node-llama-cpp 3.0.0-beta.12 → 3.0.0-beta.13

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 (181) hide show
  1. package/dist/ChatWrapper.d.ts +1 -0
  2. package/dist/ChatWrapper.js +2 -1
  3. package/dist/ChatWrapper.js.map +1 -1
  4. package/dist/TemplateChatWrapper.d.ts +67 -0
  5. package/dist/TemplateChatWrapper.js +239 -0
  6. package/dist/TemplateChatWrapper.js.map +1 -0
  7. package/dist/bindings/AddonTypes.d.ts +2 -0
  8. package/dist/bindings/Llama.d.ts +1 -3
  9. package/dist/bindings/Llama.js +10 -20
  10. package/dist/bindings/Llama.js.map +1 -1
  11. package/dist/bindings/consts.d.ts +2 -0
  12. package/dist/bindings/consts.js +11 -0
  13. package/dist/bindings/consts.js.map +1 -0
  14. package/dist/bindings/getLlama.d.ts +14 -18
  15. package/dist/bindings/getLlama.js +210 -78
  16. package/dist/bindings/getLlama.js.map +1 -1
  17. package/dist/bindings/types.d.ts +8 -5
  18. package/dist/bindings/types.js +18 -0
  19. package/dist/bindings/types.js.map +1 -1
  20. package/dist/bindings/utils/asyncEvery.d.ts +5 -0
  21. package/dist/bindings/utils/asyncEvery.js +15 -0
  22. package/dist/bindings/utils/asyncEvery.js.map +1 -0
  23. package/dist/bindings/utils/asyncSome.d.ts +5 -0
  24. package/dist/bindings/utils/asyncSome.js +27 -0
  25. package/dist/bindings/utils/asyncSome.js.map +1 -0
  26. package/dist/bindings/utils/cloneLlamaCppRepo.js +13 -3
  27. package/dist/bindings/utils/cloneLlamaCppRepo.js.map +1 -1
  28. package/dist/bindings/utils/compileLLamaCpp.js +30 -4
  29. package/dist/bindings/utils/compileLLamaCpp.js.map +1 -1
  30. package/dist/bindings/utils/detectAvailableComputeLayers.d.ts +11 -0
  31. package/dist/bindings/utils/detectAvailableComputeLayers.js +158 -0
  32. package/dist/bindings/utils/detectAvailableComputeLayers.js.map +1 -0
  33. package/dist/bindings/utils/detectGlibc.d.ts +4 -0
  34. package/dist/bindings/utils/detectGlibc.js +36 -0
  35. package/dist/bindings/utils/detectGlibc.js.map +1 -0
  36. package/dist/bindings/utils/getBestComputeLayersAvailable.d.ts +9 -0
  37. package/dist/bindings/utils/getBestComputeLayersAvailable.js +29 -0
  38. package/dist/bindings/utils/getBestComputeLayersAvailable.js.map +1 -0
  39. package/dist/bindings/utils/getBuildFolderNameForBuildOptions.js +12 -6
  40. package/dist/bindings/utils/getBuildFolderNameForBuildOptions.js.map +1 -1
  41. package/dist/bindings/utils/getGpuTypesToUseForOption.d.ts +11 -0
  42. package/dist/bindings/utils/getGpuTypesToUseForOption.js +30 -0
  43. package/dist/bindings/utils/getGpuTypesToUseForOption.js.map +1 -0
  44. package/dist/bindings/utils/getLinuxDistroInfo.d.ts +9 -0
  45. package/dist/bindings/utils/getLinuxDistroInfo.js +46 -0
  46. package/dist/bindings/utils/getLinuxDistroInfo.js.map +1 -0
  47. package/dist/bindings/utils/getPlatformInfo.d.ts +5 -0
  48. package/dist/bindings/utils/getPlatformInfo.js +28 -0
  49. package/dist/bindings/utils/getPlatformInfo.js.map +1 -0
  50. package/dist/bindings/utils/hasFileInPath.d.ts +2 -0
  51. package/dist/bindings/utils/hasFileInPath.js +34 -0
  52. package/dist/bindings/utils/hasFileInPath.js.map +1 -0
  53. package/dist/bindings/utils/logBinaryUsageExampleToConsole.d.ts +1 -1
  54. package/dist/bindings/utils/logBinaryUsageExampleToConsole.js +3 -9
  55. package/dist/bindings/utils/logBinaryUsageExampleToConsole.js.map +1 -1
  56. package/dist/bindings/utils/logDistroInstallInstruction.d.ts +13 -0
  57. package/dist/bindings/utils/logDistroInstallInstruction.js +38 -0
  58. package/dist/bindings/utils/logDistroInstallInstruction.js.map +1 -0
  59. package/dist/bindings/utils/resolveChatWrapperBasedOnWrapperTypeName.d.ts +9 -2
