f16 : {false, true}) {
785
825
  for (auto src1_f16 : {false, true}) {
786
826
  for (auto dst_f16 : {false, true}) {
787
- for (auto rte : {false, true}) {
788
827
  auto source = op == "add_rms" ? std::string("add") : op;
789
- auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
828
+ auto name = op + get_suffix(src0_f16, src1_f16, dst_f16);
790
829
  auto add_rms = op == "add_rms" ? "1" : "0";
791
- 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}});
792
- }
830
+ 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}});
793
831
  }
794
832
  }
795
833
  }
@@ -814,107 +852,109 @@ void process_shaders() {
814
852
 
815
853
  string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
816
854
 
817
- string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
855
+ string_to_spv("repeat_i32", "repeat.comp", {{"A_TYPE", "int32_t"}, {"D_TYPE", "int32_t"}});
818
856
  string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
857
+ string_to_spv("get_rows_back_f32", "get_rows_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
819
858
 
820
- string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
821
-
822
- string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
823
-
824
- string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
825
-
826
- string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
859
+ string_to_spv("repeat_i16", "repeat.comp", {{"A_TYPE", "int16_t"}, {"D_TYPE", "int16_t"}});
827
860
 
828
- string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
829
-
830
- string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
861
+ string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
831
862
 
832
863
  string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
833
864
 
834
- string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
835
- string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
836
- string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
865
+ string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
866
+ string_to_spv("concat_i16", "concat.comp", {{"A_TYPE", "uint16_t"}, {"B_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
867
+ string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "uint"}, {"B_TYPE", "uint"}, {"D_TYPE", "uint"}});
868
+ string_to_spv("concat_i64", "concat.comp", {{"A_TYPE", "uvec2"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "uvec2"}});
837
869
 
838
870
  string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
839
871
 
840
- for (auto rte : {false, true}) {
841
- std::string suffix = rte ? "_rte" : "";
842
- string_to_spv("exp_f16" + suffix, "exp.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
843
- string_to_spv("exp_f32" + suffix, "exp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"} , {"RTE16", rte ? "1" : "0"}});
844
-
845
- string_to_spv("log_f16" + suffix, "log.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
846
- string_to_spv("log_f32" + suffix, "log.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
847
- }
848
- string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
849
- string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
850
- string_to_spv("gelu_erf_f16", "gelu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
851
- string_to_spv("gelu_erf_f32", "gelu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
852
- string_to_spv("gelu_quick_f16", "gelu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
853
- string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
854
- string_to_spv("silu_f16", "silu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
855
- string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
856
- string_to_spv("relu_f16", "relu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
857
- string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
858
- string_to_spv("neg_f16", "neg.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
859
- string_to_spv("neg_f32", "neg.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
860
- string_to_spv("tanh_f16", "tanh.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
861
- string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
862
- string_to_spv("sigmoid_f16", "sigmoid.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
863
- string_to_spv("sigmoid_f32", "sigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
864
- string_to_spv("hardsigmoid_f16","hardsigmoid.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
865
- string_to_spv("hardsigmoid_f32","hardsigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
866
- string_to_spv("hardswish_f16", "hardswish.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
867
- string_to_spv("hardswish_f32", "hardswish.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
868
- string_to_spv("abs_f16", "abs.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
869
- string_to_spv("abs_f32", "abs.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
870
- string_to_spv("elu_f16", "elu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
871
- string_to_spv("elu_f32", "elu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
872
- string_to_spv("xielu_f16", "xielu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
873
- string_to_spv("xielu_f32", "xielu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
874
- string_to_spv("sgn_f16", "sgn.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
875
- string_to_spv("sgn_f32", "sgn.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
872
+ string_to_spv("exp_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_exp"}});
873
+ string_to_spv("exp_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_exp"}});
874
+ string_to_spv("expm1_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_expm1"}});
875
+ string_to_spv("expm1_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_expm1"}});
876
+
877
+ string_to_spv("log_f16", "log.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
878
+ string_to_spv("log_f32", "log.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
879
+ string_to_spv("gelu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_gelu"}});
880
+ string_to_spv("gelu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_gelu"}});
881
+ string_to_spv("gelu_erf_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_gelu_erf"}});
882
+ string_to_spv("gelu_erf_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_gelu_erf"}});
883
+ string_to_spv("gelu_quick_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_gelu_quick"}});
884
+ string_to_spv("gelu_quick_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_gelu_quick"}});
885
+ string_to_spv("silu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_silu"}});
886
+ string_to_spv("silu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_silu"}});
887
+ string_to_spv("relu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_relu"}});
888
+ string_to_spv("relu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_relu"}});
889
+ string_to_spv("sqr_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_sqr"}});
890
+ string_to_spv("sqr_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_sqr"}});
891
+ string_to_spv("sqrt_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_sqrt"}});
892
+ string_to_spv("sqrt_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_sqrt"}});
893
+ string_to_spv("sin_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_sin"}});
894
+ string_to_spv("sin_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_sin"}});
895
+ string_to_spv("cos_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_cos"}});
896
+ string_to_spv("cos_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_cos"}});
897
+ string_to_spv("clamp_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_clamp"}});
898
+ string_to_spv("clamp_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_clamp"}});
899
+ string_to_spv("leaky_relu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_leaky_relu"}});
900
+ string_to_spv("leaky_relu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_leaky_relu"}});
901
+ string_to_spv("neg_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_neg"}});
902
+ string_to_spv("neg_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_neg"}});
903
+ string_to_spv("tanh_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_tanh"}});
904
+ string_to_spv("tanh_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_tanh"}});
905
+ string_to_spv("sigmoid_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_sigmoid"}});
906
+ string_to_spv("sigmoid_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_sigmoid"}});
907
+ string_to_spv("hardsigmoid_f16","unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_hardsigmoid"}});
908
+ string_to_spv("hardsigmoid_f32","unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_hardsigmoid"}});
909
+ string_to_spv("hardswish_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_hardswish"}});
910
+ string_to_spv("hardswish_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_hardswish"}});
911
+ string_to_spv("abs_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_abs"}});
912
+ string_to_spv("abs_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_abs"}});
913
+ string_to_spv("elu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_elu"}});
914
+ string_to_spv("elu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_elu"}});
915
+ string_to_spv("xielu_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_xielu"}});
916
+ string_to_spv("xielu_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_xielu"}});
917
+ string_to_spv("sgn_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_sgn"}});
918
+ string_to_spv("sgn_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_sgn"}});
876
919
 
