@novastera-oss/llamarn 0.3.1 → 0.4.1

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 (347) hide show
  1. package/README.md +86 -3
  2. package/RNLlamaCpp.podspec +1 -1
  3. package/android/CMakeLists.txt +11 -3
  4. package/android/generated/jni/react/renderer/components/RNLlamaCppSpec/RNLlamaCppSpecJSI.h +49 -4
  5. package/android/src/main/cpp/include/llama.h +53 -114
  6. package/android/src/main/jniLibs/arm64-v8a/libggml-base.so +0 -0
  7. package/android/src/main/jniLibs/arm64-v8a/libggml-cpu.so +0 -0
  8. package/android/src/main/jniLibs/arm64-v8a/libggml.so +0 -0
  9. package/android/src/main/jniLibs/arm64-v8a/libllama.so +0 -0
  10. package/android/src/main/jniLibs/armeabi-v7a/libggml-base.so +0 -0
  11. package/android/src/main/jniLibs/armeabi-v7a/libggml-cpu.so +0 -0
  12. package/android/src/main/jniLibs/armeabi-v7a/libggml.so +0 -0
  13. package/android/src/main/jniLibs/armeabi-v7a/libllama.so +0 -0
  14. package/android/src/main/jniLibs/x86/libggml-base.so +0 -0
  15. package/android/src/main/jniLibs/x86/libggml-cpu.so +0 -0
  16. package/android/src/main/jniLibs/x86/libggml.so +0 -0
  17. package/android/src/main/jniLibs/x86/libllama.so +0 -0
  18. package/android/src/main/jniLibs/x86_64/libggml-base.so +0 -0
  19. package/android/src/main/jniLibs/x86_64/libggml-cpu.so +0 -0
  20. package/android/src/main/jniLibs/x86_64/libggml.so +0 -0
  21. package/android/src/main/jniLibs/x86_64/libllama.so +0 -0
  22. package/cpp/LlamaCppModel.cpp +2 -10
  23. package/cpp/PureCppImpl.cpp +71 -4
  24. package/cpp/SystemUtils.cpp +3 -7
  25. package/cpp/build-info.cpp +2 -2
  26. package/cpp/llama.cpp/CMakeLists.txt +2 -0
  27. package/cpp/llama.cpp/CODEOWNERS +1 -1
  28. package/cpp/llama.cpp/Makefile +6 -1605
  29. package/cpp/llama.cpp/README.md +5 -1
  30. package/cpp/llama.cpp/common/arg.cpp +230 -51
  31. package/cpp/llama.cpp/common/chat-parser.cpp +9 -1
  32. package/cpp/llama.cpp/common/chat.cpp +539 -8
  33. package/cpp/llama.cpp/common/chat.h +8 -1
  34. package/cpp/llama.cpp/common/common.cpp +60 -15
  35. package/cpp/llama.cpp/common/common.h +64 -15
  36. package/cpp/llama.cpp/common/speculative.cpp +135 -54
  37. package/cpp/llama.cpp/common/speculative.h +8 -1
  38. package/cpp/llama.cpp/convert_hf_to_gguf.py +1216 -109
  39. package/cpp/llama.cpp/convert_hf_to_gguf_update.py +19 -6
  40. package/cpp/llama.cpp/convert_lora_to_gguf.py +1 -1
  41. package/cpp/llama.cpp/flake.nix +0 -5
  42. package/cpp/llama.cpp/ggml/CMakeLists.txt +6 -3
  43. package/cpp/llama.cpp/ggml/cmake/ggml-config.cmake.in +71 -70
  44. package/cpp/llama.cpp/ggml/include/ggml-opt.h +25 -6
  45. package/cpp/llama.cpp/ggml/include/ggml-zdnn.h +16 -0
  46. package/cpp/llama.cpp/ggml/include/ggml.h +90 -3
  47. package/cpp/llama.cpp/ggml/src/CMakeLists.txt +13 -1
  48. package/cpp/llama.cpp/ggml/src/ggml-alloc.c +1 -0
  49. package/cpp/llama.cpp/ggml/src/ggml-backend-reg.cpp +10 -0
  50. package/cpp/llama.cpp/ggml/src/ggml-backend.cpp +113 -17
  51. package/cpp/llama.cpp/ggml/src/ggml-blas/ggml-blas.cpp +4 -4
  52. package/cpp/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +14 -0
  53. package/cpp/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +701 -585
  54. package/cpp/llama.cpp/ggml/src/ggml-cann/aclnn_ops.h +13 -3
  55. package/cpp/llama.cpp/ggml/src/ggml-cann/common.h +52 -0
  56. package/cpp/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +274 -91
  57. package/cpp/llama.cpp/ggml/src/ggml-common.h +17 -0
  58. package/cpp/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +2 -2
  59. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/arm/quants.c +132 -596
  60. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/arm/repack.cpp +14 -286
  61. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/loongarch/quants.c +90 -569
  62. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/powerpc/quants.c +162 -589
  63. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/riscv/quants.c +55 -341
  64. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/riscv/repack.cpp +3 -58
  65. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/s390/quants.c +371 -298
  66. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/wasm/quants.c +54 -314
  67. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/x86/quants.c +184 -675
  68. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch/x86/repack.cpp +4679 -1657
  69. package/cpp/llama.cpp/ggml/src/ggml-cpu/arch-fallback.h +33 -2
  70. package/cpp/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +8 -0
  71. package/cpp/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +26 -1
  72. package/cpp/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +21 -24
  73. package/cpp/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +16 -7
  74. package/cpp/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +232 -123
  75. package/cpp/llama.cpp/ggml/src/ggml-cpu/ops.cpp +428 -23
  76. package/cpp/llama.cpp/ggml/src/ggml-cpu/ops.h +4 -8
  77. package/cpp/llama.cpp/ggml/src/ggml-cpu/quants.c +35 -0
  78. package/cpp/llama.cpp/ggml/src/ggml-cpu/quants.h +8 -0
  79. package/cpp/llama.cpp/ggml/src/ggml-cpu/repack.cpp +458 -46
  80. package/cpp/llama.cpp/ggml/src/ggml-cpu/repack.h +22 -0
  81. package/cpp/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h +39 -14
  82. package/cpp/llama.cpp/ggml/src/ggml-cpu/traits.cpp +2 -2
  83. package/cpp/llama.cpp/ggml/src/ggml-cpu/traits.h +1 -1
  84. package/cpp/llama.cpp/ggml/src/ggml-cpu/vec.cpp +20 -1
  85. package/cpp/llama.cpp/ggml/src/ggml-cpu/vec.h +122 -5
  86. package/cpp/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +9 -11
  87. package/cpp/llama.cpp/ggml/src/ggml-cuda/add-id.cu +58 -0
  88. package/cpp/llama.cpp/ggml/src/ggml-cuda/add-id.cuh +3 -0
  89. package/cpp/llama.cpp/ggml/src/ggml-cuda/binbcast.cu +275 -170
  90. package/cpp/llama.cpp/ggml/src/ggml-cuda/binbcast.cuh +2 -0
  91. package/cpp/llama.cpp/ggml/src/ggml-cuda/common.cuh +103 -65
  92. package/cpp/llama.cpp/ggml/src/ggml-cuda/conv-transpose-1d.cu +1 -4
  93. package/cpp/llama.cpp/ggml/src/ggml-cuda/conv2d.cu +171 -0
  94. package/cpp/llama.cpp/ggml/src/ggml-cuda/conv2d.cuh +5 -0
  95. package/cpp/llama.cpp/ggml/src/ggml-cuda/convert.cu +33 -7
  96. package/cpp/llama.cpp/ggml/src/ggml-cuda/convert.cuh +13 -0
  97. package/cpp/llama.cpp/ggml/src/ggml-cuda/cpy-utils.cuh +2 -10
  98. package/cpp/llama.cpp/ggml/src/ggml-cuda/cpy.cu +3 -4
  99. package/cpp/llama.cpp/ggml/src/ggml-cuda/dequantize.cuh +14 -40
  100. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-common.cuh +83 -27
  101. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh +116 -57
  102. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-tile-f16.cu +45 -18
  103. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-tile-f32.cu +56 -29
  104. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-vec-f16.cuh +61 -39
  105. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-vec-f32.cuh +70 -49
  106. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn-wmma-f16.cu +70 -21
  107. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn.cu +162 -50
  108. package/cpp/llama.cpp/ggml/src/ggml-cuda/fattn.cuh +2 -0
  109. package/cpp/llama.cpp/ggml/src/ggml-cuda/getrows.cu +5 -4
  110. package/cpp/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu +208 -97
  111. package/cpp/llama.cpp/ggml/src/ggml-cuda/im2col.cu +46 -35
  112. package/cpp/llama.cpp/ggml/src/ggml-cuda/mean.cu +56 -2
  113. package/cpp/llama.cpp/ggml/src/ggml-cuda/mma.cuh +95 -51
  114. package/cpp/llama.cpp/ggml/src/ggml-cuda/mmf.cu +427 -0
  115. package/cpp/llama.cpp/ggml/src/ggml-cuda/mmf.cuh +5 -0
  116. package/cpp/llama.cpp/ggml/src/ggml-cuda/mmq.cu +204 -57
  117. package/cpp/llama.cpp/ggml/src/ggml-cuda/mmq.cuh +252 -168
  118. package/cpp/llama.cpp/ggml/src/ggml-cuda/{mmv.cu → mmvf.cu} +53 -53
  119. package/cpp/llama.cpp/ggml/src/ggml-cuda/{mmv.cuh → mmvf.cuh} +3 -3
  120. package/cpp/llama.cpp/ggml/src/ggml-cuda/mmvq.cu +10 -5
  121. package/cpp/llama.cpp/ggml/src/ggml-cuda/norm.cu +192 -19
  122. package/cpp/llama.cpp/ggml/src/ggml-cuda/norm.cuh +5 -0
  123. package/cpp/llama.cpp/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
  124. package/cpp/llama.cpp/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
  125. package/cpp/llama.cpp/ggml/src/ggml-cuda/pad_reflect_1d.cu +82 -0
  126. package/cpp/llama.cpp/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
  127. package/cpp/llama.cpp/ggml/src/ggml-cuda/reduce_rows.cuh +53 -0
  128. package/cpp/llama.cpp/ggml/src/ggml-cuda/roll.cu +67 -0
  129. package/cpp/llama.cpp/ggml/src/ggml-cuda/roll.cuh +5 -0
  130. package/cpp/llama.cpp/ggml/src/ggml-cuda/set-rows.cu +1 -8
  131. package/cpp/llama.cpp/ggml/src/ggml-cuda/softcap.cu +34 -0
  132. package/cpp/llama.cpp/ggml/src/ggml-cuda/softcap.cuh +5 -0
  133. package/cpp/llama.cpp/ggml/src/ggml-cuda/softmax.cu +16 -10
  134. package/cpp/llama.cpp/ggml/src/ggml-cuda/ssm-scan.cu +153 -71
  135. package/cpp/llama.cpp/ggml/src/ggml-cuda/sum.cu +6 -10
  136. package/cpp/llama.cpp/ggml/src/ggml-cuda/sumrows.cu +21 -4
  137. package/cpp/llama.cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
  138. package/cpp/llama.cpp/ggml/src/ggml-cuda/unary.cu +75 -0
  139. package/cpp/llama.cpp/ggml/src/ggml-cuda/unary.cuh +2 -0
  140. package/cpp/llama.cpp/ggml/src/ggml-cuda/vecdotq.cuh +110 -22
  141. package/cpp/llama.cpp/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
  142. package/cpp/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +14 -25
  143. package/cpp/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +2 -1
  144. package/cpp/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +10 -2
  145. package/cpp/llama.cpp/ggml/src/ggml-impl.h +61 -0
  146. package/cpp/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +31 -20
  147. package/cpp/llama.cpp/ggml/src/ggml-metal/ggml-metal.m +342 -131
  148. package/cpp/llama.cpp/ggml/src/ggml-metal/ggml-metal.metal +464 -134
  149. package/cpp/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +0 -4
  150. package/cpp/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +8 -0
  151. package/cpp/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +1108 -176
  152. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl +107 -0
  153. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  154. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl +66 -0
  155. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +343 -0
  156. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +343 -0
  157. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +346 -0
  158. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl +41 -0
  159. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl +49 -0
  160. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl +73 -0
  161. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +132 -0
  162. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +133 -0
  163. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  164. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  165. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl +80 -0
  166. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +10 -2
  167. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +10 -2
  168. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl +10 -2
  169. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl +10 -2
  170. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl +66 -0
  171. package/cpp/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl +20 -0
  172. package/cpp/llama.cpp/ggml/src/ggml-opt.cpp +97 -41
  173. package/cpp/llama.cpp/ggml/src/ggml-quants.c +110 -16
  174. package/cpp/llama.cpp/ggml/src/ggml-quants.h +6 -0
  175. package/cpp/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +22 -9
  176. package/cpp/llama.cpp/ggml/src/ggml-sycl/backend.hpp +1 -0
  177. package/cpp/llama.cpp/ggml/src/ggml-sycl/cpy.cpp +0 -212
  178. package/cpp/llama.cpp/ggml/src/ggml-sycl/cpy.hpp +213 -1
  179. package/cpp/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +117 -238
  180. package/cpp/llama.cpp/ggml/src/ggml-sycl/quantize.hpp +133 -0
  181. package/cpp/llama.cpp/ggml/src/ggml-sycl/set_rows.cpp +94 -0
  182. package/cpp/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1666 -633
  183. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +41 -1
  184. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
  185. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +13 -4
  186. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +39 -29
  187. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +107 -43
  188. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +2 -2
  189. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +18 -0
  190. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +21 -0
  191. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
  192. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +20 -0
  193. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +21 -0
  194. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +16 -1
  195. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +44 -8
  196. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +44 -16
  197. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +26 -1
  198. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +2 -17
  199. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +2 -0
  200. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +37 -1
  201. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +11 -7
  202. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +109 -55
  203. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +71 -41
  204. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +6 -0
  205. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +111 -0
  206. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
  207. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +49 -11
  208. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
  209. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +9 -3
  210. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +17 -0
  211. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +38 -5
  212. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
  213. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +55 -0
  214. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp +25 -0
  215. package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +75 -20
  216. package/cpp/llama.cpp/ggml/src/ggml-webgpu/CMakeLists.txt +2 -2
  217. package/cpp/llama.cpp/ggml/src/ggml-webgpu/ggml-webgpu.cpp +807 -412
  218. package/cpp/llama.cpp/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +72 -22
  219. package/cpp/llama.cpp/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +8 -8
  220. package/cpp/llama.cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +1794 -0
  221. package/cpp/llama.cpp/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +82 -0
  222. package/cpp/llama.cpp/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  223. package/cpp/llama.cpp/ggml/src/ggml-zdnn/ggml-zdnn-impl.h +97 -0
  224. package/cpp/llama.cpp/ggml/src/ggml-zdnn/ggml-zdnn.cpp +846 -0
  225. package/cpp/llama.cpp/ggml/src/ggml.c +204 -50
  226. package/cpp/llama.cpp/gguf-py/gguf/constants.py +187 -2