  60. package/dist/bindings/utils/resolveChatWrapperBasedOnWrapperTypeName.js +10 -4
  61. package/dist/bindings/utils/resolveChatWrapperBasedOnWrapperTypeName.js.map +1 -1
  62. package/dist/bindings/utils/testBindingBinary.d.ts +1 -0
  63. package/dist/bindings/utils/testBindingBinary.js +98 -0
  64. package/dist/bindings/utils/testBindingBinary.js.map +1 -0
  65. package/dist/chatWrappers/ChatMLChatWrapper.js +1 -1
  66. package/dist/chatWrappers/ChatMLChatWrapper.js.map +1 -1
  67. package/dist/chatWrappers/GemmaChatWrapper.d.ts +18 -0
  68. package/dist/chatWrappers/GemmaChatWrapper.js +86 -0
  69. package/dist/chatWrappers/GemmaChatWrapper.js.map +1 -0
  70. package/dist/chatWrappers/resolveChatWrapperBasedOnModel.js +3 -0
  71. package/dist/chatWrappers/resolveChatWrapperBasedOnModel.js.map +1 -1
  72. package/dist/cli/cli.js +2 -0
  73. package/dist/cli/cli.js.map +1 -1
  74. package/dist/cli/commands/BuildCommand.d.ts +5 -5
  75. package/dist/cli/commands/BuildCommand.js +78 -60
  76. package/dist/cli/commands/BuildCommand.js.map +1 -1
  77. package/dist/cli/commands/DebugCommand.js +3 -9
  78. package/dist/cli/commands/DebugCommand.js.map +1 -1
  79. package/dist/cli/commands/DownloadCommand.d.ts +5 -5
  80. package/dist/cli/commands/DownloadCommand.js +97 -56
  81. package/dist/cli/commands/DownloadCommand.js.map +1 -1
  82. package/dist/cli/commands/InspectCommand.d.ts +7 -0
  83. package/dist/cli/commands/InspectCommand.js +113 -0
  84. package/dist/cli/commands/InspectCommand.js.map +1 -0
  85. package/dist/cli/utils/logUsedGpuTypeOption.d.ts +2 -0
  86. package/dist/cli/utils/logUsedGpuTypeOption.js +9 -0
  87. package/dist/cli/utils/logUsedGpuTypeOption.js.map +1 -0
  88. package/dist/config.d.ts +3 -3
  89. package/dist/config.js +10 -11
  90. package/dist/config.js.map +1 -1
  91. package/dist/evaluator/LlamaChatSession/utils/defineChatSessionFunction.d.ts +3 -0
  92. package/dist/evaluator/LlamaChatSession/utils/defineChatSessionFunction.js +3 -0
  93. package/dist/evaluator/LlamaChatSession/utils/defineChatSessionFunction.js.map +1 -1
  94. package/dist/gguf/GGUFInsights.d.ts +28 -0
  95. package/dist/gguf/GGUFInsights.js +58 -0
  96. package/dist/gguf/GGUFInsights.js.map +1 -0
  97. package/dist/gguf/GGUFMetadata.d.ts +19 -0
  98. package/dist/gguf/GGUFMetadata.js +38 -0
  99. package/dist/gguf/GGUFMetadata.js.map +1 -0
  100. package/dist/gguf/errors/InvalidGGUFMagicError.d.ts +3 -0
  101. package/dist/gguf/errors/InvalidGGUFMagicError.js +6 -0
  102. package/dist/gguf/errors/InvalidGGUFMagicError.js.map +1 -0
  103. package/dist/gguf/errors/MetadataNotParsedYetError.d.ts +3 -0
  104. package/dist/gguf/errors/MetadataNotParsedYetError.js +6 -0
  105. package/dist/gguf/errors/MetadataNotParsedYetError.js.map +1 -0
  106. package/dist/gguf/errors/MissingNodeLlamaError.d.ts +3 -0
  107. package/dist/gguf/errors/MissingNodeLlamaError.js +6 -0
  108. package/dist/gguf/errors/MissingNodeLlamaError.js.map +1 -0
  109. package/dist/gguf/errors/ModelScore/NotEnoughVRamError.d.ts +5 -0
  110. package/dist/gguf/errors/ModelScore/NotEnoughVRamError.js +12 -0
  111. package/dist/gguf/errors/ModelScore/NotEnoughVRamError.js.map +1 -0
  112. package/dist/gguf/errors/UnsupportedMetadataTypeError.d.ts +4 -0
  113. package/dist/gguf/errors/UnsupportedMetadataTypeError.js +8 -0
  114. package/dist/gguf/errors/UnsupportedMetadataTypeError.js.map +1 -0
  115. package/dist/gguf/ggufParser/GGUFParser.d.ts +18 -0
  116. package/dist/gguf/ggufParser/GGUFParser.js +123 -0
  117. package/dist/gguf/ggufParser/GGUFParser.js.map +1 -0
  118. package/dist/gguf/ggufParser/GGUFTypes.d.ts +257 -0
  119. package/dist/gguf/ggufParser/GGUFTypes.js +2 -0
  120. package/dist/gguf/ggufParser/GGUFTypes.js.map +1 -0