877
920
  string_to_spv("tri_f16", "tri.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
878
921
  string_to_spv("tri_f32", "tri.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
879
922
  string_to_spv("diag_f16", "diag.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
880
923
  string_to_spv("diag_f32", "diag.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
881
924
 
882
- string_to_spv("softplus_f16", "softplus.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
883
- string_to_spv("softplus_f32", "softplus.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
925
+ string_to_spv("softplus_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_softplus"}});
926
+ string_to_spv("softplus_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_softplus"}});
884
927
 
885
928
  string_to_spv("add1_f16_f16", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
886
929
  string_to_spv("add1_f16_f32", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
887
930
  string_to_spv("add1_f32_f32", "add1.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
888
931
  string_to_spv("arange_f32", "arange.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
889
932
  string_to_spv("fill_f32", "fill.comp", {{"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
890
- string_to_spv("step_f16", "step.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
891
- string_to_spv("step_f32", "step.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
892
- string_to_spv("round_f16", "round.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
893
- string_to_spv("round_f32", "round.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
894
- string_to_spv("ceil_f16", "ceil.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
895
- string_to_spv("ceil_f32", "ceil.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
896
- string_to_spv("floor_f16", "floor.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
897
- string_to_spv("floor_f32", "floor.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
898
- string_to_spv("trunc_f16", "trunc.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
899
- string_to_spv("trunc_f32", "trunc.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
900
-
901
- for (auto rte : {false, true}) {
902
- std::string suffix = rte ? "_rte" : "";
903
- string_to_spv("geglu_f16" + suffix, "geglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
904
- string_to_spv("geglu_f32" + suffix, "geglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
905
- string_to_spv("reglu_f16" + suffix, "reglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
906
- string_to_spv("reglu_f32" + suffix, "reglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
907
- string_to_spv("swiglu_f16" + suffix, "swiglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
908
- string_to_spv("swiglu_f32" + suffix, "swiglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
909
- string_to_spv("swiglu_oai_f16" + suffix, "swiglu_oai.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
910
- string_to_spv("swiglu_oai_f32" + suffix, "swiglu_oai.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
911
- string_to_spv("geglu_erf_f16" + suffix, "geglu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
912
- string_to_spv("geglu_erf_f32" + suffix, "geglu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
913
- string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
914
- string_to_spv("geglu_quick_f32" + suffix,"geglu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
915
- }
933
+ string_to_spv("fill_f16", "fill.comp", {{"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
934
+ string_to_spv("step_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_step"}});
935
+ string_to_spv("step_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_step"}});
936
+ string_to_spv("round_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_round"}});
937
+ string_to_spv("round_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_round"}});
938
+ string_to_spv("ceil_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_ceil"}});
939
+ string_to_spv("ceil_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_ceil"}});
940
+ string_to_spv("floor_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_floor"}});
941
+ string_to_spv("floor_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_floor"}});
942
+ string_to_spv("trunc_f16", "unary.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OP", "op_trunc"}});
943
+ string_to_spv("trunc_f32", "unary.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"OP", "op_trunc"}});
944
+
945
+ string_to_spv("geglu_f16", "geglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
946
+ string_to_spv("geglu_f32", "geglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
947
+ string_to_spv("reglu_f16", "reglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
948
+ string_to_spv("reglu_f32", "reglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
949
+ string_to_spv("swiglu_f16", "swiglu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
950
+ string_to_spv("swiglu_f32", "swiglu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
951
+ string_to_spv("swiglu_oai_f16", "swiglu_oai.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
952
+ string_to_spv("swiglu_oai_f32", "swiglu_oai.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
953
+ string_to_spv("geglu_erf_f16", "geglu_erf.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
954
+ string_to_spv("geglu_erf_f32", "geglu_erf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
955
+ string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
956
+ string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
916
957
 