  227. package/cpp/llama.cpp/gguf-py/gguf/gguf_writer.py +11 -2
  228. package/cpp/llama.cpp/gguf-py/gguf/quants.py +53 -4
  229. package/cpp/llama.cpp/gguf-py/gguf/scripts/gguf_convert_endian.py +67 -63
  230. package/cpp/llama.cpp/gguf-py/gguf/scripts/gguf_new_metadata.py +7 -1
  231. package/cpp/llama.cpp/gguf-py/gguf/tensor_mapping.py +120 -16
  232. package/cpp/llama.cpp/gguf-py/gguf/utility.py +5 -1
  233. package/cpp/llama.cpp/gguf-py/gguf/vocab.py +284 -1
  234. package/cpp/llama.cpp/gguf-py/tests/test_quants.py +14 -5
  235. package/cpp/llama.cpp/include/llama.h +53 -114
  236. package/cpp/llama.cpp/models/templates/ByteDance-Seed-OSS.jinja +171 -0
  237. package/cpp/llama.cpp/models/templates/README.md +2 -1
  238. package/cpp/llama.cpp/models/templates/ibm-granite-granite-3.3-2B-Instruct.jinja +59 -0
  239. package/cpp/llama.cpp/models/templates/openai-gpt-oss-120b.jinja +331 -0
  240. package/cpp/llama.cpp/models/templates/unsloth-mistral-Devstral-Small-2507.jinja +105 -0
  241. package/cpp/llama.cpp/requirements/requirements-convert_hf_to_gguf.txt +3 -1
  242. package/cpp/llama.cpp/requirements/requirements-convert_hf_to_gguf_update.txt +0 -6
  243. package/cpp/llama.cpp/requirements/requirements-pydantic.txt +1 -1
  244. package/cpp/llama.cpp/src/CMakeLists.txt +2 -2
  245. package/cpp/llama.cpp/src/llama-adapter.cpp +68 -4
  246. package/cpp/llama.cpp/src/llama-adapter.h +3 -0
  247. package/cpp/llama.cpp/src/llama-arch.cpp +192 -2
  248. package/cpp/llama.cpp/src/llama-arch.h +18 -0
  249. package/cpp/llama.cpp/src/llama-batch.cpp +2 -2
  250. package/cpp/llama.cpp/src/llama-chat.cpp +47 -6
  251. package/cpp/llama.cpp/src/llama-chat.h +3 -0
  252. package/cpp/llama.cpp/src/llama-context.cpp +61 -252
  253. package/cpp/llama.cpp/src/llama-context.h +10 -15
  254. package/cpp/llama.cpp/src/llama-cparams.h +0 -1
  255. package/cpp/llama.cpp/src/llama-graph.cpp +180 -85
  256. package/cpp/llama.cpp/src/llama-graph.h +90 -51
  257. package/cpp/llama.cpp/src/llama-hparams.cpp +34 -3
  258. package/cpp/llama.cpp/src/llama-hparams.h +21 -6
  259. package/cpp/llama.cpp/src/{llama-kv-cache-unified-iswa.cpp → llama-kv-cache-iswa.cpp} +79 -56
  260. package/cpp/llama.cpp/src/{llama-kv-cache-unified-iswa.h → llama-kv-cache-iswa.h} +30 -28
  261. package/cpp/llama.cpp/src/{llama-kv-cache-unified.cpp → llama-kv-cache.cpp} +240 -632
  262. package/cpp/llama.cpp/src/{llama-kv-cache-unified.h → llama-kv-cache.h} +39 -74
  263. package/cpp/llama.cpp/src/llama-kv-cells.h +21 -21
  264. package/cpp/llama.cpp/src/llama-memory-hybrid.cpp +41 -35
  265. package/cpp/llama.cpp/src/llama-memory-hybrid.h +26 -29
  266. package/cpp/llama.cpp/src/llama-memory-recurrent.cpp +13 -9
  267. package/cpp/llama.cpp/src/llama-memory-recurrent.h +10 -14
  268. package/cpp/llama.cpp/src/llama-memory.h +13 -10
  269. package/cpp/llama.cpp/src/llama-model-loader.cpp +2 -0
  270. package/cpp/llama.cpp/src/llama-model-loader.h +3 -2
  271. package/cpp/llama.cpp/src/llama-model.cpp +1959 -419
  272. package/cpp/llama.cpp/src/llama-model.h +28 -4
  273. package/cpp/llama.cpp/src/llama-quant.cpp +40 -4
  274. package/cpp/llama.cpp/src/llama-vocab.cpp +51 -2
  275. package/cpp/llama.cpp/src/llama-vocab.h +1 -0
  276. package/cpp/llama.cpp/vendor/minja/chat-template.hpp +16 -7
  277. package/cpp/llama.cpp/vendor/minja/minja.hpp +47 -12
  278. package/cpp/rn-completion.cpp +3 -27
  279. package/ios/generated/RNLlamaCppSpec/RNLlamaCppSpec.h +30 -0
  280. package/ios/generated/RNLlamaCppSpecJSI.h +49 -4
  281. package/ios/include/chat.h +8 -1
  282. package/ios/include/common/minja/chat-template.hpp +16 -7
  283. package/ios/include/common/minja/minja.hpp +47 -12
  284. package/ios/include/common.h +64 -15
  285. package/ios/include/llama.h +53 -114
  286. package/ios/include/speculative.h +8 -1
  287. package/ios/libs/llama.xcframework/Info.plist +18 -18
  288. package/ios/libs/llama.xcframework/ios-arm64/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  289. package/ios/libs/llama.xcframework/ios-arm64/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5557 -5267
  290. package/ios/libs/llama.xcframework/ios-arm64/llama.framework/Headers/ggml-opt.h +25 -6
  291. package/ios/libs/llama.xcframework/ios-arm64/llama.framework/Headers/ggml.h +90 -3
  292. package/ios/libs/llama.xcframework/ios-arm64/llama.framework/Headers/llama.h +53 -114
  293. package/ios/libs/llama.xcframework/ios-arm64/llama.framework/llama +0 -0
  294. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  295. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5520 -5238
  296. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/x86_64/llama.yml +4241 -4014
  297. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/llama.framework/Headers/ggml-opt.h +25 -6
  298. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/llama.framework/Headers/ggml.h +90 -3
  299. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/llama.framework/Headers/llama.h +53 -114
  300. package/ios/libs/llama.xcframework/ios-arm64_x86_64-simulator/llama.framework/llama +0 -0
  301. package/ios/libs/llama.xcframework/macos-arm64_x86_64/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  302. package/ios/libs/llama.xcframework/macos-arm64_x86_64/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5519 -5238
  303. package/ios/libs/llama.xcframework/macos-arm64_x86_64/dSYMs/llama.dSYM/Contents/Resources/Relocations/x86_64/llama.yml +4242 -4016
  304. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Headers/ggml-opt.h +25 -6
  305. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Headers/ggml.h +90 -3
  306. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Headers/llama.h +53 -114
  307. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/Headers/ggml-opt.h +25 -6
  308. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/Headers/ggml.h +90 -3
  309. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/Headers/llama.h +53 -114
  310. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/llama +0 -0
  311. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/Current/Headers/ggml-opt.h +25 -6
  312. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/Current/Headers/ggml.h +90 -3
  313. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/Current/Headers/llama.h +53 -114
  314. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/Current/llama +0 -0
  315. package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/llama +0 -0
  316. package/ios/libs/llama.xcframework/tvos-arm64/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  317. package/ios/libs/llama.xcframework/tvos-arm64/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5556 -5267
  318. package/ios/libs/llama.xcframework/tvos-arm64/llama.framework/Headers/ggml-opt.h +25 -6
  319. package/ios/libs/llama.xcframework/tvos-arm64/llama.framework/Headers/ggml.h +90 -3
  320. package/ios/libs/llama.xcframework/tvos-arm64/llama.framework/Headers/llama.h +53 -114
  321. package/ios/libs/llama.xcframework/tvos-arm64/llama.framework/llama +0 -0
  322. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  323. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5519 -5238
  324. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/x86_64/llama.yml +4241 -4014
  325. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/llama.framework/Headers/ggml-opt.h +25 -6
  326. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/llama.framework/Headers/ggml.h +90 -3
  327. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/llama.framework/Headers/llama.h +53 -114
  328. package/ios/libs/llama.xcframework/tvos-arm64_x86_64-simulator/llama.framework/llama +0 -0
  329. package/ios/libs/llama.xcframework/xros-arm64/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  330. package/ios/libs/llama.xcframework/xros-arm64/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5553 -5303
  331. package/ios/libs/llama.xcframework/xros-arm64/llama.framework/Headers/ggml-opt.h +25 -6
  332. package/ios/libs/llama.xcframework/xros-arm64/llama.framework/Headers/ggml.h +90 -3
  333. package/ios/libs/llama.xcframework/xros-arm64/llama.framework/Headers/llama.h +53 -114
  334. package/ios/libs/llama.xcframework/xros-arm64/llama.framework/llama +0 -0
  335. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/DWARF/llama +0 -0
  336. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/aarch64/llama.yml +5515 -5274
  337. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/dSYMs/llama.dSYM/Contents/Resources/Relocations/x86_64/llama.yml +4238 -4044
  338. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/llama.framework/Headers/ggml-opt.h +25 -6
  339. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/llama.framework/Headers/ggml.h +90 -3
  340. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/llama.framework/Headers/llama.h +53 -114
  341. package/ios/libs/llama.xcframework/xros-arm64_x86_64-simulator/llama.framework/llama +0 -0
  342. package/lib/module/NativeRNLlamaCpp.js.map +1 -1
  343. package/lib/typescript/src/NativeRNLlamaCpp.d.ts +5 -0
  344. package/lib/typescript/src/NativeRNLlamaCpp.d.ts.map +1 -1
  345. package/package.json +1 -2
  346. package/src/NativeRNLlamaCpp.ts +7 -0
  347. package/cpp/llama.cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -56
@@ -0,0 +1,1794 @@
1
+ #define(VARIANTS)
2
+
3
+ [
4
+ {
5
+ "REPLS": {
6
+ "SRC0_TYPE" : "f32",
7
+ "SRC1_TYPE" : "f32",
8
+ "BLOCK_SIZE" : 1
9
+ },
10
+ "DECLS" : ["FLOAT"]
11
+ },
12
+ {
13
+ "REPLS": {
14
+ "SRC0_TYPE" : "f16",
15
+ "SRC1_TYPE" : "f16",
16
+ "BLOCK_SIZE" : 1
17
+ },
18
+ "DECLS" : ["FLOAT"]
19
+ },
20
+ {
21
+ "REPLS": {
22
+ "SRC0_TYPE" : "f16",
23
+ "SRC1_TYPE" : "f32",
24
+ "BLOCK_SIZE" : 1
25
+ },
26
+ "DECLS" : ["FLOAT"]
27
+ },
28
+ {
29
+ "REPLS": {
30
+ "SRC0_TYPE": "q4_0",
31
+ "SRC1_TYPE": "f32",
32
+ "BLOCK_SIZE": 32
33
+ },
34
+ "DECLS": ["BYTE_HELPERS", "Q4_0"]
35
+ },
36
+ {
37
+ "REPLS": {
38
+ "SRC0_TYPE": "q4_1",
39
+ "SRC1_TYPE": "f32",
40
+ "BLOCK_SIZE": 32
41
+ },
42
+ "DECLS": ["BYTE_HELPERS", "Q4_1"]
43
+ },
44
+ {
45
+ "REPLS": {
46
+ "SRC0_TYPE": "q5_0",
47
+ "SRC1_TYPE": "f32",
48
+ "BLOCK_SIZE": 32
49
+ },
50
+ "DECLS": ["BYTE_HELPERS", "Q5_0"]
51
+ },
52
+ {
53
+ "REPLS": {
54
+ "SRC0_TYPE": "q5_1",
55
+ "SRC1_TYPE": "f32",
56
+ "BLOCK_SIZE": 32
57
+ },
58
+ "DECLS": ["BYTE_HELPERS", "Q5_1"]
59
+ },
60
+ {
61
+ "REPLS": {
62
+ "SRC0_TYPE": "q8_0",
63
+ "SRC1_TYPE": "f32",
64
+ "BLOCK_SIZE": 32
65
+ },
66
+ "DECLS": ["BYTE_HELPERS", "Q8_0"]
67
+ },
68
+ {
69
+ "REPLS": {
70
+ "SRC0_TYPE": "q2_k",
71
+ "SRC1_TYPE": "f32",
72
+ "BLOCK_SIZE": 256
73
+ },
74
+ "DECLS": ["BYTE_HELPERS", "Q2_K"]
75
+ },
76
+ {
77
+ "REPLS": {
78
+ "SRC0_TYPE": "q3_k",
79
+ "SRC1_TYPE": "f32",
80
+ "BLOCK_SIZE": 256
81
+ },
82
+ "DECLS": ["BYTE_HELPERS", "Q3_K"]
83
+ },
84
+ {
85
+ "REPLS": {
86
+ "SRC0_TYPE": "q4_k",
87
+ "SRC1_TYPE": "f32",
88
+ "BLOCK_SIZE": 256
89
+ },
90
+ "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K"]
91
+ },
92
+ {
93
+ "REPLS": {
94
+ "SRC0_TYPE": "q5_k",
95
+ "SRC1_TYPE": "f32",
96
+ "BLOCK_SIZE": 256
97
+ },
98
+ "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K"]
99
+ },
100
+ {
101
+ "REPLS": {
102
+ "SRC0_TYPE": "q6_k",
103
+ "SRC1_TYPE": "f32",
104
+ "BLOCK_SIZE": 256
105
+ },
106
+ "DECLS": ["BYTE_HELPERS", "Q6_K"]
107
+ },
108
+ {
109
+ "REPLS": {
110
+ "SRC0_TYPE": "iq2_xxs",
111
+ "SRC1_TYPE": "f32",
112
+ "BLOCK_SIZE": 256
113
+ },
114
+ "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS"]
115
+ },
116
+ {
117
+ "REPLS": {
118
+ "SRC0_TYPE": "iq2_xs",
119
+ "SRC1_TYPE": "f32",
120
+ "BLOCK_SIZE": 256
121
+ },
122
+ "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS"]
123
+ },
124
+ {
125
+ "REPLS": {
126
+ "SRC0_TYPE": "iq2_s",
127
+ "SRC1_TYPE": "f32",
128
+ "BLOCK_SIZE": 256
129
+ },
130
+ "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S"]
131
+ },
132
+ {
133
+ "REPLS": {
134
+ "SRC0_TYPE": "iq3_xxs",
135
+ "SRC1_TYPE": "f32",
136
+ "BLOCK_SIZE": 256
137
+ },
138
+ "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS"]
139
+ },
140
+ {
141
+ "REPLS": {
142
+ "SRC0_TYPE": "iq3_s",
143
+ "SRC1_TYPE": "f32",
144
+ "BLOCK_SIZE": 256
145
+ },
146
+ "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S"]
147
+ },
148
+ {
149
+ "REPLS": {
150
+ "SRC0_TYPE": "iq1_s",
151
+ "SRC1_TYPE": "f32",
152
+ "BLOCK_SIZE": 256
153
+ },
154
+ "DECLS": ["BYTE_HELPERS", "IQ1_TABLE","IQ1_S"]
155
+ },
156
+ {
157
+ "REPLS": {
158
+ "SRC0_TYPE": "iq1_m",
159
+ "SRC1_TYPE": "f32",
160
+ "BLOCK_SIZE": 256
161
+ },
162
+ "DECLS": ["BYTE_HELPERS", "IQ1_TABLE","IQ1_M"]
163
+ },
164
+ {
165
+ "REPLS": {
166
+ "SRC0_TYPE": "iq4_nl",
167
+ "SRC1_TYPE": "f32",
168
+ "BLOCK_SIZE": 32,
169
+ },
170
+ "DECLS": ["BYTE_HELPERS", "IQ4_TABLE", "IQ4_NL"]
171
+ },
172
+ {
173
+ "REPLS": {
174
+ "SRC0_TYPE": "iq4_xs",
175
+ "SRC1_TYPE": "f32",
176
+ "BLOCK_SIZE": 256,
177
+ },
178
+ "DECLS": ["BYTE_HELPERS", "IQ4_TABLE", "IQ4_XS"]
179
+ }
180
+ ]
181
+
182
+ #end(VARIANTS)
183
+
184
+ #define(DECLS)
185
+
186
+ #decl(BYTE_HELPERS)
187
+
188
+ fn get_byte(value: u32, index: u32) -> u32 {
189
+ return (value >> (index * 8)) & 0xFF;
190
+ }
191
+
192
+ fn get_byte_i32(value: u32, index: u32) -> i32 {
193
+ return bitcast<i32>(((value >> (index * 8)) & 0xFF) << 24) >> 24;
194
+ }
195
+
196
+ #enddecl(BYTE_HELPERS)
197
+
198
+ #decl(FLOAT)
199
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
200
+ return f32(src0[src0_idx_base + offset]) * f32(src1[src1_idx_base + offset]);
201
+ }
202
+ #enddecl(FLOAT)
203
+
204
+ #decl(Q4_0)
205
+ struct q4_0 {
206
+ d: f16,
207
+ qs: array<f16, 8>
208
+ };
209
+
210
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
211
+ let block_q4_0 = src0[src0_idx_base + offset];
212
+ let d = f32(block_q4_0.d);
213
+ var sum: f32 = 0.0;
214
+ for (var j: u32 = 0; j < 4; j++) {
215
+ let q_packed = bitcast<u32>(vec2(block_q4_0.qs[2 * j], block_q4_0.qs[2 * j + 1]));
216
+ for (var k: u32 = 0; k < 4; k++) {
217
+ let q_byte = get_byte(q_packed, k);
218
+ let q_hi = (f32((q_byte >> 4) & 0xF) - 8.0f) * d;
219
+ let q_lo = (f32(q_byte & 0xF) - 8.0f) * d;
220
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
221
+ sum += q_lo * f32(src1[src1_offset]);
222
+ sum += q_hi * f32(src1[src1_offset + 16]);
223
+ }
224
+ }
225
+ return sum;
226
+ }
227
+ #enddecl(Q4_0)
228
+
229
+ #decl(Q4_1)
230
+ struct q4_1 {
231
+ d: f16,
232
+ m: f16,
233
+ qs: array<u32, 4>
234
+ };
235
+
236
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
237
+ let block_q4_1 = src0[src0_idx_base + offset];
238
+ let d = f32(block_q4_1.d);
239
+ let m = f32(block_q4_1.m);
240
+ var sum: f32 = 0.0;
241
+ for (var j: u32 = 0; j < 4; j++) {
242
+ let q_packed = block_q4_1.qs[j];
243
+ for (var k: u32 = 0; k < 4; k++) {
244
+ let q_byte = get_byte(q_packed, k);
245
+ let q_hi = f32((q_byte >> 4) & 0xF) * d + m;
246
+ let q_lo = f32(q_byte & 0xF) * d + m;
247
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
248
+ sum += q_lo * f32(src1[src1_offset]);
249
+ sum += q_hi * f32(src1[src1_offset + 16]);
250
+ }
251
+ }
252
+ return sum;
253
+ }
254
+ #enddecl(Q4_1)
255
+
256
+ #decl(Q5_0)
257
+ struct q5_0 {
258
+ d: f16,
259
+ qh: array<f16, 2>,
260
+ qs: array<f16, 8>
261
+ };
262
+
263
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