  121. package/dist/gguf/ggufParser/checkArchitecture.d.ts +14 -0
  122. package/dist/gguf/ggufParser/checkArchitecture.js +74 -0
  123. package/dist/gguf/ggufParser/checkArchitecture.js.map +1 -0
  124. package/dist/gguf/ggufParser/stream/GGUFBaseStream.d.ts +38 -0
  125. package/dist/gguf/ggufParser/stream/GGUFBaseStream.js +83 -0
  126. package/dist/gguf/ggufParser/stream/GGUFBaseStream.js.map +1 -0
  127. package/dist/gguf/ggufParser/stream/GGUFFetchStream.d.ts +14 -0
  128. package/dist/gguf/ggufParser/stream/GGUFFetchStream.js +35 -0
  129. package/dist/gguf/ggufParser/stream/GGUFFetchStream.js.map +1 -0
  130. package/dist/gguf/ggufParser/stream/GGUFReadStream.d.ts +15 -0
  131. package/dist/gguf/ggufParser/stream/GGUFReadStream.js +40 -0
  132. package/dist/gguf/ggufParser/stream/GGUFReadStream.js.map +1 -0
  133. package/dist/index.d.ts +3 -1
  134. package/dist/index.js +3 -1
  135. package/dist/index.js.map +1 -1
  136. package/dist/utils/LlamaText.js +2 -2
  137. package/dist/utils/LlamaText.js.map +1 -1
  138. package/dist/utils/cmake.js +23 -10
  139. package/dist/utils/cmake.js.map +1 -1
  140. package/dist/utils/getBuildDefaults.d.ts +1 -3
  141. package/dist/utils/getBuildDefaults.js +2 -4
  142. package/dist/utils/getBuildDefaults.js.map +1 -1
  143. package/dist/utils/getConsoleLogPrefix.d.ts +1 -1
  144. package/dist/utils/getConsoleLogPrefix.js +2 -2
  145. package/dist/utils/getConsoleLogPrefix.js.map +1 -1
  146. package/dist/utils/mergeUnionTypes.d.ts +6 -0
  147. package/dist/utils/mergeUnionTypes.js +2 -0
  148. package/dist/utils/mergeUnionTypes.js.map +1 -0
  149. package/dist/utils/parseTextTemplate.d.ts +66 -0
  150. package/dist/utils/parseTextTemplate.js +116 -0
  151. package/dist/utils/parseTextTemplate.js.map +1 -0
  152. package/llama/CMakeLists.txt +11 -5
  153. package/llama/addon.cpp +31 -7
  154. package/llama/binariesGithubRelease.json +1 -1
  155. package/llama/gitRelease.bundle +0 -0
  156. package/llama/llama.cpp.info.json +1 -1
  157. package/llamaBins/linux-arm64/.buildMetadata.json +1 -1
  158. package/llamaBins/linux-arm64/llama-addon.node +0 -0
  159. package/llamaBins/linux-armv7l/.buildMetadata.json +1 -1
  160. package/llamaBins/linux-armv7l/llama-addon.node +0 -0
  161. package/llamaBins/linux-x64/.buildMetadata.json +1 -1
  162. package/llamaBins/linux-x64/llama-addon.node +0 -0
  163. package/llamaBins/linux-x64-cuda/.buildMetadata.json +1 -1
  164. package/llamaBins/linux-x64-cuda/llama-addon.node +0 -0
  165. package/llamaBins/linux-x64-vulkan/.buildMetadata.json +1 -1
  166. package/llamaBins/linux-x64-vulkan/llama-addon.node +0 -0
  167. package/llamaBins/mac-arm64-metal/.buildMetadata.json +1 -1
  168. package/llamaBins/mac-arm64-metal/ggml-metal.metal +815 -106
  169. package/llamaBins/mac-arm64-metal/llama-addon.node +0 -0
  170. package/llamaBins/mac-x64/.buildMetadata.json +1 -1
  171. package/llamaBins/mac-x64/llama-addon.node +0 -0
  172. package/llamaBins/win-x64/.buildMetadata.json +1 -1
  173. package/llamaBins/win-x64/llama-addon.node +0 -0
  174. package/llamaBins/win-x64-cuda/.buildMetadata.json +1 -1
  175. package/llamaBins/win-x64-cuda/llama-addon.node +0 -0
  176. package/llamaBins/win-x64-vulkan/.buildMetadata.json +1 -1
  177. package/llamaBins/win-x64-vulkan/llama-addon.node +0 -0
  178. package/package.json +6 -4
  179. package/dist/cli/utils/logEnabledComputeLayers.d.ts +0 -8
  180. package/dist/cli/utils/logEnabledComputeLayers.js +0 -11
  181. package/dist/cli/utils/logEnabledComputeLayers.js.map +0 -1
@@ -2519,6 +2519,14 @@ typedef struct {
2519
2519
  } block_iq2_xs;
2520
2520
  // 74 bytes / block for QK_K = 256, so 2.3125 bpw
2521
2521
 