917
- string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
918
958
  string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
919
959
 
920
960
  string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -932,25 +972,18 @@ void process_shaders() {
932
972
 
933
973
  string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
934
974
  string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
935
- string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
936
975
  string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
937
- string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
938
976
 
939
977
  string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
940
978
  string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
941
- string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
942
979
  string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
943
- string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
944
980
 
945
981
  string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
946
982
  string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
947
- string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
948
983
  string_to_spv("rope_multi_f32_f16", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
949
- string_to_spv("rope_multi_f32_f16_rte", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
950
984
 
951
985
  string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
952
986
  string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
953
- string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
954
987
 
955
988
  string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
956
989
  string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
@@ -960,6 +993,8 @@ void process_shaders() {
960
993
 
961
994
  string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
962
995
  string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
996
+ string_to_spv("fwht_f32", "fwht.comp", {});
997
+ string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
963
998
  string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
964
999
  string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
965
1000
  string_to_spv("cumsum_multipass1_f32", "cumsum_multipass1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -973,13 +1008,19 @@ void process_shaders() {
973
1008
  std::string bda_def = bda ? "1" : "0";
974
1009
  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}}));
975
1010
  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}}));
976
- 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}}));
977
1011
  }
978
1012
  }
979
1013
 
980
1014
  string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
981
1015
 
982
1016
  string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
1017
+ string_to_spv("col2im_1d_f32", "col2im_1d.comp", {{"DATA_A_F32", "1"}, {"A_TYPE", "float"}, {"D_TYPE", "float"}});
1018
+ string_to_spv("col2im_1d_f16", "col2im_1d.comp", {{"DATA_A_F16", "1"}, {"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
1019
+ string_to_spv("col2im_1d_bf16", "col2im_1d.comp", {{"DATA_A_BF16", "1"}, {"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
1020
+
1021
+ string_to_spv("snake_f32", "snake.comp", {{"DATA_A_F32", "1"}, {"A_TYPE", "float"}, {"D_TYPE", "float"}});
1022
+ string_to_spv("snake_f16", "snake.comp", {{"DATA_A_F16", "1"}, {"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
1023
+ string_to_spv("snake_bf16", "snake.comp", {{"DATA_A_BF16", "1"}, {"DATA_D_BF16", "1"}, {"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
983
1024
 
984
1025
  string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
985
1026
 
@@ -987,7 +1028,9 @@ void process_shaders() {
987
1028
 
988
1029
  string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
989
1030
 
990
- string_to_spv("gated_delta_net_f32", "gated_delta_net.comp", merge_maps(base_dict, {{"FLOAT_TYPE", "float"}}));
1031
+ 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"}}));
1032
+ 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"}}));
1033
+ 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"}}));
991
1034
 
992
1035
  string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
993
1036
  string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
@@ -1007,14 +1050,47 @@ void process_shaders() {
1007
1050
  string_to_spv(name + (unroll ? "_unroll" : ""), "conv2d_mm.comp", defines);
1008
1051
  #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
1009
1052
  if (unroll) {
1010
- defines["COOPMAT2"] = "1";
1011
- string_to_spv(name, "conv2d_mm.comp", defines, true, false, true);
1053
+ auto cm2_defines = defines;
1054
+ cm2_defines["COOPMAT2"] = "1";
1055
+ string_to_spv(name, "conv2d_mm.comp", cm2_defines, true, false, true);
1056
+ }
1057
+ #endif
1058
+ #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
1059
+ if (unroll) {
1060
+ auto cm1_defines = defines;
1061
+ cm1_defines["COOPMAT"] = "1";
1062
+ string_to_spv(name, "conv2d_mm.comp", cm1_defines, true, true, false);
1012
1063
  }
1013
1064
  #endif
1014
1065
  }
1015
1066
  }
1016
1067
  }
1017
1068
 