264
+ let block_q5_0 = src0[src0_idx_base + offset];
265
+ let d = f32(block_q5_0.d);
266
+ var sum: f32 = 0.0;
267
+ let qh_packed = bitcast<u32>(vec2(block_q5_0.qh[0], block_q5_0.qh[1]));
268
+ for (var j: u32 = 0; j < 4; j++) {
269
+ let q_packed = bitcast<u32>(vec2(block_q5_0.qs[2 * j], block_q5_0.qs[2 * j + 1]));
270
+ for (var k: u32 = 0; k < 4; k++) {
271
+ let q_byte = get_byte(q_packed, k);
272
+ let qh_hi = (qh_packed >> (j * 4 + k + 12)) & 0x10;
273
+ let q_hi = (f32(((q_byte >> 4) & 0xF) | qh_hi) - 16.0) * d;
274
+ let qh_lo = ((qh_packed >> (j * 4 + k)) << 4) & 0x10;
275
+ let q_lo = (f32((q_byte & 0xF) | qh_lo) - 16.0) * d;
276
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
277
+ sum += q_lo * f32(src1[src1_offset]);
278
+ sum += q_hi * f32(src1[src1_offset + 16]);
279
+ }
280
+ }
281
+ return sum;
282
+ }
283
+ #enddecl(Q5_0)
284
+
285
+ #decl(Q5_1)
286
+ struct q5_1 {
287
+ d: f16,
288
+ m: f16,
289
+ qh: u32,
290
+ qs: array<u32, 4>
291
+ };
292
+
293
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
294
+ let block_q5_1 = src0[src0_idx_base + offset];
295
+ let d = f32(block_q5_1.d);
296
+ let m = f32(block_q5_1.m);
297
+ var sum: f32 = 0.0;
298
+ for (var j: u32 = 0; j < 4; j++) {
299
+ let q_packed = block_q5_1.qs[j];
300
+ for (var k: u32 = 0; k < 4; k++) {
301
+ let q_byte = get_byte(q_packed, k);
302
+ let qh_hi = (block_q5_1.qh >> (j * 4 + k + 12)) & 0x10;
303
+ let q_hi = f32(((q_byte >> 4) & 0xF) | qh_hi) * d + m;
304
+ let qh_lo = ((block_q5_1.qh >> (j * 4 + k)) << 4) & 0x10;
305
+ let q_lo = f32((q_byte & 0xF) | qh_lo) * d + m;
306
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
307
+ sum += q_lo * f32(src1[src1_offset]);
308
+ sum += q_hi * f32(src1[src1_offset + 16]);
309
+ }
310
+ }
311
+ return sum;
312
+ }
313
+ #enddecl(Q5_1)
314
+
315
+ #decl(Q8_0)
316
+ struct q8_0 {
317
+ d: f16,
318
+ qs: array<f16, 16>
319
+ };
320
+
321
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
322
+ let block_q8_0 = src0[src0_idx_base + offset];
323
+ let d = f32(block_q8_0.d);
324
+ var sum: f32 = 0.0;
325
+ for (var j: u32 = 0; j < 8; j++) {
326
+ let q_packed = bitcast<u32>(vec2(block_q8_0.qs[2 * j], block_q8_0.qs[2 * j + 1]));
327
+ for (var k: u32 = 0; k < 4; k++) {
328
+ let q_byte = get_byte_i32(q_packed, k);
329
+ let q_val = f32(q_byte) * d;
330
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
331
+ sum += q_val * f32(src1[src1_offset]);
332
+ }
333
+ }
334
+ return sum;
335
+ }
336
+ #enddecl(Q8_0)
337
+
338
+ #decl(Q8_1)
339
+ struct q8_1 {
340
+ d: f16,
341
+ m: f16,
342
+ qs: array<u32, 8>
343
+ };
344
+
345
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
346
+ let block_q8_1 = src0[src0_idx_base + offset];
347
+ let d = f32(block_q8_1.d);
348
+ let m = f32(block_q8_1.m);
349
+ var sum: f32 = 0.0;
350
+ for (var j: u32 = 0; j < 8; j++) {
351
+ let q_packed = block_q8_1.qs[j];
352
+ for (var k: u32 = 0; k < 4; k++) {
353
+ let q_byte = get_byte_i32(q_packed, k);
354
+ let q_val = f32(q_byte) * d + m;
355
+ let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
356
+ sum += q_val * f32(src1[src1_offset]);
357
+ }
358
+ }
359
+ return sum;
360
+ }
361
+ #enddecl(Q8_1)
362
+
363
+ #decl(Q2_K)
364
+ // 16 blocks of 16 elements each
365
+ struct q2_k {
366
+ scales: array<u32, 4>,
367
+ qs: array<u32, 16>,
368
+ d: f16,
369
+ dmin: f16
370
+ };
371
+
372
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
373
+ let block = src0[src0_idx_base + offset];
374
+ let d = f32(block.d);
375
+ let m = f32(block.dmin);
376
+ var sum = 0.0;
377
+ var src1_i = src1_idx_base + offset * 256;
378
+ var is: u32 = 0;
379
+ // 2 halves of the block (128 elements each)
380
+ for (var q_b_idx: u32 = 0; q_b_idx < 64; q_b_idx += 32) {
381
+ // 4 groups (each group has 2 blocks of 16 elements)
382
+ for (var shift: u32 = 0; shift < 8; shift += 2) {
383
+ // 2 blocks
384
+ for (var k: u32 = 0; k < 32; k += 16) {
385
+ let sc = get_byte(block.scales[is / 4], is % 4);
386
+ is++;
387
+ let dl = d * f32(sc & 0xF);
388
+ let ml = m * f32(sc >> 4);
389
+ for (var l: u32 = 0u; l < 16; l++) {
390
+ let q_idx = q_b_idx + k + l;
391
+ let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
392
+ let qs_val = (q_byte >> shift) & 3;
393
+ sum += (f32(qs_val) * dl - ml) * src1[src1_i];
394
+ src1_i++;
395
+ }
396
+ }
397
+ }
398
+ }
399
+ return sum;
400
+ }
401
+
402
+ #enddecl(Q2_K)
403
+
404
+ #decl(Q3_K)
405
+ // 16 blocks of 16 elements each
406
+ struct q3_k {
407
+ hmask: array<f16, 16>,
408
+ qs: array<f16, 32>,
409
+ scales: array<f16, 6>, // 6-bit quantized values
410
+ d: f16
411
+ };
412
+
413
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
414
+ let block = src0[src0_idx_base + offset];
415
+ let d = f32(block.d);
416
+
417
+ // extract 6-bit scales, which consist of 4-bits from first 8 bytes of scale,
418
+ // and 2-bits from the last 4 bytes
419
+ let kmask1: u32 = 0x03030303;
420
+ let kmask2: u32 = 0x0f0f0f0f;
421
+ var scale_vals: array<u32, 4>;
422
+ for (var i: u32 = 0; i < 4; i++) {
423
+ scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
424
+ }
425
+ var tmp: u32 = scale_vals[2];
426
+ scale_vals[2] = ((scale_vals[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
427
+ scale_vals[3] = ((scale_vals[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
428
+ scale_vals[0] = (scale_vals[0] & kmask2) | ((tmp & kmask1) << 4);
429
+ scale_vals[1] = (scale_vals[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
430
+
431
+ // convert arrays of f16 -> u32
432
+ var hmask_vals: array<u32, 8>;
433
+ for (var i: u32 = 0; i < 8; i++) {
434
+ hmask_vals[i] = bitcast<u32>(vec2(block.hmask[2 * i], block.hmask[2 * i + 1]));
435
+ }
436
+ var qs_vals: array<u32, 16>;
437
+ for (var i: u32 = 0; i < 16; i++) {
438
+ qs_vals[i] = bitcast<u32>(vec2(block.qs[2 * i], block.qs[2 * i + 1]));
439
+ }
440
+
441
+ var sum = 0.0;
442
+ var src1_i = src1_idx_base + offset * 256;
443
+ var is: u32 = 0;
444
+ var m: u32 = 1;
445
+ // 2 halves of the block (128 elements each)
446
+ for (var q_b_idx: u32 = 0; q_b_idx < 64; q_b_idx += 32) {
447
+ // 4 groups (each group has 2 blocks of 16 elements)
448
+ for (var shift: u32 = 0; shift < 8; shift += 2) {
449
+ // 2 blocks
450
+ for (var k: u32 = 0; k < 32; k += 16) {
451
+ let sc = get_byte(scale_vals[is / 4], is % 4);
452
+ is++;
453
+ let dl = d * (f32(sc) - 32.0);
454
+ for (var l: u32 = 0u; l < 16u; l++) {
455
+ let q_idx = q_b_idx + k + l;
456
+ let hm_idx = k + l;
457
+ let q_byte = get_byte(qs_vals[q_idx / 4], q_idx % 4);
458
+ let hmask_byte = get_byte(hmask_vals[hm_idx / 4], hm_idx % 4);
459
+ let hm = select(4.0, 0.0, (hmask_byte & m) != 0);
460
+ let qs_val = (q_byte >> shift) & 3;
461
+ sum += ((f32(qs_val) - hm) * dl) * src1[src1_i];
462
+ src1_i++;
463
+ }
464
+ }
465
+ m <<= 1;
466
+ }
467
+ }
468
+ return sum;
469
+ }
470
+
471
+ #enddecl(Q3_K)
472
+
473
+ #decl(Q45_K_SCALE_MIN)
474
+
475
+ fn get_scale_min(is: u32, scales: array<u32, 3>) -> vec2<f32> {
476
+ if (is < 4) {
477
+ let sc_byte = get_byte(scales[is / 4], is % 4);
478
+ let min_byte = get_byte(scales[(is + 4) / 4], is % 4);
479
+ return vec2(f32(sc_byte & 63), f32(min_byte & 63));
480
+ } else {
481
+ let sc_min_lo = get_byte(scales[(is + 4) / 4], (is + 4) % 4);
482
+ let sc_hi = get_byte(scales[(is - 4) / 4], (is - 4) % 4);
483
+ let min_hi = get_byte(scales[is / 4], is % 4);
484
+ let sc = (sc_min_lo & 0xF) | ((sc_hi >> 6) << 4);
485
+ let m = (sc_min_lo >> 4) | ((min_hi >> 6) << 4);
486
+ return vec2(f32(sc), f32(m));
487
+ }
488
+ }
489
+
490
+ #enddecl(Q45_K_SCALE_MIN)
491
+
492
+ #decl(Q4_K)
493
+ // 8 blocks of 32 elements each
494
+ struct q4_k {
495
+ d: f16,
496
+ dmin: f16,
497
+ scales: array<u32, 3>,
498
+ qs: array<u32, 32>
499
+ };
500
+
501
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
502
+ let block = src0[src0_idx_base + offset];
503
+ let d = f32(block.d);
504
+ let m = f32(block.dmin);
505
+ var sum = 0.0;
506
+ var src1_i = src1_idx_base + offset * 256;
507
+ var is: u32 = 0;
508
+ // 2 blocks each iteration
509
+ for (var q_b_idx: u32 = 0; q_b_idx < 128; q_b_idx += 32) {
510
+ for (var shift: u32 = 0; shift < 8; shift += 4) {
511
+ let scale_min = get_scale_min(is, block.scales);
512
+ is++;
513
+ let dl = d * scale_min.x;
514
+ let ml = m * scale_min.y;
515
+ for (var l: u32 = 0; l < 32; l++) {
516
+ let q_idx = q_b_idx + l;
517
+ let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
518
+ let qs_val = (q_byte >> shift) & 0xF;
519
+ sum += (f32(qs_val) * dl - ml) * src1[src1_i];
520
+ src1_i++;
521
+ }
522
+ }
523
+ }
524
+ return sum;
525
+ }
526
+
527
+ #enddecl(Q4_K)
528
+
529
+ #decl(Q5_K)
530
+ // 8 blocks of 32 elements each
531
+ struct q5_k {
532
+ d: f16,
533
+ dmin: f16,
534
+ scales: array<u32, 3>,
535
+ qh: array<u32, 8>,
536
+ qs: array<u32, 32>
537
+ };
538
+
539
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
540
+ let block = src0[src0_idx_base + offset];
541
+ let d = f32(block.d);
542
+ let m = f32(block.dmin);
543
+ var sum = 0.0;
544
+ var src1_i = src1_idx_base + offset * 256;
545
+ var is: u32 = 0;
546
+ var u: u32 = 1;
547
+ // 2 blocks each iteration
548
+ for (var q_b_idx: u32 = 0; q_b_idx < 128; q_b_idx += 32) {
549
+ for (var shift: u32 = 0; shift < 8; shift += 4) {
550
+ let scale_min = get_scale_min(is, block.scales);
551
+ is++;
552
+ let dl = d * scale_min.x;
553
+ let ml = m * scale_min.y;
554
+ for (var l: u32 = 0; l < 32; l++) {
555
+ let q_idx = q_b_idx + l;
556
+ let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
557
+ let qh_byte = get_byte(block.qh[l / 4], l % 4);
558
+ let qs_val = (q_byte >> shift) & 0xF;
559
+ let qh_val = select(0.0, 16.0, (qh_byte & u) != 0);
560
+ sum += ((f32(qs_val) + qh_val) * dl - ml) * src1[src1_i];
561
+ src1_i++;
562
+ }
563
+ u <<= 1;
564
+ }
565
+ }
566
+ return sum;
567
+ }
568
+
569
+ #enddecl(Q5_K)
570
+
571
+ #decl(Q6_K)
572
+ // 16 blocks of 16 elements each
573
+ struct q6_k {
574
+ ql: array<f16, 64>,
575
+ qh: array<f16, 32>,
576
+ scales: array<f16, 8>,
577
+ d: f16
578
+ };
579
+
580
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
581
+ let block = src0[src0_idx_base + offset];
582
+ let d = f32(block.d);
583
+
584
+ // convert arrays of f16 -> u32
585
+ var ql_vals: array<u32, 32>;
586
+ for (var i: u32 = 0; i < 32; i++) {
587
+ ql_vals[i] = bitcast<u32>(vec2(block.ql[2 * i], block.ql[2 * i + 1]));
588
+ }
589
+ var qh_vals: array<u32, 16>;
590
+ for (var i: u32 = 0; i < 16; i++) {
591
+ qh_vals[i] = bitcast<u32>(vec2(block.qh[2 * i], block.qh[2 * i + 1]));
592
+ }
593
+ var scale_vals: array<u32, 4>;
594
+ for (var i: u32 = 0; i < 4; i++) {
595
+ scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
596
+ }
597
+
598
+ var sum = 0.0;
599
+ var src1_i = src1_idx_base + offset * 256;
600
+ var qh_b_idx: u32 = 0;
601
+ var sc_b_idx: u32 = 0;
602
+ for (var ql_b_idx: u32 = 0; ql_b_idx < 128; ql_b_idx += 64) {
603
+ for (var l: u32 = 0; l < 32; l++) {
604
+ let ql13_b = get_byte(ql_vals[(ql_b_idx + l) / 4], (ql_b_idx + l) % 4);
605
+ let ql24_b = get_byte(ql_vals[(ql_b_idx + l + 32) / 4], (ql_b_idx + l + 32) % 4);
606
+ let qh_b = get_byte(qh_vals[(qh_b_idx + l) / 4], (qh_b_idx + l) % 4);
607
+
608
+ let q1 = f32((ql13_b & 0xF) | ((qh_b & 3) << 4)) - 32.0;
609
+ let q2 = f32((ql24_b & 0xF) | (((qh_b >> 2) & 3) << 4)) - 32.0;
610
+ let q3 = f32((ql13_b >> 4) | (((qh_b >> 4) & 3) << 4)) - 32.0;
611
+ let q4 = f32((ql24_b >> 4) | (((qh_b >> 6) & 3) << 4)) - 32.0;
612
+
613
+ let is = l/16;
614
+ let is1 = sc_b_idx + is;
615
+ let sc1 = get_byte_i32(scale_vals[is1 / 4], is1 % 4);
616
+ let is2 = sc_b_idx + is + 2;
617
+ let sc2 = get_byte_i32(scale_vals[is2 / 4], is2 % 4);
618
+ let is3 = sc_b_idx + is + 4;
619
+ let sc3 = get_byte_i32(scale_vals[is3 / 4], is3 % 4);
620
+ let is4 = sc_b_idx + is + 6;
621
+ let sc4 = get_byte_i32(scale_vals[is4 / 4], is4 % 4);
622
+
623
+ sum += d * f32(sc1) * q1 * src1[src1_i + l];
624
+ sum += d * f32(sc2) * q2 * src1[src1_i + l + 32];
625
+ sum += d * f32(sc3) * q3 * src1[src1_i + l + 64];
626
+ sum += d * f32(sc4) * q4 * src1[src1_i + l + 96];
627
+ }
628
+ src1_i += 128;
629
+ qh_b_idx += 32;
630
+ sc_b_idx += 8;
631
+ }
632
+ return sum;
633
+ }
634
+
635
+ #enddecl(Q6_K)
636
+
637
+ #decl(IQ23_TABLES)
638
+ const kmask_iq2xs : array<u32, 2> = array<u32, 2>(
639
+ 0x08040201u, // 1, 2, 4, 8
640
+ 0x80402010u // 16, 32, 64, 128
641
+ );
642
+
643
+ const ksigns_iq2xs: array<u32, 32> = array<u32, 32>(
644
+ 0x03828100,0x87060584,0x8b0a0988,0x0f8e8d0c,
645
+ 0x93121190,0x17969514,0x1b9a9918,0x9f1e1d9c,
646
+ 0xa32221a0,0x27a6a524,0x2baaa928,0xaf2e2dac,
647
+ 0x33b2b130,0xb73635b4,0xbb3a39b8,0x3fbebd3c,
648
+ 0xc34241c0,0x47c6c544,0x4bcac948,0xcf4e4dcc,
649
+ 0x53d2d150,0xd75655d4,0xdb5a59d8,0x5fdedd5c,
650
+ 0x63e2e160,0xe76665e4,0xeb6a69e8,0x6feeed6c,
651
+ 0xf37271f0,0x77f6f574,0x7bfaf978,0xff7e7dfc
652
+ );
653
+ #enddecl(IQ23_TABLES)
654
+
655
+ #decl(IQ2_XXS)
656
+
657
+ const iq2xxs_grid = array<u32, 512>(
658
+ 0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
659
+ 0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x082b0808, 0x08080808,
660
+ 0x082b082b, 0x08080808, 0x082b2b08, 0x08080808, 0x082b2b2b, 0x08080808, 0x19080819, 0x08080808,
661
+ 0x19081908, 0x08080808, 0x19190808, 0x08080808, 0x19192b08, 0x08080808, 0x192b0819, 0x08080808,
662
+ 0x192b1908, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b082b2b, 0x08080808,
663
+ 0x2b2b082b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819, 0x08190808, 0x08080819,
664
+ 0x08191919, 0x08080819, 0x19080808, 0x08080819, 0x2b081908, 0x08080819, 0x2b192b08, 0x08080819,
665
+ 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x082b082b, 0x0808082b, 0x2b08082b, 0x0808082b,
666
+ 0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x08190808, 0x08081908, 0x082b0819, 0x08081908,
667
+ 0x082b1908, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19082b08, 0x08081908,
668
+ 0x192b0808, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908,
669
+ 0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919, 0x08082b08, 0x08081919,
670
+ 0x082b0808, 0x08081919, 0x1908192b, 0x08081919, 0x192b2b19, 0x08081919, 0x2b080808, 0x08081919,
671
+ 0x2b190819, 0x08081919, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b, 0x19080808, 0x0808192b,
672
+ 0x2b081908, 0x0808192b, 0x2b2b1908, 0x0808192b, 0x08080808, 0x08082b08, 0x08081919, 0x08082b08,
673
+ 0x08082b08, 0x08082b08, 0x08191908, 0x08082b08, 0x082b2b08, 0x08082b08, 0x19080819, 0x08082b08,
674
+ 0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x2b082b08, 0x08082b08,
675
+ 0x08081908, 0x08082b19, 0x19080808, 0x08082b19, 0x0808082b, 0x08082b2b, 0x08191908, 0x08082b2b,
676
+ 0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x08190808, 0x08190808, 0x082b0819, 0x08190808,
677
+ 0x19080808, 0x08190808, 0x192b0808, 0x08190808, 0x2b081908, 0x08190808, 0x2b190808, 0x08190808,
678
+ 0x2b191919, 0x08190808, 0x08080808, 0x08190819, 0x08082b08, 0x08190819, 0x082b0808, 0x08190819,
679
+ 0x19190808, 0x08190819, 0x19192b2b, 0x08190819, 0x2b080808, 0x08190819, 0x082b1908, 0x0819082b,
680
+ 0x19081919, 0x0819082b, 0x08080808, 0x08191908, 0x08082b08, 0x08191908, 0x082b0808, 0x08191908,
681
+ 0x082b1919, 0x08191908, 0x19082b19, 0x08191908, 0x2b080808, 0x08191908, 0x08192b08, 0x08191919,
682