2522
+ // 2.5625 bpw quants
2523
+ typedef struct {
2524
+ half d;
2525
+ uint8_t qs[QK_K/4];
2526
+ uint8_t qh[QK_K/32];
2527
+ uint8_t scales[QK_K/32];
2528
+ } block_iq2_s;
2529
+
2522
2530
  typedef struct {
2523
2531
  half d;
2524
2532
  uint8_t qs[3*QK_K/8];
@@ -2552,6 +2560,17 @@ typedef struct {
2552
2560
  uint8_t qs[QK4_NL/2];
2553
2561
  } block_iq4_nl;
2554
2562
 
2563
+ #if QK_K == 64
2564
+ #define block_iq4_xs block_iq4_nl
2565
+ #else
2566
+ typedef struct {
2567
+ half d;
2568
+ uint16_t scales_h;
2569
+ uint8_t scales_l[QK_K/64];
2570
+ uint8_t qs[QK_K/2];
2571
+ } block_iq4_xs;
2572
+ #endif
2573
+
2555
2574
  //====================================== dot products =========================
2556
2575
 
2557
2576
  void kernel_mul_mv_q2_K_f32_impl(
@@ -3774,6 +3793,265 @@ constexpr constant static uint64_t iq2xs_grid[512] = {
3774
3793
  0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
3775
3794
  };
3776
3795
 