1069
+ for (auto unroll : {false, true}) {
1070
+ for (auto a_f16 : {false, true}) {
1071
+ std::map<std::string, std::string> defines = {
1072
+ {"A_TYPE", a_f16 ? "float16_t" : "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"},
1073
+ {"UNROLL", unroll ? "[[unroll]]" : ""},
1074
+ };
1075
+ std::string name = std::string("conv3d") + (a_f16 ? "_f16" : "") + "_f32";
1076
+ string_to_spv(name + (unroll ? "_unroll" : ""), "conv3d_mm.comp", defines);
1077
+ #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
1078
+ if (unroll) {
1079
+ auto cm2_defines = defines;
1080
+ cm2_defines["COOPMAT2"] = "1";
1081
+ string_to_spv(name, "conv3d_mm.comp", cm2_defines, true, false, true);
1082
+ }
1083
+ #endif
1084
+ #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
1085
+ if (unroll) {
1086
+ auto cm1_defines = defines;
1087
+ cm1_defines["COOPMAT"] = "1";
1088
+ string_to_spv(name, "conv3d_mm.comp", cm1_defines, true, true, false);
1089
+ }
1090
+ #endif
1091
+ }
1092
+ }
1093
+
1018
1094
  string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
1019
1095
  string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
1020
1096
  string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
@@ -1024,8 +1100,8 @@ void process_shaders() {
1024
1100
 
1025
1101
  string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
1026
1102
 
1027
- 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"}});
1028
- 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"}});
1103
+ string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"ADD_RMS" , "0"}});
1104
+ 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"}});
1029
1105
 
1030
1106
  string_to_spv("ssm_scan_f32", "ssm_scan.comp", {{"A_TYPE", "float"}});
1031
1107
  string_to_spv("ssm_scan_subgroup_f32", "ssm_scan.comp", {{"A_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
@@ -1078,8 +1154,8 @@ void write_output_files() {
1078
1154
 
1079
1155
  std::string suffixes[2] = {"_f32", "_f16"};
1080
1156
  for (std::string op : {"add", "sub", "mul", "div", "add_rms"}) {
1081
- hdr << "extern const void * " << op << "_data[2][2][2][2];\n";
1082
- hdr << "extern const uint64_t " << op << "_len[2][2][2][2];\n";
1157
+ hdr << "extern const void * " << op << "_data[2][2][2];\n";
1158
+ hdr << "extern const uint64_t " << op << "_len[2][2][2];\n";
1083
1159
 
1084
1160
  std::string op_file = op == "add_rms" ? "add.comp" : std::string(op) + ".comp";
1085
1161
  if (basename(input_filepath) != op_file) {
@@ -1087,8 +1163,8 @@ void write_output_files() {
1087
1163
  }
1088
1164
  std::stringstream data = make_generic_stringstream();
1089
1165
  std::stringstream len = make_generic_stringstream();
1090
- data << "const void * " << op << "_data[2][2][2][2] = ";
1091
- len << "const uint64_t " << op << "_len[2][2][2][2] = ";
1166
+ data << "const void * " << op << "_data[2][2][2] = ";
1167
+ len << "const uint64_t " << op << "_len[2][2][2] = ";
1092
1168
  for (uint32_t t0 = 0; t0 < 2; ++t0) {
1093
1169
  if (t0 == 0) {
1094
1170
  data << "{";
@@ -1104,20 +1180,10 @@ void write_output_files() {
1104
1180
  data << "{";
1105
1181
  len << "{";
1106
1182
  }
1107
- for (uint32_t rte = 0; rte < 2; ++rte) {
1108
- if (rte == 0) {
1109
- data << "{";
1110
- len << "{";
1111
- }
1112
- data << op << suffixes[t0] << suffixes[t1] << suffixes[t2] << ((rte != 0) ? "_rte" : "");
1113
- len << op << suffixes[t0] << suffixes[t1] << suffixes[t2] << ((rte != 0) ? "_rte" : "");
1114
- data << "_data,";
1115
- len << "_len,";
1116
- if (rte == 1) {
1117
- data << "}, ";
1118
- len << "}, ";
1119
- }
1120
- }
1183
+ data << op << suffixes[t0] << suffixes[t1] << suffixes[t2];
1184
+ len << op << suffixes[t0] << suffixes[t1] << suffixes[t2];
1185
+ data << "_data,";
1186
+ len << "_len,";
1121
1187
  if (t2 == 1) {
1122
1188
  data << "}, ";
1123
1189
  len << "}, ";
@@ -1216,6 +1282,11 @@ int main(int argc, char** argv) {
1216
1282
 
1217
1283
  process_shaders();
1218
1284
 
1285
+ if (compile_failed) {
1286
+ std::cerr << "vulkan-shaders-gen: one or more shaders failed to compile" << std::endl;
1287
+ return EXIT_FAILURE;
1288
+ }
1289
+
1219
1290
  write_output_files();
1220
1291
 
1221
1292
  return EXIT_SUCCESS;