+ 0x192b082b, 0x08191919, 0x08080808, 0x0819192b, 0x0819192b, 0x0819192b, 0x08080819, 0x08192b08,
683
+ 0x08081908, 0x08192b08, 0x08190808, 0x08192b08, 0x19080808, 0x08192b08, 0x2b080819, 0x08192b08,
684
+ 0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x2b2b0808, 0x08192b19, 0x19190819, 0x08192b2b,
685
+ 0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08082b2b, 0x082b0808, 0x19081908, 0x082b0808,
686
+ 0x192b0819, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b08082b, 0x082b0808, 0x082b2b19, 0x082b0819,
687
+ 0x19082b08, 0x082b0819, 0x08080808, 0x082b082b, 0x0808082b, 0x082b082b, 0x08080819, 0x082b1908,
688
+ 0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x19080808, 0x082b1908, 0x1919192b, 0x082b1908,
689
+ 0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x192b1908, 0x082b1919, 0x2b190808, 0x082b192b,
690
+ 0x08082b08, 0x082b2b08, 0x082b0808, 0x082b2b08, 0x2b191908, 0x082b2b08, 0x19081908, 0x082b2b2b,
691
+ 0x08080819, 0x19080808, 0x08081908, 0x19080808, 0x08190808, 0x19080808, 0x08192b08, 0x19080808,
692
+ 0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x19080808, 0x19080808, 0x19082b08, 0x19080808,
693
+ 0x1919192b, 0x19080808, 0x192b0808, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808,
694
+ 0x2b190808, 0x19080808, 0x08080808, 0x19080819, 0x082b0808, 0x19080819, 0x192b0819, 0x19080819,
695
+ 0x2b080808, 0x19080819, 0x2b081919, 0x19080819, 0x08080819, 0x1908082b, 0x08190808, 0x1908082b,
696
+ 0x19082b08, 0x1908082b, 0x1919192b, 0x1908082b, 0x192b2b08, 0x1908082b, 0x08080808, 0x19081908,
697
+ 0x08082b08, 0x19081908, 0x082b0808, 0x19081908, 0x2b080808, 0x19081908, 0x2b192b19, 0x19081908,
698
+ 0x0819082b, 0x19081919, 0x082b1908, 0x19081919, 0x08080808, 0x1908192b, 0x08080819, 0x19082b08,
699
+ 0x08081908, 0x19082b08, 0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08,
700
+ 0x08080808, 0x19082b19, 0x19192b08, 0x19082b19, 0x192b0819, 0x19082b19, 0x2b08082b, 0x19082b19,
701
+ 0x19081919, 0x19082b2b, 0x2b190808, 0x19082b2b, 0x08080808, 0x19190808, 0x08082b08, 0x19190808,
702
+ 0x08190819, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x2b080808, 0x19190808,
703
+ 0x2b082b08, 0x19190808, 0x08081908, 0x19190819, 0x1908082b, 0x19190819, 0x2b2b1908, 0x19190819,
704
+ 0x2b190819, 0x1919082b, 0x2b190808, 0x19191908, 0x2b19082b, 0x19191908, 0x08082b2b, 0x19191919,
705
+ 0x08080819, 0x1919192b, 0x19191908, 0x1919192b, 0x08080808, 0x19192b08, 0x08190819, 0x19192b08,
706
+ 0x08192b19, 0x19192b08, 0x192b1908, 0x19192b08, 0x19080808, 0x19192b19, 0x08082b08, 0x19192b2b,
707
+ 0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808, 0x192b2b08, 0x192b0808,
708
+ 0x08080808, 0x192b0819, 0x19191919, 0x192b0819, 0x08192b08, 0x192b082b, 0x192b0808, 0x192b082b,
709
+ 0x08080808, 0x192b1908, 0x08081919, 0x192b1908, 0x08190808, 0x192b1919, 0x0819082b, 0x192b1919,
710
+ 0x2b081908, 0x192b1919, 0x1908082b, 0x192b2b08, 0x08080808, 0x2b080808, 0x0808082b, 0x2b080808,
711
+ 0x08082b2b, 0x2b080808, 0x19080819, 0x2b080808, 0x2b08082b, 0x2b080808, 0x08081908, 0x2b080819,
712
+ 0x08192b08, 0x2b080819, 0x19080808, 0x2b080819, 0x08190819, 0x2b08082b, 0x08080819, 0x2b081908,
713
+ 0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x08191919, 0x2b081908, 0x19080808, 0x2b081908,
714
+ 0x192b0808, 0x2b081908, 0x08080808, 0x2b081919, 0x1908192b, 0x2b081919, 0x2b191908, 0x2b081919,
715
+ 0x08082b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x192b0808, 0x2b08192b, 0x0808082b, 0x2b082b08,
716
+ 0x08081908, 0x2b082b19, 0x08190819, 0x2b082b2b, 0x08081908, 0x2b190808, 0x08190808, 0x2b190808,
717
+ 0x082b1908, 0x2b190808, 0x19080808, 0x2b190808, 0x2b2b0819, 0x2b190808, 0x0819192b, 0x2b190819,
718
+ 0x2b080808, 0x2b190819, 0x19081919, 0x2b19082b, 0x08080808, 0x2b191908, 0x082b082b, 0x2b191908,
719
+ 0x19081908, 0x2b191908, 0x19190819, 0x2b191919, 0x2b080819, 0x2b192b08, 0x082b0808, 0x2b192b19,
720
+ 0x0808082b, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b081919, 0x2b2b0808, 0x08082b19, 0x2b2b0819,
721
+ 0x08080808, 0x2b2b082b, 0x08192b08, 0x2b2b1908, 0x19190808, 0x2b2b2b08, 0x08081908, 0x2b2b2b19
722
+ );
723
+
724
+ struct iq2_xxs {
725
+ d: f16,
726
+ qs: array<f16, 32>
727
+ };
728
+
729
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
730
+ let block = src0[src0_idx_base + offset];
731
+ let d = f32(block.d);
732
+ var src1_i = src1_idx_base + offset * 256;
733
+ var sum = 0.0;
734
+ for (var ib: u32 = 0; ib < 32; ib += 4) {
735
+ let aux0 = bitcast<u32>(vec2(block.qs[ib], block.qs[ib + 1]));
736
+ let aux1 = bitcast<u32>(vec2(block.qs[ib + 2], block.qs[ib + 3]));
737
+ let db = d * (0.5 + f32(aux1 >> 28)) * 0.25;
738
+ for (var l: u32 = 0; l < 4; l++) {
739
+ let ig = get_byte(aux0, l) * 8;
740
+ let is = (aux1 >> (7 * l)) & 127;
741
+ let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
742
+ for (var j: u32 = 0; j < 8; j++) {
743
+ let g = get_byte(iq2xxs_grid[(ig + j) / 4], (ig + j) % 4);
744
+ let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
745
+ sum += db * f32(g) * m * src1[src1_i];
746
+ src1_i++;
747
+ }
748
+ }
749
+ }
750
+ return sum;
751
+ }
752
+
753
+ #enddecl(IQ2_XXS)
754
+
755
+ #decl(IQ2_XS)
756
+ const iq2xs_grid = array<u32, 1024>(
757
+ 0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
758
+ 0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
759
+ 0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
760
+ 0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
761
+ 0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
762
+ 0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x2b080808, 0x08080808,
763
+ 0x2b08082b, 0x08080808, 0x2b081919, 0x08080808, 0x2b082b08, 0x08080808, 0x2b190819, 0x08080808,
764
+ 0x2b191908, 0x08080808, 0x2b192b19, 0x08080808, 0x2b2b0808, 0x08080808, 0x08080819, 0x08080819,
765
+ 0x08081908, 0x08080819, 0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819,
766
+ 0x0819082b, 0x08080819, 0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x08192b2b, 0x08080819,
767
+ 0x082b0819, 0x08080819, 0x082b1908, 0x08080819, 0x19080808, 0x08080819, 0x1908082b, 0x08080819,
768
+ 0x19081919, 0x08080819, 0x19082b08, 0x08080819, 0x19190819, 0x08080819, 0x19191908, 0x08080819,
769
+ 0x192b0808, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819, 0x2b081908, 0x08080819,
770
+ 0x2b190808, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x08081919, 0x0808082b,
771
+ 0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b, 0x082b0808, 0x0808082b,
772
+ 0x19080819, 0x0808082b, 0x19081908, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
773
+ 0x2b080808, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908,
774
+ 0x0808192b, 0x08081908, 0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908,
775
+ 0x08191919, 0x08081908, 0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908,
776
+ 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908, 0x19082b08, 0x08081908,
777
+ 0x19190819, 0x08081908, 0x19191908, 0x08081908, 0x1919192b, 0x08081908, 0x192b0808, 0x08081908,
778
+ 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908, 0x08080808, 0x08081919,
779
+ 0x0808082b, 0x08081919, 0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08190819, 0x08081919,
780
+ 0x08191908, 0x08081919, 0x082b0808, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
781
+ 0x19190808, 0x08081919, 0x192b0819, 0x08081919, 0x2b080808, 0x08081919, 0x08080819, 0x0808192b,
782
+ 0x08081908, 0x0808192b, 0x08190808, 0x0808192b, 0x082b192b, 0x0808192b, 0x19080808, 0x0808192b,
783
+ 0x1908082b, 0x0808192b, 0x2b081908, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
784
+ 0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08082b2b, 0x08082b08, 0x08190819, 0x08082b08,
785
+ 0x08191908, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08, 0x19080819, 0x08082b08,
786
+ 0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x19192b08, 0x08082b08, 0x2b080808, 0x08082b08,
787
+ 0x2b2b0808, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19, 0x08081908, 0x08082b19,
788
+ 0x08190808, 0x08082b19, 0x19080808, 0x08082b19, 0x2b080819, 0x08082b19, 0x2b082b19, 0x08082b19,
789
+ 0x08080808, 0x08082b2b, 0x082b0808, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x2b19192b, 0x08082b2b,
790
+ 0x2b2b0808, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x0808192b, 0x08190808,
791
+ 0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808, 0x08191919, 0x08190808,
792
+ 0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808, 0x19080808, 0x08190808,
793
+ 0x1908082b, 0x08190808, 0x19081919, 0x08190808, 0x19082b08, 0x08190808, 0x19190819, 0x08190808,
794
+ 0x19191908, 0x08190808, 0x192b0808, 0x08190808, 0x192b2b2b, 0x08190808, 0x2b080819, 0x08190808,
795
+ 0x2b081908, 0x08190808, 0x2b190808, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819,
796
+ 0x08081919, 0x08190819, 0x08082b08, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
797
+ 0x082b0808, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819, 0x19190808, 0x08190819,
798
+ 0x2b080808, 0x08190819, 0x2b191908, 0x08190819, 0x2b19192b, 0x08190819, 0x08080819, 0x0819082b,
799
+ 0x08081908, 0x0819082b, 0x0808192b, 0x0819082b, 0x08190808, 0x0819082b, 0x19080808, 0x0819082b,
800
+ 0x192b0808, 0x0819082b, 0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908,
801
+ 0x08082b08, 0x08191908, 0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x082b0808, 0x08191908,
802
+ 0x19080819, 0x08191908, 0x19081908, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
803
+ 0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x08080819, 0x08191919, 0x08081908, 0x08191919,
804
+ 0x08190808, 0x08191919, 0x19080808, 0x08191919, 0x08080808, 0x0819192b, 0x08191908, 0x0819192b,
805
+ 0x19082b19, 0x0819192b, 0x08080819, 0x08192b08, 0x08081908, 0x08192b08, 0x08190808, 0x08192b08,
806
+ 0x0819082b, 0x08192b08, 0x19080808, 0x08192b08, 0x19191908, 0x08192b08, 0x2b08192b, 0x08192b08,
807
+ 0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x192b192b, 0x08192b19, 0x19190819, 0x08192b2b,
808
+ 0x2b2b2b19, 0x08192b2b, 0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808,
809
+ 0x08082b08, 0x082b0808, 0x08082b2b, 0x082b0808, 0x08190819, 0x082b0808, 0x08191908, 0x082b0808,
810
+ 0x082b0808, 0x082b0808, 0x19080819, 0x082b0808, 0x19081908, 0x082b0808, 0x19190808, 0x082b0808,
811
+ 0x2b080808, 0x082b0808, 0x2b2b0808, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819,
812
+ 0x08190808, 0x082b0819, 0x19080808, 0x082b0819, 0x19082b08, 0x082b0819, 0x192b1919, 0x082b0819,
813
+ 0x08080808, 0x082b082b, 0x082b082b, 0x082b082b, 0x2b080808, 0x082b082b, 0x2b2b2b08, 0x082b082b,
814
+ 0x08080819, 0x082b1908, 0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x082b2b19, 0x082b1908,
815
+ 0x19080808, 0x082b1908, 0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x1919082b, 0x082b1919,
816
+ 0x2b192b19, 0x082b1919, 0x08080819, 0x082b192b, 0x08192b2b, 0x082b192b, 0x2b2b192b, 0x082b192b,
817
+ 0x08080808, 0x082b2b08, 0x08082b08, 0x082b2b08, 0x08082b2b, 0x082b2b08, 0x082b0808, 0x082b2b08,
818
+ 0x19191919, 0x082b2b08, 0x2b082b08, 0x082b2b08, 0x2b2b082b, 0x082b2b08, 0x192b2b08, 0x082b2b19,
819
+ 0x2b190808, 0x082b2b19, 0x08082b08, 0x082b2b2b, 0x082b0808, 0x082b2b2b, 0x2b08082b, 0x082b2b2b,
820
+ 0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808, 0x08081908, 0x19080808,
821
+ 0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808, 0x0819082b, 0x19080808,
822
+ 0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x082b0819, 0x19080808, 0x082b1908, 0x19080808,
823
+ 0x19080808, 0x19080808, 0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808,
824
+ 0x19082b2b, 0x19080808, 0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x192b0808, 0x19080808,
825
+ 0x192b1919, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808, 0x2b190808, 0x19080808,
826
+ 0x08080808, 0x19080819, 0x0808082b, 0x19080819, 0x08081919, 0x19080819, 0x08082b08, 0x19080819,
827
+ 0x08190819, 0x19080819, 0x08191908, 0x19080819, 0x082b0808, 0x19080819, 0x19080819, 0x19080819,
828
+ 0x19081908, 0x19080819, 0x19190808, 0x19080819, 0x2b080808, 0x19080819, 0x2b081919, 0x19080819,
829
+ 0x2b2b082b, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b, 0x08190808, 0x1908082b,
830
+ 0x0819082b, 0x1908082b, 0x082b2b19, 0x1908082b, 0x19080808, 0x1908082b, 0x08080808, 0x19081908,
831
+ 0x0808082b, 0x19081908, 0x08081919, 0x19081908, 0x08082b08, 0x19081908, 0x08190819, 0x19081908,
832
+ 0x08191908, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x19080819, 0x19081908,
833
+ 0x19081908, 0x19081908, 0x19190808, 0x19081908, 0x2b080808, 0x19081908, 0x2b191908, 0x19081908,
834
+ 0x08080819, 0x19081919, 0x08081908, 0x19081919, 0x08190808, 0x19081919, 0x082b1908, 0x19081919,
835
+ 0x19080808, 0x19081919, 0x2b192b2b, 0x19081919, 0x08080808, 0x1908192b, 0x08082b2b, 0x1908192b,
836
+ 0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x08080819, 0x19082b08, 0x08081908, 0x19082b08,
837
+ 0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08, 0x19191908, 0x19082b08,
838
+ 0x192b082b, 0x19082b08, 0x08080808, 0x19082b19, 0x08190819, 0x19082b19, 0x19081908, 0x19082b19,
839
+ 0x19190808, 0x19082b19, 0x192b2b19, 0x19082b19, 0x08081908, 0x19082b2b, 0x08080808, 0x19190808,
840
+ 0x0808082b, 0x19190808, 0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808,
841
+ 0x08191908, 0x19190808, 0x082b0808, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808,
842
+ 0x19081908, 0x19190808, 0x19190808, 0x19190808, 0x2b080808, 0x19190808, 0x08080819, 0x19190819,
843
+ 0x08081908, 0x19190819, 0x08190808, 0x19190819, 0x08191919, 0x19190819, 0x19080808, 0x19190819,
844
+ 0x1908082b, 0x19190819, 0x08080808, 0x1919082b, 0x19081908, 0x1919082b, 0x2b2b2b2b, 0x1919082b,
845
+ 0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x08190808, 0x19191908, 0x082b0819, 0x19191908,
846
+ 0x19080808, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b2b0819, 0x19191908,
847
+ 0x08080808, 0x19191919, 0x08082b08, 0x19191919, 0x2b080808, 0x19191919, 0x2b082b08, 0x19191919,
848
+ 0x082b0819, 0x1919192b, 0x192b2b08, 0x1919192b, 0x2b2b0819, 0x1919192b, 0x08080808, 0x19192b08,
849
+ 0x08191908, 0x19192b08, 0x19080819, 0x19192b08, 0x19190808, 0x19192b08, 0x2b192b19, 0x19192b08,
850
+ 0x08192b2b, 0x19192b19, 0x19080808, 0x19192b19, 0x1908082b, 0x19192b19, 0x2b081919, 0x19192b2b,
851
+ 0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808,
852
+ 0x19191908, 0x192b0808, 0x192b082b, 0x192b0808, 0x2b08192b, 0x192b0808, 0x2b2b2b19, 0x192b0808,
853
+ 0x08080808, 0x192b0819, 0x082b1908, 0x192b082b, 0x19082b2b, 0x192b082b, 0x2b19082b, 0x192b082b,
854
+ 0x08080808, 0x192b1908, 0x0819192b, 0x192b1908, 0x08190808, 0x192b1919, 0x19080808, 0x192b1919,
855
+ 0x19081919, 0x192b1919, 0x2b2b1908, 0x192b1919, 0x08080819, 0x192b2b08, 0x192b2b2b, 0x192b2b08,
856
+ 0x082b1919, 0x192b2b19, 0x0808192b, 0x192b2b2b, 0x19191908, 0x192b2b2b, 0x192b082b, 0x192b2b2b,
857
+ 0x08080808, 0x2b080808, 0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808,
858
+ 0x08190819, 0x2b080808, 0x08191908, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b2b2b, 0x2b080808,
859
+ 0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x2b080808, 0x2b080808,
860
+ 0x2b08082b, 0x2b080808, 0x2b2b2b08, 0x2b080808, 0x2b2b2b2b, 0x2b080808, 0x08080819, 0x2b080819,
861