3796
+ constexpr constant static uint64_t iq2s_grid[1024] = {
3797
+ 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
3798
+ 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
3799
+ 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
3800
+ 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
3801
+ 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
3802
+ 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
3803
+ 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
3804
+ 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
3805
+ 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
3806
+ 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
3807
+ 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
3808
+ 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
3809
+ 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
3810
+ 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
3811
+ 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
3812
+ 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
3813
+ 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
3814
+ 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
3815
+ 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
3816
+ 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
3817
+ 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
3818
+ 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
3819
+ 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
3820
+ 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
3821
+ 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
3822
+ 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
3823
+ 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
3824
+ 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
3825
+ 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
3826
+ 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
3827
+ 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
3828
+ 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
3829
+ 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
3830
+ 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
3831
+ 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
3832
+ 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
3833
+ 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
3834
+ 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
3835
+ 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
3836
+ 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
3837
+ 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
3838
+ 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
3839
+ 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
3840
+ 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
3841
+ 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
3842
+ 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
3843
+ 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
3844
+ 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
3845
+ 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
3846
+ 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
3847
+ 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
3848
+ 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
3849
+ 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
3850
+ 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
3851
+ 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
3852
+ 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
3853
+ 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
3854
+ 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
3855
+ 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
3856
+ 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
3857
+ 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
3858
+ 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
3859
+ 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
3860
+ 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
3861
+ 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
3862
+ 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
3863
+ 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
3864
+ 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
3865
+ 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
3866
+ 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
3867
+ 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
3868
+ 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
3869
+ 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
3870
+ 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
3871
+ 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
3872
+ 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
3873
+ 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
3874
+ 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
3875
+ 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
3876
+ 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
3877
+ 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
3878
+ 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
3879
+ 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
3880
+ 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
3881
+ 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
3882
+ 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
3883
+ 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
3884
+ 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
3885
+ 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
3886
+ 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
3887
+ 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
3888
+ 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
3889
+ 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
3890
+ 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
3891
+ 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
3892
+ 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
3893
+ 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
3894
+ 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
3895
+ 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
3896
+ 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
3897
+ 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
3898
+ 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
3899
+ 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
3900
+ 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
3901
+ 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
3902
+ 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
3903
+ 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
3904
+ 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
3905
+ 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
3906
+ 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
3907
+ 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
3908
+ 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
3909
+ 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
3910
+ 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
3911
+ 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
3912
+ 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
3913
+ 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
3914
+ 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
3915
+ 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
3916
+ 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
3917
+ 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
3918
+ 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
3919
+ 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
3920
+ 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
3921
+ 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
3922
+ 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
3923
+ 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
3924
+ 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
3925
+ 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
3926
+ 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
3927
+ 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
3928
+ 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
3929
+ 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
3930
+ 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
3931
+ 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
3932
+ 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
3933
+ 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
3934
+ 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
3935
+ 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
3936
+ 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
3937
+ 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
3938
+ 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
3939
+ 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
3940
+ 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
3941
+ 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
3942
+ 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
3943
+ 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
3944
+ 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
3945
+ 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
3946
+ 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
3947
+ 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
3948
+ 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
3949
+ 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
3950
+ 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
3951
+ 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
3952
+ 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
3953
+ 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
3954
+ 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
3955
+ 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
3956
+ 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
3957
+ 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
3958
+ 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
3959
+ 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
3960
+ 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
3961
+ 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
3962
+ 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
3963
+ 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
3964
+ 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
3965
+ 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
3966
+ 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
3967
+ 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
3968
+ 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
3969
+ 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
3970
+ 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
3971
+ 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
3972
+ 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
3973
+ 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
3974
+ 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
3975
+ 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
3976
+ 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
3977
+ 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
3978
+ 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
3979
+ 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
3980
+ 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
3981
+ 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
3982
+ 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
3983
+ 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
3984
+ 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
3985
+ 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
3986
+ 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
3987
+ 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
3988
+ 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
3989
+ 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
3990
+ 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
3991
+ 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
3992
+ 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
3993
+ 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
3994
+ 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
3995
+ 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
3996
+ 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
3997
+ 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
3998
+ 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
3999
+ 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
4000
+ 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
4001
+ 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
4002
+ 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
4003
+ 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
4004
+ 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
4005
+ 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
4006
+ 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
4007
+ 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
4008
+ 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
4009
+ 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
4010
+ 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
4011
+ 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
4012
+ 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
4013
+ 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
4014
+ 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
4015
+ 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
4016
+ 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
4017
+ 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
4018
+ 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
4019
+ 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
4020
+ 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
4021
+ 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
4022
+ 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
4023
+ 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
4024
+ 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
4025
+ 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
4026
+ 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
4027
+ 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
4028
+ 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
4029
+ 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
4030
+ 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
4031
+ 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
4032
+ 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
4033
+ 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
4034
+ 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
4035
+ 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
4036
+ 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
4037
+ 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
4038
+ 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
4039
+ 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
4040
+ 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
4041
+ 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
4042
+ 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
4043
+ 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
4044
+ 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
4045
+ 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
4046
+ 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
4047
+ 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
4048
+ 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
4049
+ 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
4050
+ 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
4051
+ 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
4052
+ 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
4053
+ };
4054
+
3777
4055
  constexpr constant static uint32_t iq3xxs_grid[256] = {
3778
4056
  0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
3779
4057
  0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
@@ -3809,71 +4087,71 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
3809
4087
  0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
3810
4088
  };
3811
4089
 