+ 0x08081908, 0x2b080819, 0x0808192b, 0x2b080819, 0x08190808, 0x2b080819, 0x19080808, 0x2b080819,
862
+ 0x19190819, 0x2b080819, 0x19192b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x082b0808, 0x2b08082b,
863
+ 0x2b080808, 0x2b08082b, 0x2b08082b, 0x2b08082b, 0x2b2b0808, 0x2b08082b, 0x2b2b2b08, 0x2b08082b,
864
+ 0x08080819, 0x2b081908, 0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
865
+ 0x08191919, 0x2b081908, 0x19080808, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b082b19, 0x2b081908,
866
+ 0x08080808, 0x2b081919, 0x19081908, 0x2b081919, 0x2b2b1919, 0x2b081919, 0x08192b08, 0x2b08192b,
867
+ 0x192b2b2b, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08082b08, 0x2b082b08, 0x082b1919, 0x2b082b08,
868
+ 0x19192b2b, 0x2b082b08, 0x2b080808, 0x2b082b08, 0x2b08082b, 0x2b082b08, 0x2b2b2b08, 0x2b082b08,
869
+ 0x0808192b, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x2b080808, 0x2b082b2b, 0x2b082b08, 0x2b082b2b,
870
+ 0x2b19192b, 0x2b082b2b, 0x2b2b2b08, 0x2b082b2b, 0x08080819, 0x2b190808, 0x08081908, 0x2b190808,
871
+ 0x08190808, 0x2b190808, 0x19080808, 0x2b190808, 0x1919192b, 0x2b190808, 0x2b081908, 0x2b190808,
872
+ 0x08080808, 0x2b190819, 0x082b082b, 0x2b190819, 0x192b1908, 0x2b190819, 0x1919192b, 0x2b19082b,
873
+ 0x2b082b19, 0x2b19082b, 0x08080808, 0x2b191908, 0x08081919, 0x2b191908, 0x19081908, 0x2b191908,
874
+ 0x19190808, 0x2b191908, 0x19192b08, 0x2b191908, 0x082b2b19, 0x2b191919, 0x2b190808, 0x2b191919,
875
+ 0x2b19082b, 0x2b191919, 0x19080819, 0x2b19192b, 0x19190819, 0x2b192b08, 0x2b2b192b, 0x2b192b08,
876
+ 0x19082b19, 0x2b192b19, 0x08191919, 0x2b192b2b, 0x192b0808, 0x2b192b2b, 0x08080808, 0x2b2b0808,
877
+ 0x0808082b, 0x2b2b0808, 0x08082b08, 0x2b2b0808, 0x08082b2b, 0x2b2b0808, 0x082b0808, 0x2b2b0808,
878
+ 0x082b2b2b, 0x2b2b0808, 0x2b2b0808, 0x2b2b0808, 0x19190819, 0x2b2b0819, 0x19192b19, 0x2b2b0819,
879
+ 0x2b2b192b, 0x2b2b0819, 0x08080808, 0x2b2b082b, 0x0808082b, 0x2b2b082b, 0x08082b08, 0x2b2b082b,
880
+ 0x082b2b2b, 0x2b2b082b, 0x2b080808, 0x2b2b082b, 0x2b2b0808, 0x2b2b082b, 0x19080808, 0x2b2b1908,
881
+ 0x2b191919, 0x2b2b1908, 0x192b1919, 0x2b2b192b, 0x2b192b08, 0x2b2b192b, 0x08082b2b, 0x2b2b2b08,
882
+ 0x082b0808, 0x2b2b2b08, 0x082b082b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b0808, 0x2b2b2b08,
883
+ 0x2b2b2b08, 0x2b2b2b08, 0x08081908, 0x2b2b2b19, 0x2b081908, 0x2b2b2b19, 0x2b08192b, 0x2b2b2b19,
884
+ 0x082b2b08, 0x2b2b2b2b, 0x082b2b2b, 0x2b2b2b2b, 0x2b190819, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
885
+ );
886
+
887
+ struct iq2_xs {
888
+ d: f16,
889
+ qs: array<f16, 32>,
890
+ scales: array<f16, 4>
891
+ };
892
+
893
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
894
+ let block = src0[src0_idx_base + offset];
895
+ let d = f32(block.d);
896
+ var src1_i = src1_idx_base + offset * 256;
897
+ var scale_vals = array<u32, 2>(
898
+ bitcast<u32>(vec2(block.scales[0], block.scales[1])),
899
+ bitcast<u32>(vec2(block.scales[2], block.scales[3]))
900
+ );
901
+ var sum = 0.0;
902
+ for (var ib: u32 = 0; ib < 32; ib += 4) {
903
+ let s = get_byte(scale_vals[ib / 16], (ib % 16) / 4);
904
+ let db = array<f32, 2>(
905
+ d * (0.5 + f32(s & 0xF)) * 0.25,
906
+ d * (0.5 + f32(s >> 4)) * 0.25
907
+ );
908
+ for (var l: u32 = 0; l < 4; l++) {
909
+ let qs_val = bitcast<u32>(vec2(block.qs[ib + l], 0.0));
910
+ let ig = (qs_val & 511) * 8;
911
+ let is = qs_val >> 9;
912
+ let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
913
+ let dl = db[l/2];
914
+ for (var j: u32 = 0; j < 8; j++) {
915
+ let g = get_byte(iq2xs_grid[(ig + j) / 4], (ig + j) % 4);
916
+ let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
917
+ sum += dl * f32(g) * m * src1[src1_i];
918
+ src1_i++;
919
+ }
920
+ }
921
+ }
922
+ return sum;
923
+ }
924
+
925
+ #enddecl(IQ2_XS)
926
+
927
+ #decl(IQ2_S)
928
+
929
+ const iq2s_grid = array<u32, 2048>(
930
+ 0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
931
+ 0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
932
+ 0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
933
+ 0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
934
+ 0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
935
+ 0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x192b192b, 0x08080808,
936
+ 0x192b2b19, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b081919, 0x08080808,
937
+ 0x2b082b08, 0x08080808, 0x2b190819, 0x08080808, 0x2b191908, 0x08080808, 0x2b2b0808, 0x08080808,
938
+ 0x2b2b1919, 0x08080808, 0x2b2b2b2b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819,
939
+ 0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819, 0x0819082b, 0x08080819,
940
+ 0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x082b0819, 0x08080819, 0x082b1908, 0x08080819,
941
+ 0x19080808, 0x08080819, 0x1908082b, 0x08080819, 0x19081919, 0x08080819, 0x19082b08, 0x08080819,
942
+ 0x19190819, 0x08080819, 0x19191908, 0x08080819, 0x1919192b, 0x08080819, 0x19192b19, 0x08080819,
943
+ 0x192b0808, 0x08080819, 0x192b1919, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819,
944
+ 0x2b081908, 0x08080819, 0x2b190808, 0x08080819, 0x2b19082b, 0x08080819, 0x2b191919, 0x08080819,
945
+ 0x2b2b0819, 0x08080819, 0x2b2b1908, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b,
946
+ 0x08081919, 0x0808082b, 0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b,
947
+ 0x082b0808, 0x0808082b, 0x082b2b2b, 0x0808082b, 0x19080819, 0x0808082b, 0x19081908, 0x0808082b,
948
+ 0x1908192b, 0x0808082b, 0x19082b19, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
949
+ 0x2b080808, 0x0808082b, 0x2b081919, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x2b191908, 0x0808082b,
950
+ 0x2b2b082b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x0808192b, 0x08081908,
951
+ 0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908, 0x08191919, 0x08081908,
952
+ 0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908, 0x082b192b, 0x08081908,
953
+ 0x082b2b19, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908,
954
+ 0x19082b08, 0x08081908, 0x19082b2b, 0x08081908, 0x19190819, 0x08081908, 0x19191908, 0x08081908,
955
+ 0x1919192b, 0x08081908, 0x19192b19, 0x08081908, 0x192b0808, 0x08081908, 0x192b082b, 0x08081908,
956
+ 0x192b1919, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b08192b, 0x08081908,
957
+ 0x2b082b19, 0x08081908, 0x2b190808, 0x08081908, 0x2b191919, 0x08081908, 0x2b192b08, 0x08081908,
958
+ 0x2b2b0819, 0x08081908, 0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919,
959
+ 0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08082b2b, 0x08081919, 0x08190819, 0x08081919,
960
+ 0x08191908, 0x08081919, 0x0819192b, 0x08081919, 0x08192b19, 0x08081919, 0x082b0808, 0x08081919,
961
+ 0x082b1919, 0x08081919, 0x082b2b08, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
962
+ 0x1908192b, 0x08081919, 0x19082b19, 0x08081919, 0x19190808, 0x08081919, 0x1919082b, 0x08081919,
963
+ 0x19191919, 0x08081919, 0x19192b08, 0x08081919, 0x192b0819, 0x08081919, 0x192b1908, 0x08081919,
964
+ 0x2b080808, 0x08081919, 0x2b08082b, 0x08081919, 0x2b081919, 0x08081919, 0x2b082b08, 0x08081919,
965
+ 0x2b190819, 0x08081919, 0x2b191908, 0x08081919, 0x2b2b0808, 0x08081919, 0x08080819, 0x0808192b,
966
+ 0x08081908, 0x0808192b, 0x0808192b, 0x0808192b, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b,
967
+ 0x08191919, 0x0808192b, 0x19080808, 0x0808192b, 0x19081919, 0x0808192b, 0x19082b08, 0x0808192b,
968
+ 0x19190819, 0x0808192b, 0x19191908, 0x0808192b, 0x192b0808, 0x0808192b, 0x2b080819, 0x0808192b,
969
+ 0x2b081908, 0x0808192b, 0x2b190808, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
970
+ 0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08190819, 0x08082b08, 0x08191908, 0x08082b08,
971
+ 0x0819192b, 0x08082b08, 0x08192b19, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08,
972
+ 0x082b2b2b, 0x08082b08, 0x19080819, 0x08082b08, 0x19081908, 0x08082b08, 0x1908192b, 0x08082b08,
973
+ 0x19082b19, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x19191919, 0x08082b08,
974
+ 0x19192b08, 0x08082b08, 0x192b0819, 0x08082b08, 0x192b1908, 0x08082b08, 0x2b080808, 0x08082b08,
975
+ 0x2b081919, 0x08082b08, 0x2b191908, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19,
976
+ 0x08081908, 0x08082b19, 0x08190808, 0x08082b19, 0x0819082b, 0x08082b19, 0x08191919, 0x08082b19,
977
+ 0x08192b08, 0x08082b19, 0x082b0819, 0x08082b19, 0x19080808, 0x08082b19, 0x19081919, 0x08082b19,
978
+ 0x19082b08, 0x08082b19, 0x19190819, 0x08082b19, 0x19191908, 0x08082b19, 0x192b0808, 0x08082b19,
979
+ 0x2b080819, 0x08082b19, 0x2b190808, 0x08082b19, 0x08080808, 0x08082b2b, 0x08190819, 0x08082b2b,
980
+ 0x08191908, 0x08082b2b, 0x082b082b, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x082b2b2b, 0x08082b2b,
981
+ 0x19190808, 0x08082b2b, 0x2b192b19, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808,
982
+ 0x0808192b, 0x08190808, 0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808,
983
+ 0x08191919, 0x08190808, 0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808,
984
+ 0x082b192b, 0x08190808, 0x19080808, 0x08190808, 0x1908082b, 0x08190808, 0x19081919, 0x08190808,
985
+ 0x19082b08, 0x08190808, 0x19190819, 0x08190808, 0x19191908, 0x08190808, 0x1919192b, 0x08190808,
986
+ 0x19192b19, 0x08190808, 0x192b0808, 0x08190808, 0x192b082b, 0x08190808, 0x192b1919, 0x08190808,
987
+ 0x192b2b08, 0x08190808, 0x2b080819, 0x08190808, 0x2b081908, 0x08190808, 0x2b08192b, 0x08190808,
988
+ 0x2b190808, 0x08190808, 0x2b191919, 0x08190808, 0x2b192b08, 0x08190808, 0x2b2b0819, 0x08190808,
989
+ 0x2b2b1908, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819, 0x08081919, 0x08190819,
990
+ 0x08082b08, 0x08190819, 0x08082b2b, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
991
+ 0x0819192b, 0x08190819, 0x08192b19, 0x08190819, 0x082b0808, 0x08190819, 0x082b082b, 0x08190819,
992
+ 0x082b1919, 0x08190819, 0x082b2b08, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819,
993
+ 0x1908192b, 0x08190819, 0x19082b19, 0x08190819, 0x19190808, 0x08190819, 0x1919082b, 0x08190819,
994
+ 0x19191919, 0x08190819, 0x19192b08, 0x08190819, 0x192b0819, 0x08190819, 0x192b1908, 0x08190819,
995
+ 0x2b080808, 0x08190819, 0x2b08082b, 0x08190819, 0x2b081919, 0x08190819, 0x2b082b08, 0x08190819,
996
+ 0x2b190819, 0x08190819, 0x2b191908, 0x08190819, 0x08080819, 0x0819082b, 0x08081908, 0x0819082b,
997
+ 0x08082b19, 0x0819082b, 0x08190808, 0x0819082b, 0x08191919, 0x0819082b, 0x082b0819, 0x0819082b,
998
+ 0x082b1908, 0x0819082b, 0x19080808, 0x0819082b, 0x19081919, 0x0819082b, 0x19190819, 0x0819082b,
999
+ 0x19191908, 0x0819082b, 0x2b080819, 0x0819082b, 0x2b081908, 0x0819082b, 0x2b190808, 0x0819082b,
1000
+ 0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908, 0x08082b08, 0x08191908,
1001
+ 0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x0819192b, 0x08191908, 0x08192b19, 0x08191908,
1002
+ 0x082b0808, 0x08191908, 0x082b1919, 0x08191908, 0x082b2b08, 0x08191908, 0x19080819, 0x08191908,
1003
+ 0x19081908, 0x08191908, 0x1908192b, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
1004
+ 0x1919082b, 0x08191908, 0x19191919, 0x08191908, 0x19192b08, 0x08191908, 0x192b0819, 0x08191908,
1005
+ 0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x2b08082b, 0x08191908, 0x2b081919, 0x08191908,
1006
+ 0x2b082b08, 0x08191908, 0x2b190819, 0x08191908, 0x2b191908, 0x08191908, 0x2b2b0808, 0x08191908,
1007
+ 0x08080819, 0x08191919, 0x08081908, 0x08191919, 0x0808192b, 0x08191919, 0x08082b19, 0x08191919,
1008
+ 0x08190808, 0x08191919, 0x0819082b, 0x08191919, 0x08191919, 0x08191919, 0x08192b08, 0x08191919,
1009
+ 0x082b0819, 0x08191919, 0x082b1908, 0x08191919, 0x19080808, 0x08191919, 0x1908082b, 0x08191919,
1010
+ 0x19081919, 0x08191919, 0x19082b08, 0x08191919, 0x19190819, 0x08191919, 0x19191908, 0x08191919,
1011
+ 0x192b0808, 0x08191919, 0x2b080819, 0x08191919, 0x2b081908, 0x08191919, 0x2b190808, 0x08191919,
1012
+ 0x08080808, 0x0819192b, 0x08081919, 0x0819192b, 0x08082b08, 0x0819192b, 0x08190819, 0x0819192b,
1013
+ 0x08191908, 0x0819192b, 0x082b0808, 0x0819192b, 0x19080819, 0x0819192b, 0x19081908, 0x0819192b,
1014
+ 0x19190808, 0x0819192b, 0x2b080808, 0x0819192b, 0x2b2b2b2b, 0x0819192b, 0x08080819, 0x08192b08,
1015
+ 0x08081908, 0x08192b08, 0x0808192b, 0x08192b08, 0x08082b19, 0x08192b08, 0x08190808, 0x08192b08,
1016
+ 0x08191919, 0x08192b08, 0x08192b08, 0x08192b08, 0x082b0819, 0x08192b08, 0x19080808, 0x08192b08,
1017
+ 0x1908082b, 0x08192b08, 0x19081919, 0x08192b08, 0x19082b08, 0x08192b08, 0x19190819, 0x08192b08,
1018
+ 0x19191908, 0x08192b08, 0x192b0808, 0x08192b08, 0x2b080819, 0x08192b08, 0x2b081908, 0x08192b08,
1019
+ 0x08080808, 0x08192b19, 0x0808082b, 0x08192b19, 0x08081919, 0x08192b19, 0x08082b08, 0x08192b19,
1020
+ 0x08190819, 0x08192b19, 0x08191908, 0x08192b19, 0x082b0808, 0x08192b19, 0x19080819, 0x08192b19,
1021
+ 0x19081908, 0x08192b19, 0x19190808, 0x08192b19, 0x192b2b19, 0x08192b19, 0x2b2b082b, 0x08192b19,
1022
+ 0x08081908, 0x08192b2b, 0x08190808, 0x08192b2b, 0x19080808, 0x08192b2b, 0x1919192b, 0x08192b2b,
1023
+ 0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808, 0x08082b08, 0x082b0808,
1024
+ 0x08190819, 0x082b0808, 0x08191908, 0x082b0808, 0x0819192b, 0x082b0808, 0x08192b19, 0x082b0808,
1025
+ 0x082b0808, 0x082b0808, 0x082b1919, 0x082b0808, 0x082b2b2b, 0x082b0808, 0x19080819, 0x082b0808,
1026
+ 0x19081908, 0x082b0808, 0x19190808, 0x082b0808, 0x1919082b, 0x082b0808, 0x19191919, 0x082b0808,
1027
+ 0x192b1908, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b082b2b, 0x082b0808, 0x2b191908, 0x082b0808,
1028
+ 0x2b2b2b2b, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819, 0x08190808, 0x082b0819,
1029
+ 0x0819082b, 0x082b0819, 0x08191919, 0x082b0819, 0x082b0819, 0x082b0819, 0x19080808, 0x082b0819,
1030
+ 0x1908082b, 0x082b0819, 0x19081919, 0x082b0819, 0x19190819, 0x082b0819, 0x19191908, 0x082b0819,
1031
+ 0x192b0808, 0x082b0819, 0x2b080819, 0x082b0819, 0x2b081908, 0x082b0819, 0x2b190808, 0x082b0819,
1032
+ 0x08080808, 0x082b082b, 0x08082b2b, 0x082b082b, 0x082b082b, 0x082b082b, 0x082b2b08, 0x082b082b,
1033
+ 0x082b2b2b, 0x082b082b, 0x19081908, 0x082b082b, 0x19190808, 0x082b082b, 0x2b082b08, 0x082b082b,
1034
+ 0x2b082b2b, 0x082b082b, 0x2b2b2b08, 0x082b082b, 0x08080819, 0x082b1908, 0x08081908, 0x082b1908,
1035
+ 0x0808192b, 0x082b1908, 0x08082b19, 0x082b1908, 0x08190808, 0x082b1908, 0x08191919, 0x082b1908,
1036
+ 0x08192b08, 0x082b1908, 0x082b0819, 0x082b1908, 0x082b1908, 0x082b1908, 0x19080808, 0x082b1908,
1037
+ 0x1908082b, 0x082b1908, 0x19081919, 0x082b1908, 0x19082b08, 0x082b1908, 0x19190819, 0x082b1908,
1038
+ 0x19191908, 0x082b1908, 0x192b0808, 0x082b1908, 0x2b080819, 0x082b1908, 0x2b081908, 0x082b1908,
1039
+ 0x2b190808, 0x082b1908, 0x08080808, 0x082b1919, 0x08081919, 0x082b1919, 0x08082b08, 0x082b1919,
1040
+ 0x08190819, 0x082b1919, 0x08191908, 0x082b1919, 0x082b0808, 0x082b1919, 0x19080819, 0x082b1919,
1041
+ 0x19081908, 0x082b1919, 0x19190808, 0x082b1919, 0x192b192b, 0x082b1919, 0x2b080808, 0x082b1919,
1042
+ 0x08080819, 0x082b192b, 0x08081908, 0x082b192b, 0x08190808, 0x082b192b, 0x19080808, 0x082b192b,
1043
+ 0x19192b19, 0x082b192b, 0x08080808, 0x082b2b08, 0x08081919, 0x082b2b08, 0x08190819, 0x082b2b08,
1044
+ 0x08191908, 0x082b2b08, 0x19080819, 0x082b2b08, 0x19081908, 0x082b2b08, 0x19190808, 0x082b2b08,
1045
+ 0x2b082b2b, 0x082b2b08, 0x2b2b2b2b, 0x082b2b08, 0x08080819, 0x082b2b19, 0x08081908, 0x082b2b19,
1046
+ 0x08190808, 0x082b2b19, 0x2b191919, 0x082b2b19, 0x08082b2b, 0x082b2b2b, 0x082b082b, 0x082b2b2b,
1047
+ 0x192b1908, 0x082b2b2b, 0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808,
1048
+ 0x08081908, 0x19080808, 0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808,
1049
+ 0x0819082b, 0x19080808, 0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x08192b2b, 0x19080808,
1050
+ 0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x082b192b, 0x19080808, 0x19080808, 0x19080808,
1051
+ 0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808, 0x19082b2b, 0x19080808,
1052
+ 0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x1919192b, 0x19080808, 0x19192b19, 0x19080808,
1053
+ 0x192b0808, 0x19080808, 0x192b082b, 0x19080808, 0x192b1919, 0x19080808, 0x2b080819, 0x19080808,
1054
+ 0x2b081908, 0x19080808, 0x2b190808, 0x19080808, 0x2b191919, 0x19080808, 0x2b192b08, 0x19080808,
1055
+ 0x2b2b0819, 0x19080808, 0x2b2b1908, 0x19080808, 0x08080808, 0x19080819, 0x0808082b, 0x19080819,
1056
+ 0x08081919, 0x19080819, 0x08082b08, 0x19080819, 0x08190819, 0x19080819, 0x08191908, 0x19080819,
1057
+ 0x0819192b, 0x19080819, 0x08192b19, 0x19080819, 0x082b0808, 0x19080819, 0x082b082b, 0x19080819,
1058
+ 0x082b1919, 0x19080819, 0x19080819, 0x19080819, 0x19081908, 0x19080819, 0x1908192b, 0x19080819,
1059
+ 0x19082b19, 0x19080819, 0x19190808, 0x19080819, 0x1919082b, 0x19080819, 0x19191919, 0x19080819,
1060
+ 0x19192b08, 0x19080819, 0x192b0819, 0x19080819, 0x192b1908, 0x19080819, 0x2b080808, 0x19080819,
1061
+ 0x2b08082b, 0x19080819, 0x2b081919, 0x19080819, 0x2b082b08, 0x19080819, 0x2b190819, 0x19080819,
1062
+ 0x2b191908, 0x19080819, 0x2b2b0808, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b,
1063
+ 0x08190808, 0x1908082b, 0x0819082b, 0x1908082b, 0x08191919, 0x1908082b, 0x08192b08, 0x1908082b,
1064
+ 0x082b1908, 0x1908082b, 0x19080808, 0x1908082b, 0x19081919, 0x1908082b, 0x19082b08, 0x1908082b,
1065
+ 0x19190819, 0x1908082b, 0x19191908, 0x1908082b, 0x192b0808, 0x1908082b, 0x2b080819, 0x1908082b,
1066
+ 0x2b081908, 0x1908082b, 0x08080808, 0x19081908, 0x0808082b, 0x19081908, 0x08081919, 0x19081908,
1067
+ 0x08082b08, 0x19081908, 0x08082b2b, 0x19081908, 0x08190819, 0x19081908, 0x08191908, 0x19081908,
1068
+ 0x0819192b, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x082b082b, 0x19081908,
1069
+ 0x082b1919, 0x19081908, 0x082b2b08, 0x19081908, 0x19080819, 0x19081908, 0x19081908, 0x19081908,
1070
+ 0x1908192b, 0x19081908, 0x19082b19, 0x19081908, 0x19190808, 0x19081908, 0x1919082b, 0x19081908,
1071
+ 0x19191919, 0x19081908, 0x19192b08, 0x19081908, 0x192b0819, 0x19081908, 0x192b1908, 0x19081908,
1072
+ 0x2b080808, 0x19081908, 0x2b08082b, 0x19081908, 0x2b081919, 0x19081908, 0x2b082b08, 0x19081908,
1073
+ 0x2b190819, 0x19081908, 0x2b191908, 0x19081908, 0x2b2b0808, 0x19081908, 0x08080819, 0x19081919,
1074
+ 0x08081908, 0x19081919, 0x0808192b, 0x19081919, 0x08082b19, 0x19081919, 0x08190808, 0x19081919,
1075
+ 0x0819082b, 0x19081919, 0x08191919, 0x19081919, 0x08192b08, 0x19081919, 0x082b0819, 0x19081919,
1076
+ 0x082b1908, 0x19081919, 0x19080808, 0x19081919, 0x1908082b, 0x19081919, 0x19081919, 0x19081919,
1077
+ 0x19082b08, 0x19081919, 0x19190819, 0x19081919, 0x19191908, 0x19081919, 0x192b0808, 0x19081919,
1078
+ 0x192b2b2b, 0x19081919, 0x2b080819, 0x19081919, 0x2b081908, 0x19081919, 0x2b190808, 0x19081919,
1079
+ 0x08080808, 0x1908192b, 0x0808082b, 0x1908192b, 0x08081919, 0x1908192b, 0x08082b08, 0x1908192b,
1080
+ 0x08190819, 0x1908192b, 0x08191908, 0x1908192b, 0x082b0808, 0x1908192b, 0x19080819, 0x1908192b,
1081
+ 0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x2b080808, 0x1908192b, 0x2b2b1919, 0x1908192b,
1082
+ 0x08080819, 0x19082b08, 0x08081908, 0x19082b08, 0x08082b19, 0x19082b08, 0x08190808, 0x19082b08,
1083
+ 0x0819082b, 0x19082b08, 0x08191919, 0x19082b08, 0x08192b08, 0x19082b08, 0x082b0819, 0x19082b08,
1084
+ 0x082b1908, 0x19082b08, 0x19080808, 0x19082b08, 0x1908082b, 0x19082b08, 0x19081919, 0x19082b08,
1085
+ 0x19082b08, 0x19082b08, 0x19190819, 0x19082b08, 0x19191908, 0x19082b08, 0x192b0808, 0x19082b08,
1086
+ 0x2b081908, 0x19082b08, 0x2b190808, 0x19082b08, 0x08080808, 0x19082b19, 0x0808082b, 0x19082b19,
1087
+ 0x08081919, 0x19082b19, 0x08082b08, 0x19082b19, 0x08190819, 0x19082b19, 0x08191908, 0x19082b19,
1088
+ 0x082b0808, 0x19082b19, 0x19080819, 0x19082b19, 0x19081908, 0x19082b19, 0x19190808, 0x19082b19,
1089
+ 0x2b080808, 0x19082b19, 0x2b19192b, 0x19082b19, 0x08080819, 0x19082b2b, 0x08081908, 0x19082b2b,
1090
+ 0x08190808, 0x19082b2b, 0x19080808, 0x19082b2b, 0x08080808, 0x19190808, 0x0808082b, 0x19190808,
1091
+ 0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808, 0x08191908, 0x19190808,
1092
+ 0x0819192b, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x082b082b, 0x19190808,
1093
+ 0x082b1919, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808, 0x19081908, 0x19190808,
1094
+ 0x1908192b, 0x19190808, 0x19082b19, 0x19190808, 0x19190808, 0x19190808, 0x1919082b, 0x19190808,
1095
+ 0x19191919, 0x19190808, 0x19192b08, 0x19190808, 0x192b0819, 0x19190808, 0x192b1908, 0x19190808,
1096
+ 0x2b080808, 0x19190808, 0x2b08082b, 0x19190808, 0x2b081919, 0x19190808, 0x2b082b08, 0x19190808,
1097
+ 0x2b190819, 0x19190808, 0x2b191908, 0x19190808, 0x08080819, 0x19190819, 0x08081908, 0x19190819,
1098
+ 0x0808192b, 0x19190819, 0x08082b19, 0x19190819, 0x08190808, 0x19190819, 0x0819082b, 0x19190819,
1099
+ 0x08191919, 0x19190819, 0x08192b08, 0x19190819, 0x082b0819, 0x19190819, 0x082b1908, 0x19190819,
1100
+ 0x19080808, 0x19190819, 0x1908082b, 0x19190819, 0x19081919, 0x19190819, 0x19082b08, 0x19190819,
1101
+ 0x19190819, 0x19190819, 0x19191908, 0x19190819, 0x192b0808, 0x19190819, 0x2b080819, 0x19190819,
1102
+ 0x2b081908, 0x19190819, 0x2b190808, 0x19190819, 0x08080808, 0x1919082b, 0x08081919, 0x1919082b,
1103
+ 0x08082b08, 0x1919082b, 0x08190819, 0x1919082b, 0x08191908, 0x1919082b, 0x082b0808, 0x1919082b,
1104
+ 0x19080819, 0x1919082b, 0x19081908, 0x1919082b, 0x19190808, 0x1919082b, 0x192b2b19, 0x1919082b,
1105
+ 0x2b080808, 0x1919082b, 0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x0808192b, 0x19191908,
1106
+ 0x08082b19, 0x19191908, 0x08190808, 0x19191908, 0x0819082b, 0x19191908, 0x08191919, 0x19191908,
1107
+ 0x08192b08, 0x19191908, 0x082b0819, 0x19191908, 0x082b1908, 0x19191908, 0x19080808, 0x19191908,
1108
+ 0x1908082b, 0x19191908, 0x19081919, 0x19191908, 0x19082b08, 0x19191908, 0x19190819, 0x19191908,
1109
+ 0x19191908, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b081908, 0x19191908,
1110
+ 0x2b190808, 0x19191908, 0x08080808, 0x19191919, 0x0808082b, 0x19191919, 0x08081919, 0x19191919,
1111
+ 0x08082b08, 0x19191919, 0x08190819, 0x19191919, 0x08191908, 0x19191919, 0x082b0808, 0x19191919,
1112
+ 0x19080819, 0x19191919, 0x19081908, 0x19191919, 0x19190808, 0x19191919, 0x2b080808, 0x19191919,
1113
+ 0x08080819, 0x1919192b, 0x08081908, 0x1919192b, 0x08190808, 0x1919192b, 0x082b192b, 0x1919192b,
1114
+ 0x19080808, 0x1919192b, 0x08080808, 0x19192b08, 0x0808082b, 0x19192b08, 0x08081919, 0x19192b08,
1115
+ 0x08082b08, 0x19192b08, 0x08190819, 0x19192b08, 0x08191908, 0x19192b08, 0x082b0808, 0x19192b08,
1116
+ 0x19080819, 0x19192b08, 0x19081908, 0x19192b08, 0x19190808, 0x19192b08, 0x19192b2b, 0x19192b08,
1117
+ 0x2b080808, 0x19192b08, 0x08080819, 0x19192b19, 0x08081908, 0x19192b19, 0x08190808, 0x19192b19,
1118
+ 0x19080808, 0x19192b19, 0x08080808, 0x19192b2b, 0x08192b19, 0x19192b2b, 0x2b081919, 0x19192b2b,
1119
+ 0x2b2b2b08, 0x19192b2b, 0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x0808192b, 0x192b0808,
1120
+ 0x08190808, 0x192b0808, 0x0819082b, 0x192b0808, 0x08191919, 0x192b0808, 0x08192b08, 0x192b0808,
1121
+ 0x082b0819, 0x192b0808, 0x082b1908, 0x192b0808, 0x19080808, 0x192b0808, 0x19081919, 0x192b0808,
1122
+ 0x19082b08, 0x192b0808, 0x19190819, 0x192b0808, 0x19191908, 0x192b0808, 0x192b0808, 0x192b0808,
1123
+ 0x2b081908, 0x192b0808, 0x2b190808, 0x192b0808, 0x08080808, 0x192b0819, 0x0808082b, 0x192b0819,
1124
+ 0x08081919, 0x192b0819, 0x08082b08, 0x192b0819, 0x08190819, 0x192b0819, 0x08191908, 0x192b0819,
1125
+ 0x082b0808, 0x192b0819, 0x19080819, 0x192b0819, 0x19081908, 0x192b0819, 0x19190808, 0x192b0819,
1126
+ 0x2b080808, 0x192b0819, 0x2b192b19, 0x192b0819, 0x08081908, 0x192b082b, 0x08190808, 0x192b082b,
1127
+ 0x19080808, 0x192b082b, 0x1919192b, 0x192b082b, 0x2b2b0819, 0x192b082b, 0x08080808, 0x192b1908,
1128
+ 0x08081919, 0x192b1908, 0x08082b08, 0x192b1908, 0x08190819, 0x192b1908, 0x08191908, 0x192b1908,
1129
+ 0x082b0808, 0x192b1908, 0x19080819, 0x192b1908, 0x19081908, 0x192b1908, 0x19190808, 0x192b1908,
1130
+ 0x2b080808, 0x192b1908, 0x08080819, 0x192b1919, 0x08081908, 0x192b1919, 0x08190808, 0x192b1919,
1131
+ 0x19080808, 0x192b1919, 0x19082b2b, 0x192b1919, 0x192b2b08, 0x192b1919, 0x2b19082b, 0x192b1919,
1132
+ 0x08080808, 0x192b192b, 0x2b191908, 0x192b192b, 0x08080819, 0x192b2b08, 0x08081908, 0x192b2b08,
1133
+ 0x08190808, 0x192b2b08, 0x192b1919, 0x192b2b08, 0x2b192b08, 0x192b2b08, 0x08080808, 0x192b2b19,
1134
+ 0x082b2b2b, 0x192b2b19, 0x1908082b, 0x192b2b2b, 0x2b2b0819, 0x192b2b2b, 0x08080808, 0x2b080808,
1135
+ 0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808, 0x08190819, 0x2b080808,
1136
+ 0x08191908, 0x2b080808, 0x08192b19, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b1919, 0x2b080808,
1137
+ 0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x1919082b, 0x2b080808,
1138
+ 0x19191919, 0x2b080808, 0x19192b08, 0x2b080808, 0x192b0819, 0x2b080808, 0x2b080808, 0x2b080808,
1139
+ 0x2b081919, 0x2b080808, 0x2b190819, 0x2b080808, 0x2b191908, 0x2b080808, 0x08080819, 0x2b080819,
1140
+ 0x08081908, 0x2b080819, 0x08082b19, 0x2b080819, 0x08190808, 0x2b080819, 0x0819082b, 0x2b080819,
1141
+ 0x08191919, 0x2b080819, 0x08192b08, 0x2b080819, 0x082b0819, 0x2b080819, 0x082b1908, 0x2b080819,
1142
+ 0x19080808, 0x2b080819, 0x1908082b, 0x2b080819, 0x19081919, 0x2b080819, 0x19082b08, 0x2b080819,
1143
+ 0x19190819, 0x2b080819, 0x19191908, 0x2b080819, 0x2b080819, 0x2b080819, 0x2b081908, 0x2b080819,
1144
+ 0x2b190808, 0x2b080819, 0x2b2b2b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x08081919, 0x2b08082b,
1145
+ 0x08082b2b, 0x2b08082b, 0x08190819, 0x2b08082b, 0x08191908, 0x2b08082b, 0x19080819, 0x2b08082b,
1146
+ 0x19081908, 0x2b08082b, 0x19190808, 0x2b08082b, 0x08080819, 0x2b081908, 0x08081908, 0x2b081908,
1147
+ 0x0808192b, 0x2b081908, 0x08082b19, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
1148
+ 0x08191919, 0x2b081908, 0x08192b08, 0x2b081908, 0x082b0819, 0x2b081908, 0x19080808, 0x2b081908,
1149
+ 0x1908082b, 0x2b081908, 0x19081919, 0x2b081908, 0x19082b08, 0x2b081908, 0x19190819, 0x2b081908,
1150
+ 0x19191908, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b080819, 0x2b081908, 0x2b081908, 0x2b081908,
1151
+ 0x2b190808, 0x2b081908, 0x08080808, 0x2b081919, 0x0808082b, 0x2b081919, 0x08081919, 0x2b081919,
1152
+ 0x08082b08, 0x2b081919, 0x08190819, 0x2b081919, 0x08191908, 0x2b081919, 0x082b0808, 0x2b081919,
1153
+ 0x19080819, 0x2b081919, 0x19081908, 0x2b081919, 0x19190808, 0x2b081919, 0x2b080808, 0x2b081919,
1154
+ 0x2b082b2b, 0x2b081919, 0x08080819, 0x2b08192b, 0x08081908, 0x2b08192b, 0x08190808, 0x2b08192b,
1155
+ 0x082b2b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08081919, 0x2b082b08,
1156
+ 0x08190819, 0x2b082b08, 0x08191908, 0x2b082b08, 0x19080819, 0x2b082b08, 0x19081908, 0x2b082b08,
1157
+ 0x19190808, 0x2b082b08, 0x2b2b082b, 0x2b082b08, 0x08080819, 0x2b082b19, 0x08081908, 0x2b082b19,
1158
+ 0x19080808, 0x2b082b19, 0x192b1919, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x19192b08, 0x2b082b2b,
1159
+ 0x19192b2b, 0x2b082b2b, 0x2b08082b, 0x2b082b2b, 0x2b2b082b, 0x2b082b2b, 0x08080819, 0x2b190808,
1160
+ 0x08081908, 0x2b190808, 0x08082b19, 0x2b190808, 0x08190808, 0x2b190808, 0x0819082b, 0x2b190808,
1161
+ 0x08191919, 0x2b190808, 0x08192b08, 0x2b190808, 0x082b1908, 0x2b190808, 0x19080808, 0x2b190808,
1162
+ 0x1908082b, 0x2b190808, 0x19081919, 0x2b190808, 0x19082b08, 0x2b190808, 0x19190819, 0x2b190808,
1163
+ 0x19191908, 0x2b190808, 0x192b0808, 0x2b190808, 0x2b080819, 0x2b190808, 0x2b081908, 0x2b190808,
1164
+ 0x2b190808, 0x2b190808, 0x08080808, 0x2b190819, 0x08081919, 0x2b190819, 0x08190819, 0x2b190819,
1165
+ 0x08191908, 0x2b190819, 0x19080819, 0x2b190819, 0x19081908, 0x2b190819, 0x19190808, 0x2b190819,
1166
+ 0x19192b2b, 0x2b190819, 0x08080819, 0x2b19082b, 0x08081908, 0x2b19082b, 0x08190808, 0x2b19082b,
1167
+ 0x19080808, 0x2b19082b, 0x2b2b192b, 0x2b19082b, 0x08080808, 0x2b191908, 0x0808082b, 0x2b191908,
1168
+ 0x08081919, 0x2b191908, 0x08082b08, 0x2b191908, 0x08190819, 0x2b191908, 0x08191908, 0x2b191908,
1169
+ 0x082b0808, 0x2b191908, 0x19080819, 0x2b191908, 0x19081908, 0x2b191908, 0x19190808, 0x2b191908,
1170
+ 0x2b080808, 0x2b191908, 0x2b19192b, 0x2b191908, 0x08080819, 0x2b191919, 0x08081908, 0x2b191919,
1171
+ 0x08190808, 0x2b191919, 0x19080808, 0x2b191919, 0x2b192b08, 0x2b191919, 0x2b2b0819, 0x2b191919,
1172
+ 0x08080808, 0x2b19192b, 0x1908192b, 0x2b19192b, 0x192b1908, 0x2b19192b, 0x08080819, 0x2b192b08,
1173
+ 0x08081908, 0x2b192b08, 0x08190808, 0x2b192b08, 0x082b192b, 0x2b192b08, 0x19080808, 0x2b192b08,
1174
+ 0x2b2b2b19, 0x2b192b08, 0x08080808, 0x2b192b19, 0x19082b19, 0x2b192b19, 0x1919082b, 0x2b192b19,
1175
+ 0x2b190808, 0x2b192b2b, 0x08080808, 0x2b2b0808, 0x08081919, 0x2b2b0808, 0x08082b2b, 0x2b2b0808,
1176
+ 0x08191908, 0x2b2b0808, 0x082b082b, 0x2b2b0808, 0x082b2b2b, 0x2b2b0808, 0x19080819, 0x2b2b0808,
1177
+ 0x19081908, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b2b082b, 0x2b2b0808, 0x2b2b2b2b, 0x2b2b0808,
1178
+ 0x19080808, 0x2b2b0819, 0x192b1919, 0x2b2b0819, 0x0808082b, 0x2b2b082b, 0x08082b2b, 0x2b2b082b,
1179
+ 0x082b082b, 0x2b2b082b, 0x082b2b08, 0x2b2b082b, 0x082b2b2b, 0x2b2b082b, 0x2b08082b, 0x2b2b082b,
1180
+ 0x2b082b08, 0x2b2b082b, 0x2b082b2b, 0x2b2b082b, 0x2b2b2b08, 0x2b2b082b, 0x08080819, 0x2b2b1908,
1181
+ 0x08081908, 0x2b2b1908, 0x08190808, 0x2b2b1908, 0x19080808, 0x2b2b1908, 0x2b082b19, 0x2b2b1908,
1182
+ 0x2b2b1908, 0x2b2b1908, 0x08080808, 0x2b2b1919, 0x08192b19, 0x2b2b1919, 0x19190819, 0x2b2b192b,
1183
+ 0x08082b2b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b082b, 0x2b2b2b08, 0x19191908, 0x2b2b2b19,
1184
+ 0x2b08192b, 0x2b2b2b19, 0x08082b08, 0x2b2b2b2b, 0x08082b2b, 0x2b2b2b2b, 0x082b0808, 0x2b2b2b2b,
1185
+ 0x082b082b, 0x2b2b2b2b, 0x082b2b08, 0x2b2b2b2b, 0x2b082b08, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
1186
+ );
1187
+
1188
+ struct iq2_s {
1189
+ d: f16,
1190
+ qs: array<f16, 32>,
1191
+ qh: array<f16, 4>,
1192
+ scales: array<f16, 4>
1193
+ };
1194
+
1195
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1196
+ let block = src0[src0_idx_base + offset];
1197
+ let d = f32(block.d);
1198
+ var src1_i = src1_idx_base + offset * 256;
1199
+ var qs_vals : array<u32, 16>;
1200
+ for (var i: u32 = 0; i < 16; i++) {
1201
+ qs_vals[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
1202
+ }
1203
+ var qh_vals = array<u32, 2>(
1204
+ bitcast<u32>(vec2(block.qh[0], block.qh[1])),
1205
+ bitcast<u32>(vec2(block.qh[2], block.qh[3]))
1206
+ );
1207
+ var scale_vals = array<u32, 2>(
1208
+ bitcast<u32>(vec2(block.scales[0], block.scales[1])),
1209
+ bitcast<u32>(vec2(block.scales[2], block.scales[3]))
1210
+ );
1211
+ var sum = 0.0;
1212
+ for (var ib: u32 = 0; ib < 8; ib ++) {
1213
+ let s = get_byte(scale_vals[ib / 4], ib % 4);
1214