3812
- constexpr constant static uint32_t iq3xs_grid[512] = {
3813
- 0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14,
3814
- 0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414,
3815
- 0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24,
3816
- 0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c,
3817
- 0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c,
3818
- 0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34,
3819
- 0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c,
3820
- 0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414,
3821
- 0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c,
3822
- 0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404,
3823
- 0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434,
3824
- 0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c,
3825
- 0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404,
3826
- 0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414,
3827
- 0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414,
3828
- 0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404,
3829
- 0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c,
3830
- 0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c,
3831
- 0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404,
3832
- 0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e,
3833
- 0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14,
3834
- 0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c,
3835
- 0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424,
3836
- 0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c,
3837
- 0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c,
3838
- 0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e,
3839
- 0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e,
3840
- 0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e,
3841
- 0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424,
3842
- 0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e,
3843
- 0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424,
3844
- 0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404,
3845
- 0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c,
3846
- 0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e,
3847
- 0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c,
3848
- 0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c,
3849
- 0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c,
3850
- 0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404,
3851
- 0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04,
3852
- 0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c,
3853
- 0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414,
3854
- 0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c,
3855
- 0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c,
3856
- 0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424,
3857
- 0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c,
3858
- 0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c,
3859
- 0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414,
3860
- 0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c,
3861
- 0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e,
3862
- 0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04,
3863
- 0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424,
3864
- 0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14,
3865
- 0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34,
3866
- 0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c,
3867
- 0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434,
3868
- 0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c,
3869
- 0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424,
3870
- 0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24,
3871
- 0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24,
3872
- 0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e,
3873
- 0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c,
3874
- 0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c,
3875
- 0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c,
3876
- 0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404,
4090
+ constexpr constant static uint32_t iq3s_grid[512] = {
4091
+ 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
4092
+ 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
4093
+ 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
4094
+ 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
4095
+ 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
4096
+ 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
4097
+ 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
4098
+ 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
4099
+ 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
4100
+ 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
4101
+ 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
4102
+ 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
4103
+ 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
4104
+ 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
4105
+ 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
4106
+ 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
4107
+ 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
4108
+ 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
4109
+ 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
4110
+ 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
4111
+ 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
4112
+ 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
4113
+ 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
4114
+ 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
4115
+ 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
4116
+ 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
4117
+ 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
4118
+ 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
4119
+ 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
4120
+ 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
4121
+ 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
4122
+ 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
4123
+ 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
4124
+ 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
4125
+ 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
4126
+ 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
4127
+ 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
4128
+ 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
4129
+ 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
4130
+ 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
4131
+ 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
4132
+ 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
4133
+ 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
4134
+ 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
4135
+ 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
4136
+ 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
4137
+ 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
4138
+ 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
4139
+ 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
4140
+ 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
4141
+ 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
4142
+ 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
4143
+ 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
4144
+ 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
4145
+ 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
4146
+ 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
4147
+ 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
4148
+ 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
4149
+ 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
4150
+ 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
4151
+ 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
4152
+ 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
4153
+ 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
4154
+ 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
3877
4155
  };
3878
4156
 
3879
4157
  #define NGRID_IQ1S 512
@@ -4072,7 +4350,6 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
4072
4350
  threadgroup_barrier(mem_flags::mem_threadgroup);
4073
4351
  }
4074
4352
 
4075
- #if QK_K == 256
4076
4353
  const int ix = tiisg;
4077
4354
 
4078
4355
  device const float * y4 = y + 32 * ix;
@@ -4113,12 +4390,6 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
4113
4390
 
4114
4391
  y4 += 32 * 32;
4115
4392
  }
4116
- #else
4117
- (void) x;
4118
- (void) y;
4119
- (void) yl;
4120
- (void) nb32;
4121
- #endif
4122
4393
 
4123
4394
  for (int row = 0; row < N_DST; ++row) {
4124
4395
  all_sum = simd_sum(sumf[row]);
@@ -4208,7 +4479,6 @@ void kernel_mul_mv_iq2_xs_f32_impl(
4208
4479
  threadgroup_barrier(mem_flags::mem_threadgroup);
4209
4480
  }
4210
4481
 
4211
- #if QK_K == 256
4212
4482
  const int ix = tiisg;
4213
4483
 
4214
4484
  device const float * y4 = y + 32 * ix;
@@ -4259,12 +4529,6 @@ void kernel_mul_mv_iq2_xs_f32_impl(
4259
4529
 
4260
4530
  y4 += 32 * 32;
4261
4531
  }
4262
- #else
4263
- (void) x;
4264
- (void) y;
4265
- (void) yl;
4266
- (void) nb32;
4267
- #endif
4268
4532
 
4269
4533
  for (int row = 0; row < N_DST; ++row) {
4270
4534
  all_sum = simd_sum(sumf[row]);
@@ -4354,7 +4618,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
4354
4618
  threadgroup_barrier(mem_flags::mem_threadgroup);
4355
4619
  }
4356
4620
 
4357
- #if QK_K == 256
4358
4621
  const int ix = tiisg;
4359
4622
 
4360
4623
  device const float * y4 = y + 32 * ix;
@@ -4398,12 +4661,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
4398
4661
 
4399
4662
  y4 += 32 * 32;
4400
4663
  }
4401
- #else
4402
- (void) x;
4403
- (void) y;
4404
- (void) yl;
4405
- (void) nb32;
4406
- #endif
4407
4664
 
4408
4665
  for (int row = 0; row < N_DST; ++row) {
4409
4666
  all_sum = simd_sum(sumf[row]);
@@ -4485,7 +4742,7 @@ void kernel_mul_mv_iq3_s_f32_impl(
4485
4742
  {
4486
4743
  int nval = 8;
4487
4744
  int pos = (32*sgitg + tiisg)*nval;
4488
- for (int i = 0; i < nval; ++i) values[pos + i] = iq3xs_grid[pos + i];
4745
+ for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
4489
4746
  threadgroup_barrier(mem_flags::mem_threadgroup);
4490
4747
  }
4491
4748
 