+ let db = array<f32, 2>(
1215
+ d * (0.5 + f32(s & 0xF)) * 0.25,
1216
+ d * (0.5 + f32(s >> 4)) * 0.25
1217
+ );
1218
+ let qs_w = qs_vals[ib];
1219
+ for (var l: u32 = 0; l < 4; l++) {
1220
+ let qh_b = (get_byte(qh_vals[ib / 4], ib % 4) << (8 - 2 * l)) & 0x300;
1221
+ let ig = (get_byte(qs_w, l) | qh_b) * 8;
1222
+ let signs = get_byte(qs_vals[ib + 8], l);
1223
+ let dl = db[l/2];
1224
+ for (var j: u32 = 0; j < 8; j++) {
1225
+ let g = get_byte(iq2s_grid[(ig + j) / 4], (ig + j) % 4);
1226
+ let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
1227
+ sum += dl * f32(g) * m * src1[src1_i];
1228
+ src1_i++;
1229
+ }
1230
+ }
1231
+ }
1232
+ return sum;
1233
+ }
1234
+
1235
+
1236
+ #enddecl(IQ2_S)
1237
+
1238
+ #decl(IQ3_XSS)
1239
+
1240
+ const iq3xxs_grid = array<u32, 256>(
1241
+ 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
1242
+ 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
1243
+ 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
1244
+ 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
1245
+ 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
1246
+ 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
1247
+ 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
1248
+ 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
1249
+ 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
1250
+ 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
1251
+ 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
1252
+ 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
1253
+ 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
1254
+ 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
1255
+ 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
1256
+ 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
1257
+ 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
1258
+ 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
1259
+ 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
1260
+ 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
1261
+ 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
1262
+ 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
1263
+ 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
1264
+ 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
1265
+ 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
1266
+ 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
1267
+ 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
1268
+ 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
1269
+ 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
1270
+ 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
1271
+ 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
1272
+ 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04
1273
+ );
1274
+
1275
+ struct iq3_xxs {
1276
+ d: f16,
1277
+ qs: array<f16, 48>
1278
+ };
1279
+
1280
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1281
+ let block = src0[src0_idx_base + offset];
1282
+ let d = f32(block.d);
1283
+ var src1_i = src1_idx_base + offset * 256;
1284
+ var sum = 0.0;
1285
+ for (var ib: u32 = 0; ib < 16; ib += 2) {
1286
+ let sc_sign = bitcast<u32>(vec2(block.qs[ib + 32], block.qs[ib + 33]));
1287
+ let db = d * (0.5 + f32(sc_sign >> 28)) * 0.5;
1288
+ for (var l: u32 = 0; l < 4; l++) {
1289
+ let is = (sc_sign >> (7 * l)) & 127;
1290
+ let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
1291
+ let ig_val = bitcast<u32>(vec2(block.qs[ib * 2 + l], 0.0));
1292
+ let ig1 = get_byte(ig_val, 0);
1293
+ let ig2 = get_byte(ig_val, 1);
1294
+ for (var j: u32 = 0; j < 4; j++) {
1295
+ let g1 = get_byte(iq3xxs_grid[ig1], j);
1296
+ let g2 = get_byte(iq3xxs_grid[ig2], j);
1297
+ let m1 = select(1.0, -1.0, (get_byte(kmask_iq2xs[0], j) & signs) != 0);
1298
+ let m2 = select(1.0, -1.0, (get_byte(kmask_iq2xs[1], j) & signs) != 0);
1299
+ sum += db * f32(g1) * m1 * src1[src1_i];
1300
+ sum += db * f32(g2) * m2 * src1[src1_i + 4];
1301
+ src1_i++;
1302
+ }
1303
+ src1_i += 4;
1304
+ }
1305
+ }
1306
+ return sum;
1307
+ }
1308
+
1309
+ #enddecl(IQ3_XSS)
1310
+
1311
+ #decl(IQ3_S)
1312
+
1313
+ const iq3s_grid = array<u32, 512>(
1314
+ 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
1315
+ 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
1316
+ 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
1317
+ 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
1318
+ 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
1319
+ 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
1320
+ 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
1321
+ 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
1322
+ 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
1323
+ 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
1324
+ 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
1325
+ 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
1326
+ 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
1327
+ 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
1328
+ 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
1329
+ 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
1330
+ 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
1331
+ 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
1332
+ 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
1333
+ 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
1334
+ 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
1335
+ 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
1336
+ 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
1337
+ 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
1338
+ 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
1339
+ 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
1340
+ 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
1341
+ 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
1342
+ 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
1343
+ 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
1344
+ 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
1345
+ 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
1346
+ 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
1347
+ 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
1348
+ 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
1349
+ 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
1350
+ 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
1351
+ 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
1352
+ 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
1353
+ 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
1354
+ 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
1355
+ 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
1356
+ 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
1357
+ 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
1358
+ 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
1359
+ 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
1360
+ 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
1361
+ 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
1362
+ 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
1363
+ 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
1364
+ 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
1365
+ 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
1366
+ 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
1367
+ 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
1368
+ 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
1369
+ 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
1370
+ 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
1371
+ 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
1372
+ 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
1373
+ 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
1374
+ 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
1375
+ 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
1376
+ 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
1377
+ 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101
1378
+ );
1379
+
1380
+ struct iq3_s {
1381
+ d: f16,
1382
+ qs: array<f16, 32>,
1383
+ qh: array<f16, 4>,
1384
+ signs: array<f16, 16>,
1385
+ scales: array<f16, 2>
1386
+ };
1387
+
1388
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1389
+ let block = src0[src0_idx_base + offset];
1390
+ let d = f32(block.d);
1391
+ var src1_i = src1_idx_base + offset * 256;
1392
+ var qh_vals = array<u32, 2>(
1393
+ bitcast<u32>(vec2(block.qh[0], block.qh[1])),
1394
+ bitcast<u32>(vec2(block.qh[2], block.qh[3]))
1395
+ );
1396
+ var sign_vals: array<u32, 8>;
1397
+ for (var i: u32 = 0; i < 8; i++) {
1398
+ sign_vals[i] = bitcast<u32>(vec2(block.signs[i * 2], block.signs[i * 2 + 1]));
1399
+ }
1400
+ var scale_vals = bitcast<u32>(vec2(block.scales[0], block.scales[1]));
1401
+ var sum = 0.0;
1402
+ for (var ib: u32 = 0; ib < 4; ib++) {
1403
+ let s = get_byte(scale_vals, ib);
1404
+ let db = array<f32, 2>(
1405
+ d * (1.0 + 2.0 * f32(s & 0xF)),
1406
+ d * (1.0 + 2.0 * f32(s >> 4))
1407
+ );
1408
+ for (var k: u32 = 0; k < 2; k++) {
1409
+ let dl = db[k];
1410
+ let qh_byte = get_byte(qh_vals[ib / 2], (ib % 2) * 2 + k);
1411
+ let sign_w = sign_vals[ib * 2 + k];
1412
+ for (var l: u32 = 0; l < 4; l++) {
1413
+ let signs = get_byte(sign_w, l);
1414
+ let ig_val = bitcast<u32>(vec2(block.qs[ib * 8 + k * 4 + l], 0.0));
1415
+ let ig1 = get_byte(ig_val, 0) | ((qh_byte << ((8 - (2 * l)))) & 256);
1416
+ let ig2 = get_byte(ig_val, 1) | ((qh_byte << ((7 - (2 * l)))) & 256);
1417
+ for (var j: u32 = 0; j < 4; j++) {
1418
+ let g1 = get_byte(iq3s_grid[ig1], j);
1419
+ let g2 = get_byte(iq3s_grid[ig2], j);
1420
+ let m1 = select(1.0, -1.0, (get_byte(kmask_iq2xs[0], j) & signs) != 0);
1421
+ let m2 = select(1.0, -1.0, (get_byte(kmask_iq2xs[1], j) & signs) != 0);
1422
+ sum += dl * f32(g1) * m1 * src1[src1_i];
1423
+ sum += dl * f32(g2) * m2 * src1[src1_i + 4];
1424
+ src1_i++;
1425
+ }
1426
+ src1_i += 4;
1427
+ }
1428
+ }
1429
+ }
1430
+ return sum;
1431
+ }
1432
+ #enddecl(IQ3_S)
1433
+
1434
+ #decl(IQ1_TABLE)
1435
+
1436
+ const IQ1_DELTA: f32 = 0.125;
1437
+
1438
+ const iq1_grid = array<u32, 1024>(
1439
+ 0xfffdffff, 0xfff7fff0, 0xffccfff5, 0xffdfffc0, 0xffd7ffdd, 0xff30ffd5, 0xff03ff0c, 0xff10ff01,
1440
+ 0xff7dff7f, 0xff75ff77, 0xff5fff40, 0xff57ff5d, 0xfcf3ff55, 0xfcccfcf0, 0xfcc1fcc3, 0xfcc5fcc4,
1441
+ 0xfc3cfcd0, 0xfc34fc31, 0xfc00fc0d, 0xfc1cfc05, 0xfc11fc13, 0xfc70fc17, 0xfc43fc4c, 0xfc50fc41,
1442
+ 0xfdfdfdff, 0xfdf5fdf7, 0xfddffdc0, 0xfdd7fddd, 0xfd30fdd5, 0xfd04fd0c, 0xfd14fd13, 0xfd7dfd7f,
1443
+ 0xfd75fd77, 0xfd40fd4c, 0xfd5ffd44, 0xfd57fd5d, 0xf3ccfd55, 0xf3c1f3c3, 0xf33cf3d0, 0xf300f334,
1444
+ 0xf313f305, 0xf34cf310, 0xf350f344, 0xf0f3f0fc, 0xf0f1f0f0, 0xf0c7f0c0, 0xf0d4f0c5, 0xf030f03f,
1445
+ 0xf00ff035, 0xf003f00c, 0xf001f000, 0xf01ff004, 0xf010f01d, 0xf015f017, 0xf04cf07c, 0xf047f040,
1446
+ 0xf05cf045, 0xf050f053, 0xf054f051, 0xf1c4f1c3, 0xf133f13c, 0xf10df10f, 0xf107f100, 0xf11cf11f,
1447
+ 0xf114f111, 0xf14cf170, 0xf144f143, 0xf7fdf7ff, 0xf7f5f7f7, 0xf7dff7c0, 0xf7d7f7dd, 0xf730f7d5,
1448
+ 0xf701f70c, 0xf77ff710, 0xf777f77d, 0xf740f775, 0xf75df75f, 0xf755f757, 0xf4ccf4f0, 0xf4c4f4c3,
1449
+ 0xf4d0f4d3, 0xf40ff43c, 0xf400f40c, 0xf413f41c, 0xf44cf414, 0xf441f443, 0xf450f444, 0xf5fdf5ff,
1450
+ 0xf5f5f5f7, 0xf5dff5c0, 0xf5d7f5dd, 0xf530f5d5, 0xf504f50c, 0xf510f51c, 0xf57df57f, 0xf577f570,
1451
+ 0xf540f575, 0xf55df55f, 0xf555f557, 0xcfcccfcf, 0xcfc4cfc3, 0xcfd0cfd3, 0xcf33cf3c, 0xcf00cf0f,
1452
+ 0xcf1ccf07, 0xcf10cf13, 0xcf4ccf14, 0xcf41cf43, 0xcf50cf5c, 0xccf3ccfc, 0xccf4ccf1, 0xcccdcccf,
1453
+ 0xccc7ccc0, 0xccd3ccdc, 0xcc30ccd4, 0xcc0fcc35, 0xcc0dcc0c, 0xcc00cc03, 0xcc04cc01, 0xcc10cc1f,
1454
+ 0xcc4dcc73, 0xcc5ccc40, 0xcdcccc53, 0xcdc1cdc3, 0xcd3fcdd0, 0xcd34cd31, 0xcd00cd0d, 0xcd05cd07,
1455
+ 0xcd11cd13, 0xcd4ccd70, 0xcd41cd43, 0xc3fccd50, 0xc3f4c3f1, 0xc3c0c3c3, 0xc3c4c3c7, 0xc3d1c3dc,
1456
+ 0xc330c33c, 0xc337c331, 0xc30cc335, 0xc300c303, 0xc304c301, 0xc310c31d, 0xc373c317, 0xc34fc374,
1457
+ 0xc340c343, 0xc344c347, 0xc35cc345, 0xc350c353, 0xc0fdc354, 0xc0f5c0f0, 0xc0c3c0cc, 0xc0c1c0c0,
1458
+ 0xc0dfc0c4, 0xc0d0c0dd, 0xc0d5c0d7, 0xc033c03c, 0xc031c030, 0xc00dc00c, 0xc000c003, 0xc004c001,
1459
+ 0xc01cc005, 0xc010c013, 0xc014c011, 0xc07dc07f, 0xc070c073, 0xc075c077, 0xc04cc04f, 0xc040c043,
1460
+ 0xc044c041, 0xc05fc045, 0xc050c05d, 0xc1f3c1fc, 0xc1f1c1f0, 0xc1c1c1c0, 0xc1c5c1c7, 0xc1d1c1dc,
1461
+ 0xc13dc13f, 0xc130c133, 0xc135c137, 0xc100c10c, 0xc107c101, 0xc11cc104, 0xc110c113, 0xc114c117,
1462
+ 0xc171c115, 0xc14dc175, 0xc153c140, 0xc7ccc154, 0xc7d0c7c1, 0xc733c73c, 0xc734c731, 0xc700c70f,
1463
+ 0xc705c707, 0xc71cc71f, 0xc711c713, 0xc770c714, 0xc743c74c, 0xc4cfc750, 0xc4c0c4cd, 0xc4dcc4c5,
1464
+ 0xc43dc4d0, 0xc430c433, 0xc40cc437, 0xc400c403, 0xc404c401, 0xc41fc405, 0xc415c410, 0xc44cc474,
1465
+ 0xc440c44d, 0xc45cc447, 0xc454c451, 0xc5c1c5f4, 0xc5d1c5d3, 0xc531c533, 0xc50fc534, 0xc500c50d,
1466
+ 0xc51cc507, 0xc514c511, 0xc54cc570, 0xc545c541, 0xdffddfff, 0xdff5dff7, 0xdfdfdfc0, 0xdfd0dfdd,
1467
+ 0xdfd5dfd7, 0xdf0cdf30, 0xdf1cdf04, 0xdf7fdf10, 0xdf77df7d, 0xdf40df75, 0xdf5ddf5f, 0xdf57df50,
1468
+ 0xdcf0df55, 0xdcc3dccc, 0xdcd0dcc4, 0xdc33dc3d, 0xdc00dc34, 0xdc05dc07, 0xdc13dc1c, 0xdc11dc10,
1469
+ 0xdc4fdc70, 0xdc44dc41, 0xddfcdc50, 0xddf5ddf7, 0xddc0ddcc, 0xdddddddf, 0xddd5ddd7, 0xdd0cdd30,
1470
+ 0xdd04dd01, 0xdd7cdd10, 0xdd75dd77, 0xdd40dd4c, 0xdd5ddd5f, 0xdd55dd57, 0xd3c3d3f0, 0xd3c4d3c1,
1471
+ 0xd333d3d0, 0xd331d330, 0xd30dd334, 0xd307d300, 0xd311d305, 0xd34cd370, 0xd344d343, 0xd350d35c,
1472
+ 0xd0c0d0f4, 0xd0d4d0dc, 0xd030d03f, 0xd00cd037, 0xd000d003, 0xd01dd004, 0xd017d010, 0xd04fd074,
1473
+ 0xd040d043, 0xd045d047, 0xd053d05c, 0xd054d051, 0xd1cfd1f0, 0xd1c4d1cd, 0xd13cd1d0, 0xd100d134,
1474
+ 0xd11cd11f, 0xd173d114, 0xd14fd171, 0xd7ffd145, 0xd7f7d7fd, 0xd7c0d7f5, 0xd7ddd7df, 0xd7d5d7d7,
1475
+ 0xd70cd730, 0xd710d703, 0xd77dd77f, 0xd775d777, 0xd75dd75f, 0xd755d757, 0xd4ccd4f4, 0xd4c4d4c3,
1476
+ 0xd431d4d0, 0xd40dd434, 0xd41cd400, 0xd411d413, 0xd470d414, 0xd441d44f, 0xd453d444, 0xd5ffd450,
1477
+ 0xd5f7d5fd, 0xd5dfd5f5, 0xd5d7d5dd, 0xd530d5d5, 0xd501d50c, 0xd510d504, 0xd57dd57f, 0xd575d577,
1478
+ 0xd55fd540, 0xd557d55d, 0x3ff0d555, 0x3fc13fcc, 0x3f343fd0, 0x3f003f0d, 0x3f053f07, 0x3f133f1c,
1479
+ 0x3f433f11, 0x3f5c3f44, 0x3cff3f51, 0x3cf33cfc, 0x3cf43cf1, 0x3cc03ccd, 0x3cc73cc1, 0x3cdc3cc5,
1480
+ 0x3cd43cd1, 0x3c373c30, 0x3c0c3c35, 0x3c003c03, 0x3c043c01, 0x3c103c05, 0x3c153c17, 0x3c733c7c,
1481
+ 0x3c4f3c71, 0x3c403c4d, 0x3c5c3c5f, 0x3df03c5d, 0x3dc33dcc, 0x3dd03dc1, 0x3d0d3d3c, 0x3d053d00,
1482
+ 0x3d143d13, 0x3d433d74, 0x33fc3d50, 0x33c433c0, 0x333033d4, 0x33353337, 0x3303330c, 0x33013300,
1483
+ 0x331d331c, 0x33173310, 0x337c3315, 0x33743371, 0x334d334f, 0x335f3340, 0x3354335c, 0x30fd30fc,
1484
+ 0x30f530f0, 0x30c330cc, 0x30c130c0, 0x30df30c4, 0x30d530d0, 0x3033303c, 0x30313030, 0x300f3034,
1485
+ 0x3003300c, 0x30013000, 0x30043007, 0x3013301c, 0x30113010, 0x307d3014, 0x30703073, 0x304c3077,
1486
+ 0x30403043, 0x30443041, 0x30503045, 0x30553057, 0x31f031fc, 0x31c331f4, 0x31c731c0, 0x31dc31c5,
1487
+ 0x31d431d3, 0x313d313f, 0x31373130, 0x310c310f, 0x3100310d, 0x31043101, 0x3110311d, 0x317c3117,
1488
+ 0x31753170, 0x31403143, 0x3153315c, 0x37f03151, 0x37c037cc, 0x37d037c5, 0x3734373d, 0x3700370f,
1489
+ 0x371c3707, 0x37113713, 0x37703714, 0x3743374c, 0x37443741, 0x34fc3750, 0x34f134f0, 0x34cf34f5,
1490
+ 0x34c034c3, 0x34dc34c7, 0x34d134d3, 0x3430343f, 0x340c3435, 0x3403340d, 0x34013400, 0x341f3404,
1491
+ 0x3410341d, 0x34153411, 0x34743471, 0x3440344d, 0x34473441, 0x3453345c, 0x34543451, 0x353335c1,
1492
+ 0x35343531, 0x35073500, 0x35133505, 0x35433514, 0x0ffc3550, 0x0ff00ff3, 0x0ff40ff1, 0x0fc00fcd,
1493
+ 0x0fdc0fc5, 0x0fd40fd3, 0x0f300f3f, 0x0f0c0f37, 0x0f000f03, 0x0f040f01, 0x0f170f10, 0x0f740f71,
1494
+ 0x0f470f40, 0x0f5c0f5f, 0x0f540f51, 0x0cf70cf0, 0x0cf50cf4, 0x0cc30ccc, 0x0cc10cc0, 0x0cc40cc7,
1495
+ 0x0cd00cdf, 0x0cd70cd1, 0x0c3c0cd5, 0x0c300c33, 0x0c340c31, 0x0c0c0c0f, 0x0c030c0d, 0x0c010c00,
1496
+ 0x0c040c07, 0x0c1c0c05, 0x0c100c13, 0x0c140c11, 0x0c700c7d, 0x0c430c4c, 0x0c410c40, 0x0c5f0c44,
1497
+ 0x0c550c50, 0x0df10dfc, 0x0dc00dcd, 0x0ddc0dc5, 0x0d3d0dd3, 0x0d350d30, 0x0d030d0c, 0x0d010d00,
1498
+ 0x0d1d0d04, 0x0d700d10, 0x0d4d0d4f, 0x0d440d40, 0x0d530d45, 0x03f003f3, 0x03c303cc, 0x03c103c0,
1499
+ 0x03c403c7, 0x03d003dc, 0x03d503d7, 0x0333033c, 0x03310330, 0x03350334, 0x030c030f, 0x03000303,
1500
+ 0x03070301, 0x03050304, 0x031d031c, 0x03100313, 0x03140311, 0x0377037f, 0x034c0375, 0x03400343,
1501
+ 0x03440341, 0x0353035c, 0x03550350, 0x00fd00fc, 0x00f000f3, 0x00f400f1, 0x00cc00cf, 0x00c300cd,
1502
+ 0x00c100c0, 0x00c500c4, 0x00d300dc, 0x00d100d0, 0x003f00d4, 0x003d003c, 0x00300033, 0x00370031,
1503
+ 0x000f0034, 0x000d000c, 0x00000003, 0x00070001, 0x00050004, 0x001c001f, 0x00100013, 0x00170011,
1504
+ 0x00150014, 0x0073007c, 0x00740070, 0x004f0075, 0x0043004c, 0x00410040, 0x00440047, 0x0053005c,
1505
+ 0x00510050, 0x01ff0054, 0x01fd01fc, 0x01f101f3, 0x01f401f7, 0x01c301cc, 0x01c701c0, 0x01df01c4,
1506
+ 0x01dd01dc, 0x01d001d3, 0x01d701d1, 0x013c01d4, 0x01310130, 0x01340137, 0x010f0135, 0x010d010c,
1507
+ 0x01000103, 0x01070101, 0x01050104, 0x0113011c, 0x01140110, 0x0170017d, 0x01770171, 0x01750174,
1508
+ 0x0140014c, 0x015d0145, 0x01510150, 0x01540157, 0x07f007f3, 0x07f407f1, 0x07c007cf, 0x07dc07c7,
1509
+ 0x073007d5, 0x07350737, 0x0703070c, 0x07010700, 0x07040707, 0x071d071f, 0x07100713, 0x0774077d,
1510
+ 0x074d074f, 0x07470740, 0x0754075c, 0x04fd04fc, 0x04f504f0, 0x04c304cc, 0x04c104c0, 0x04d004c4,
1511
+ 0x0433043c, 0x04310430, 0x040f0434, 0x040d040c, 0x04000403, 0x04070401, 0x04050404, 0x0413041c,
1512
+ 0x04110410, 0x047c0414, 0x04740470, 0x0443044c, 0x04410440, 0x04440447, 0x05f30450, 0x05c005f7,
1513
+ 0x05df05c5, 0x05d105d0, 0x053005d4, 0x05340537, 0x0500050c, 0x05070501, 0x051d0504, 0x05170510,
1514
+ 0x057c0515, 0x054d0575, 0x05410540, 0x05450547, 0x1ff0055c, 0x1fc11fc3, 0x1fd01fc4, 0x1f0f1f33,
1515
+ 0x1f011f00, 0x1f051f07, 0x1f131f1c, 0x1f141f11, 0x1f411f7c, 0x1cfc1f50, 0x1cf11cf3, 0x1ccd1cf4,
1516
+ 0x1cdc1cc0, 0x1cd11cdd, 0x1c301cd4, 0x1c0c1c34, 0x1c011c00, 0x1c101c04, 0x1c151c11, 0x1c751c73,
1517
+ 0x1c401c4d, 0x1c511c5c, 0x1dcc1c54, 0x1dc41dc1, 0x1d3c1d3f, 0x1d001d31, 0x1d071d01, 0x1d701d1f,
1518
+ 0x1d411d4c, 0x13cc1d50, 0x13c013cd, 0x13c513c1, 0x13d113dc, 0x133f13d4, 0x1330133d, 0x13351337,
1519
+ 0x1303130c, 0x13011300, 0x13051304, 0x131d131f, 0x13731310, 0x13741370, 0x134d134f, 0x13401343,
1520
+ 0x13471341, 0x135c1345, 0x13541353, 0x10f710f0, 0x10cc10f5, 0x10c110c0, 0x103310c4, 0x10311030,
1521
+ 0x100f1034, 0x1003100c, 0x10011000, 0x101c1004, 0x10101013, 0x10141011, 0x10741071, 0x104c1075,
1522
+ 0x10411040, 0x10451044, 0x1050105d, 0x10571051, 0x11f411fd, 0x11df11c0, 0x11d711d1, 0x113f11d4,
1523
+ 0x11371130, 0x110c1135, 0x11001103, 0x11071101, 0x111f1105, 0x11171110, 0x117d117f, 0x11751170,
1524
+ 0x11411143, 0x11441147, 0x1153115f, 0x11551151, 0x17c417c1, 0x173c17d0, 0x1700170d, 0x171c1705,
1525
+ 0x17701714, 0x1747174c, 0x14fc1751, 0x14cf14f3, 0x14dc14c0, 0x14d114d3, 0x143f14d4, 0x1430143c,
1526
+ 0x14371431, 0x1403140c, 0x14011400, 0x141f1404, 0x14151410, 0x1473147d, 0x14401475, 0x1453145c,
1527
+ 0x14541450, 0x15c115cc, 0x153c15c7, 0x15341533, 0x1500150f, 0x15051507, 0x15101513, 0x15711514,
1528
+ 0x15471543, 0x15511545, 0x7ffd7fff, 0x7ff57ff7, 0x7fdd7fdf, 0x7fd57fd7, 0x7f0f7f30, 0x7f037f0c,
1529
+ 0x7f047f01, 0x7f7f7f10, 0x7f777f7d, 0x7f407f75, 0x7f5d7f5f, 0x7f557f57, 0x7ccc7cf0, 0x7cc17cc3,
1530
+ 0x7cd07cc4, 0x7c337c3c, 0x7c0f7c34, 0x7c007c0d, 0x7c077c01, 0x7c137c04, 0x7c147c11, 0x7c747c70,
1531
+ 0x7c417c43, 0x7c507c44, 0x7dfd7dff, 0x7df57df7, 0x7ddf7dc0, 0x7dd77ddd, 0x7d0c7dd5, 0x7d047d03,
1532
+ 0x7d7f7d10, 0x7d777d7d, 0x7d407d75, 0x7d5d7d5f, 0x7d557d57, 0x73c473c3, 0x7333733c, 0x7300730c,
1533
+ 0x731c7305, 0x73147313, 0x73447343, 0x70f470fc, 0x70c070cd, 0x70d170c5, 0x703f70d4, 0x7030703c,
1534
+ 0x700c7037, 0x70007003, 0x70047001, 0x70107005, 0x70177011, 0x707c7015, 0x70717073, 0x704f7074,
1535
+ 0x7040704d, 0x70517047, 0x71c171cc, 0x71d071c4, 0x7133713c, 0x71357134, 0x7100710f, 0x71057104,
1536
+ 0x7111711c, 0x71707115, 0x7145714c, 0x77ff7153, 0x77f777fd, 0x77c077f5, 0x77dd77df, 0x77d577d7,
1537
+ 0x7730773c, 0x7703770c, 0x77107704, 0x777f7714, 0x7777777d, 0x77407775, 0x775d775f, 0x77557757,
1538
+ 0x74f174f0, 0x74c374cc, 0x74d074c1, 0x7433743c, 0x74347431, 0x740d740f, 0x74057400, 0x7413741c,
1539
+ 0x74417470, 0x74507444, 0x75fd75ff, 0x75f575f7, 0x75df75c0, 0x75d775dd, 0x753075d5, 0x7503750c,
1540
+ 0x757f7501, 0x7577757d, 0x75407575, 0x755d755f, 0x75557557, 0x4fcc4ff0, 0x4fc74fc1, 0x4fd04fc4,
1541
+ 0x4f314f3c, 0x4f004f34, 0x4f054f07, 0x4f154f14, 0x4f4c4f70, 0x4f414f43, 0x4f504f44, 0x4cf34cfc,
1542
+ 0x4cf44cf1, 0x4cc04ccf, 0x4cc54cc7, 0x4cd34cdc, 0x4cd44cd1, 0x4c304c3f, 0x4c0c4c0f, 0x4c004c03,
1543
+ 0x4c044c01, 0x4c104c1d, 0x4c714c73, 0x4c404c4d, 0x4c5c4c47, 0x4c514c53, 0x4df04c54, 0x4dc34dcc,
1544
+ 0x4dd04dc4, 0x4d314d33, 0x4d0f4d34, 0x4d004d0d, 0x4d114d07, 0x4d704d14, 0x4d414d43, 0x43fc4d54,
1545
+ 0x43f143f3, 0x43c043cf, 0x43d143c7, 0x4335433f, 0x4303430c, 0x43014300, 0x43044307, 0x431c431f,
1546
+ 0x4310431d, 0x43714373, 0x4343434d, 0x43474340, 0x4354435c, 0x40f040ff, 0x40f540f7, 0x40cc40cf,
1547
+ 0x40c040c3, 0x40c440c1, 0x40d040dc, 0x40d540d4, 0x4033403c, 0x40314030, 0x400f4034, 0x400d400c,
1548
+ 0x40004003, 0x40074001, 0x40054004, 0x4013401c, 0x40114010, 0x407c4014, 0x40774070, 0x404d404c,
1549
+ 0x40404043, 0x40444041, 0x405f4045, 0x4050405d, 0x40554057, 0x41f341fc, 0x41c041cf, 0x41df41c4,
1550
+ 0x41d441d1, 0x41374130, 0x410c4134, 0x4100410d, 0x41044101, 0x41174110, 0x4173417d, 0x41754174,
1551
+ 0x4143414d, 0x41534140, 0x41544151, 0x47c147f0, 0x47d047c4, 0x4731473c, 0x470d470f, 0x47014700,
1552
+ 0x47134705, 0x47704710, 0x4741474c, 0x47504744, 0x44f144f3, 0x44cf44f4, 0x44c044cd, 0x44c544c7,
1553
+ 0x44dc44df, 0x44d144d3, 0x443d443f, 0x44374430, 0x440c4435, 0x44004403, 0x44044401, 0x4410441d,
1554
+ 0x44154411, 0x4473447c, 0x444d444f, 0x44454440, 0x4451445c, 0x45c045f0, 0x453345d0, 0x45344531,
1555
+ 0x4500450f, 0x451c4507, 0x454c4570, 0x45404543, 0x5fff4541, 0x5ff75ffd, 0x5fc05ff5, 0x5fdd5fdf,
1556
+ 0x5fd55fd7, 0x5f0c5f30, 0x5f015f03, 0x5f7f5f04, 0x5f775f7d, 0x5f405f75, 0x5f5d5f5f, 0x5f555f57,
1557
+ 0x5cf45cf0, 0x5cc35ccc, 0x5cc45cc1, 0x5c315cc5, 0x5c0c5c34, 0x5c075c00, 0x5c1c5c05, 0x5c705c13,
1558
+ 0x5c4d5c4f, 0x5c445c41, 0x5df75dfd, 0x5dcf5df5, 0x5ddd5dc4, 0x5dd55dd7, 0x5d0c5d30, 0x5d045d01,
1559
+ 0x5d7f5d10, 0x5d775d7d, 0x5d405d75, 0x5d5d5d5f, 0x5d555d57, 0x53d053c4, 0x5333533c, 0x5303530f,
1560
+ 0x53075300, 0x531c5305, 0x53115310, 0x53145317, 0x50f15370, 0x50cf50f4, 0x50c050cd, 0x50d150c7,
1561
+ 0x503d50d4, 0x500c5030, 0x50005003, 0x50045001, 0x50155010, 0x5073507c, 0x50715070, 0x504d5074,
1562
+ 0x50475040, 0x51cc51f0, 0x51c551c1, 0x51d051dc, 0x51315133, 0x510d5135, 0x51015100, 0x511f5107,
1563
+ 0x5171511d, 0x5140514f, 0x51445141, 0x5153515c, 0x57ff5151, 0x57f757fd, 0x57df57f5, 0x57d757dd,
1564
+ 0x570c57d5, 0x57015703, 0x577f5704, 0x5777577d, 0x57405775, 0x575d575f, 0x57555757, 0x54c354f0,
1565
+ 0x54dc54c4, 0x543c54d0, 0x5400540f, 0x541c5405, 0x54145411, 0x5441544f, 0x55fd55ff, 0x55f555f7,
1566
+ 0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
1567
+ );
1568
+
1569
+ #enddecl(IQ1_TABLE)
1570
+
1571
+ #decl(IQ1_S)
1572
+
1573
+ struct iq1_s {
1574
+ d: f16,
1575
+ qs: array<f16, 16>,
1576
+ qh: array<f16, 8>
1577
+ };
1578
+
1579
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1580
+ let block = src0[src0_idx_base + offset];
1581
+ let d = f32(block.d);
1582
+ var src1_i = src1_idx_base + offset * 256;
1583
+ var sum = 0.0;
1584
+ for (var ib: u32 = 0; ib < 8; ib++) {
1585
+ let qh = bitcast<u32>(vec2(block.qh[ib], 0.0));
1586
+ let dl = d * (2 * f32((qh >> 12) & 7) + 1);
1587
+ let delta = select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x8000) != 0);
1588
+ let qs_w = bitcast<u32>(vec2(block.qs[ib * 2], block.qs[ib * 2 + 1]));
1589
+ for (var l: u32 = 0; l < 4; l++) {
1590
+ let ig = (get_byte(qs_w, l) | (((qh >> (3 * l)) & 7) << 8)) * 8;
1591
+ for (var j: u32 = 0; j < 8; j++) {
1592
+ let gw = iq1_grid[(ig + j) / 16];
1593
+ let g = (gw >> (((ig + j) % 16) * 2)) & 3;
1594
+ let gs = bitcast<i32>(g << 30) >> 30;
1595
+ sum += dl * (f32(gs) + delta) * src1[src1_i];
1596
+ src1_i++;
1597
+ }
1598
+ }
1599
+ }
1600
+ return sum;
1601
+ }
1602
+
1603
+ #enddecl(IQ1_S)
1604
+
1605
+ #decl(IQ1_M)
1606
+
1607
+ struct iq1_m {
1608
+ qs: array<u32, 8>,
1609
+ qh: array<u32, 4>,
1610
+ scales: array<u32, 2>
1611
+ };
1612
+
1613
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1614
+ let block = src0[src0_idx_base + offset];
1615
+
1616
+ let scale = ((block.scales[0] >> 12) & 0xF) | ((block.scales[0] >> 24) & 0x00F0) | ((block.scales[1] >> 4) & 0x0F00) | ((block.scales[1] >> 16) & 0xF000);
1617
+ let d = f32(bitcast<vec2<f16>>(scale).x);
1618
+ var src1_i = src1_idx_base + offset * 256;
1619
+ var sum = 0.0;
1620
+ for (var ib: u32 = 0; ib < 8; ib++) {
1621
+ let sw = (block.scales[ib / 4] >> (16 * ((ib / 2) % 2))) & 0xFFFF;
1622
+ let s1 : u32 = (sw >> (6 * (ib % 2))) & 0x7;
1623
+ let s2 : u32 = (sw >> (6 * (ib % 2) + 3)) & 0x7;
1624
+ var dl = array<f32, 2>(
1625
+ d * f32(2 * s1 + 1),
1626
+ d * f32(2 * s2 + 1)
1627
+ );
1628
+
1629
+ let qh = block.qh[ib / 2] >> (16 * (ib % 2));
1630
+ var idx = array<u32, 4>(
1631
+ get_byte(block.qs[ib], 0) | ((qh << 8) & 0x700),
1632
+ get_byte(block.qs[ib], 1) | ((qh << 4) & 0x700),
1633
+ get_byte(block.qs[ib], 2) | ((qh) & 0x700),
1634
+ get_byte(block.qs[ib], 3) | ((qh >> 4) & 0x700)
1635
+ );
1636
+ var delta = array<f32, 4>(
1637
+ select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x08) != 0),
1638
+ select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x80) != 0),
1639
+ select(IQ1_DELTA, -IQ1_DELTA, ((qh >> 8) & 0x08) != 0),
1640
+ select(IQ1_DELTA, -IQ1_DELTA, ((qh >> 8) & 0x80) != 0)
1641
+ );
1642
+ for (var l: u32 = 0; l < 4; l++) {
1643
+ let ig = idx[l] * 8;
1644
+ for (var j: u32 = 0; j < 8; j++) {
1645
+ let gw = iq1_grid[(ig + j) / 16];
1646
+ let g = (gw >> (((ig + j) % 16) * 2)) & 3;
1647
+ let gs = bitcast<i32>(g << 30) >> 30;
1648
+ sum += dl[l/2] * (f32(gs) + delta[l]) * src1[src1_i];
1649
+ src1_i++;
1650
+ }
1651
+ }
1652
+ }
1653
+ return sum;
1654
+ }
1655
+
1656
+ #enddecl(IQ1_M)
1657
+
1658
+ #decl(IQ4_TABLE)
1659
+
1660
+ const kvalues_iq4nl = array<i32, 16>(
1661
+ -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113
1662
+ );
1663
+
1664
+ #enddecl(IQ4_TABLE)
1665
+
1666
+ #decl(IQ4_NL)
1667
+
1668
+ struct iq4_nl {
1669
+ d: f16,
1670
+ qs: array<f16, 8>,
1671
+ }
1672
+
1673
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1674
+ let block = src0[src0_idx_base + offset];
1675
+ let d = f32(block.d);
1676
+ var src1_i = src1_idx_base + offset * 32;
1677
+ var sum = 0.0;
1678
+ var qs: array<u32, 4>;
1679
+ for (var i: u32 = 0; i < 4; i++) {
1680
+ qs[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
1681
+ }
1682
+ for (var j: u32 = 0; j < 16; j++) {
1683
+ let qsb = get_byte(qs[j / 4], j % 4);
1684
+ sum += d * f32(kvalues_iq4nl[qsb & 0xF]) * src1[src1_i];
1685
+ sum += d * f32(kvalues_iq4nl[qsb >> 4]) * src1[src1_i + 16];
1686
+ src1_i++;
1687
+ }
1688
+ return sum;
1689
+ }
1690
+
1691
+ #enddecl(IQ4_NL)
1692
+
1693
+ #decl(IQ4_XS)
1694
+
1695
+ struct iq4_xs {
1696
+ d: f16,
1697
+ scales_h: f16,
1698
+ scales_l: u32,
1699
+ qs: array<u32, 32>
1700
+ };
1701
+
1702
+ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
1703
+ let block = src0[src0_idx_base + offset];
1704
+ let d = f32(block.d);
1705
+ let scales_h = bitcast<u32>(vec2(block.scales_h, 0.0));
1706
+ var src1_i = src1_idx_base + offset * 256;
1707
+ var sum = 0.0;
1708
+ for (var ib: u32 = 0; ib < 8; ib++) {
1709
+ let ls = ((get_byte(block.scales_l, ib / 2) >> (4 * (ib % 2))) & 0xF) | (((scales_h >> (2 * ib)) & 3) << 4);
1710
+ let dl = d * (f32(ls) - 32.0);
1711
+ for (var j: u32 = 0; j < 16; j++) {
1712
+ let iqs = ib * 16 + j;
1713
+ let qsb = get_byte(block.qs[iqs / 4], iqs % 4);
1714
+ sum += dl * f32(kvalues_iq4nl[qsb & 0xF]) * src1[src1_i];
1715
+ sum += dl * f32(kvalues_iq4nl[qsb >> 4]) * src1[src1_i + 16];
1716
+ src1_i++;
1717
+ }
1718
+ src1_i += 16;
1719
+ }
1720
+ return sum;
1721
+ }
1722
+
1723
+ #enddecl(IQ4_XS)
1724
+
1725
+ #end(DECLS)
1726
+
1727
+ #define(SHADER)
1728
+
1729
+ enable f16;
1730
+
1731
+ DECLS
1732
+
1733
+ struct MulMatParams {
1734
+ offset_src0: u32, // in elements/blocks
1735
+ offset_src1: u32, // in elements/blocks
1736
+ offset_dst: u32, // in elements/blocks
1737
+ m: u32,
1738
+ n: u32,
1739
+ k: u32,
1740
+ // all strides are in elements/blocks
1741
+ stride_01: u32,
1742
+ stride_11: u32,
1743
+ stride_02: u32,
1744
+ stride_12: u32,
1745
+ stride_03: u32,
1746
+ stride_13: u32,
1747
+
1748
+ bs02: u32,
1749
+ bs03: u32,
1750
+ broadcast2: u32,
1751
+ broadcast3: u32
1752
+ };
1753
+
1754
+ @group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // N rows, K columns
1755
+ @group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // M rows, K columns (transposed)
1756
+ @group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
1757
+
1758
+ @group(0) @binding(3) var<uniform> params: MulMatParams;
1759
+
1760
+ @compute @workgroup_size(64)
1761
+ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
1762
+ let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
1763
+ if (global_id.x >= total) {
1764
+ return;
1765
+ }
1766
+
1767
+ let dst2_stride = params.m * params.n;
1768
+ let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
1769
+
1770
+ let dst3_idx = global_id.x / dst3_stride;
1771
+ let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
1772
+ let src13_idx = dst3_idx; // src1 is not broadcast
1773
+ let dst3_rem = global_id.x % dst3_stride;
1774
+
1775
+ let dst2_idx = dst3_rem / dst2_stride;
1776
+ let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
1777
+ let src12_idx = dst2_idx; // src1 is not broadcast
1778
+
1779
+ let dst2_rem = dst3_rem % dst2_stride;
1780
+
1781
+ let row = dst2_rem / params.n; // output row
1782
+ let col = dst2_rem % params.n; // output column
1783
+
1784
+ let src0_idx_base = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01;
1785
+ let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11;
1786
+
1787
+ var sum = 0.0;
1788
+ for (var i: u32 = 0u; i < params.k/{{BLOCK_SIZE}}; i = i + 1u) {
1789
+ sum += multiply_add(src0_idx_base, src1_idx_base, i);
1790
+ }
1791
+ dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
1792
+ }
1793
+
1794
+ #end(SHADER)