@@ -4512,12 +4769,14 @@ void kernel_mul_mv_iq3_s_f32_impl(
4512
4769
  for (int row = 0; row < N_DST; row++) {
4513
4770
 
4514
4771
  const float db = dh[0];
4515
- const float d = db * (0.5f + ((sc[0] >> 4*(ib%2)) & 0xf));
4772
+ const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf));
4516
4773
 
4517
4774
  float2 sum = {0};
4518
4775
  for (int l = 0; l < 4; ++l) {
4519
- const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
4520
- const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
4776
+ const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
4777
+ const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
4778
+ const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
4779
+ const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
4521
4780
  for (int j = 0; j < 4; ++j) {
4522
4781
  sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
4523
4782
  sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
@@ -4538,7 +4797,7 @@ void kernel_mul_mv_iq3_s_f32_impl(
4538
4797
  for (int row = 0; row < N_DST; ++row) {
4539
4798
  all_sum = simd_sum(sumf[row]);
4540
4799
  if (tiisg == 0) {
4541
- dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
4800
+ dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
4542
4801
  }
4543
4802
  }
4544
4803
  }
@@ -4572,6 +4831,139 @@ kernel void kernel_mul_mv_iq3_s_f32(
4572
4831
  kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
4573
4832
  }
4574
4833
 
4834
+ void kernel_mul_mv_iq2_s_f32_impl(
4835
+ device const void * src0,
4836
+ device const float * src1,
4837
+ device float * dst,
4838
+ constant int64_t & ne00,
4839
+ constant int64_t & ne01,
4840
+ constant int64_t & ne02,
4841
+ constant int64_t & ne10,
4842
+ constant int64_t & ne12,
4843
+ constant int64_t & ne0,
4844
+ constant int64_t & ne1,
4845
+ constant uint & r2,
4846
+ constant uint & r3,
4847
+ threadgroup int8_t * shared_values [[threadgroup(0)]],
4848
+ uint3 tgpig[[threadgroup_position_in_grid]],
4849
+ uint tiisg[[thread_index_in_simdgroup]],
4850
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
4851
+
4852
+ const int nb = ne00/QK_K;
4853
+ const int r0 = tgpig.x;
4854
+ const int r1 = tgpig.y;
4855
+ const int im = tgpig.z;
4856
+
4857
+ const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
4858
+ const int ib_row = first_row * nb;
4859
+
4860
+ const uint i12 = im%ne12;
4861
+ const uint i13 = im/ne12;
4862
+
4863
+ const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
4864
+
4865
+ device const block_iq2_s * x = (device const block_iq2_s *) src0 + ib_row + offset0;
4866
+ device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
4867
+
4868
+ float yl[32];
4869
+ float sumf[N_DST]={0.f}, all_sum;
4870
+
4871
+ const int nb32 = nb * (QK_K / 32);
4872
+
4873
+ //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
4874
+ //{
4875
+ // int nval = 32;
4876
+ // int pos = (32*sgitg + tiisg)*nval;
4877
+ // for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
4878
+ // threadgroup_barrier(mem_flags::mem_threadgroup);
4879
+ //}
4880
+
4881
+ const int ix = tiisg;
4882
+
4883
+ device const float * y4 = y + 32 * ix;
4884
+
4885
+ for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
4886
+
4887
+ for (int i = 0; i < 32; ++i) {
4888
+ yl[i] = y4[i];
4889
+ }
4890
+
4891
+ const int ibl = ib32 / (QK_K / 32);
4892
+ const int ib = ib32 % (QK_K / 32);
4893
+
4894
+ device const block_iq2_s * xr = x + ibl;
4895
+ device const uint8_t * qs = xr->qs + 4 * ib;
4896
+ device const uint8_t * qh = xr->qh + ib;
4897
+ device const uint8_t * sc = xr->scales + ib;
4898
+ device const uint8_t * signs = qs + QK_K/8;
4899
+ device const half * dh = &xr->d;
4900
+
4901
+ for (int row = 0; row < N_DST; row++) {
4902
+
4903
+ const float db = dh[0];
4904
+ const float d1 = db * (0.5f + (sc[0] & 0xf));
4905
+ const float d2 = db * (0.5f + (sc[0] >> 4));
4906
+
4907
+ float2 sum = {0};
4908
+ for (int l = 0; l < 2; ++l) {
4909
+ //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
4910
+ //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
4911
+ constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
4912
+ constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
4913
+ for (int j = 0; j < 8; ++j) {
4914
+ sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]);
4915
+ sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]);
4916
+ }
4917
+ }
4918
+ sumf[row] += d1 * sum[0] + d2 * sum[1];
4919
+
4920
+ dh += nb*sizeof(block_iq2_s)/2;
4921
+ qs += nb*sizeof(block_iq2_s);
4922
+ qh += nb*sizeof(block_iq2_s);
4923
+ sc += nb*sizeof(block_iq2_s);
4924
+ signs += nb*sizeof(block_iq2_s);
4925
+ }
4926
+
4927
+ y4 += 32 * 32;
4928
+ }
4929
+
4930
+ for (int row = 0; row < N_DST; ++row) {
4931
+ all_sum = simd_sum(sumf[row]);
4932
+ if (tiisg == 0) {
4933
+ dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
4934
+ }
4935
+ }
4936
+ }
4937
+
4938
+ [[host_name("kernel_mul_mv_iq2_s_f32")]]
4939
+ kernel void kernel_mul_mv_iq2_s_f32(
4940
+ device const void * src0,
4941
+ device const float * src1,
4942
+ device float * dst,
4943
+ constant int64_t & ne00,
4944
+ constant int64_t & ne01,
4945
+ constant int64_t & ne02,
4946
+ constant uint64_t & nb00,
4947
+ constant uint64_t & nb01,
4948
+ constant uint64_t & nb02,
4949
+ constant int64_t & ne10,
4950
+ constant int64_t & ne11,
4951
+ constant int64_t & ne12,
4952
+ constant uint64_t & nb10,
4953
+ constant uint64_t & nb11,
4954
+ constant uint64_t & nb12,
4955
+ constant int64_t & ne0,
4956
+ constant int64_t & ne1,
4957
+ constant uint & r2,
4958
+ constant uint & r3,
4959
+ threadgroup int8_t * shared_values [[threadgroup(0)]],
4960
+ uint3 tgpig[[threadgroup_position_in_grid]],
4961
+ uint tiisg[[thread_index_in_simdgroup]],
4962
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
4963
+
4964
+ kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
4965
+ }
4966
+
4575
4967
  void kernel_mul_mv_iq1_s_f32_impl(
4576
4968
  device const void * src0,
4577
4969
  device const float * src1,
@@ -4609,7 +5001,6 @@ void kernel_mul_mv_iq1_s_f32_impl(
4609
5001
 
4610
5002
  const int nb32 = nb * (QK_K / 32);
4611
5003
 
4612
- #if QK_K == 256
4613
5004
  const int ix = tiisg/2;
4614
5005
  const int il = tiisg%2;
4615
5006
 
@@ -4648,12 +5039,6 @@ void kernel_mul_mv_iq1_s_f32_impl(
4648
5039
 
4649
5040
  y4 += 16 * 32;
4650
5041
  }
4651
- #else
4652
- (void) x;
4653
- (void) y;
4654
- (void) yl;
4655
- (void) nb32;
4656
- #endif
4657
5042
 
4658
5043
  for (int row = 0; row < N_DST; ++row) {
4659
5044
  all_sum = simd_sum(sumf[row]);
@@ -4760,6 +5145,102 @@ void kernel_mul_mv_iq4_nl_f32_impl(
4760
5145
  }
4761
5146
  }
4762
5147
 
5148
+ #if QK_K != 64
5149
+ void kernel_mul_mv_iq4_xs_f32_impl(
5150
+ device const void * src0,
5151
+ device const float * src1,
5152
+ device float * dst,
5153
+ constant int64_t & ne00,
5154
+ constant int64_t & ne01,
5155
+ constant int64_t & ne02,
5156
+ constant int64_t & ne10,
5157
+ constant int64_t & ne12,
5158
+ constant int64_t & ne0,
5159
+ constant int64_t & ne1,
5160
+ constant uint & r2,
5161
+ constant uint & r3,
5162
+ threadgroup float * shared_values [[threadgroup(0)]],
5163
+ uint3 tgpig[[threadgroup_position_in_grid]],
5164
+ uint tiisg[[thread_index_in_simdgroup]],
5165
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
5166
+
5167
+ const int nb = ne00/QK_K;
5168
+ const int r0 = tgpig.x;
5169
+ const int r1 = tgpig.y;
5170
+ const int im = tgpig.z;
5171
+ const int first_row = (r0 * 2 + sgitg) * 2;
5172
+ const int ib_row = first_row * nb;
5173
+
5174
+ const uint i12 = im%ne12;
5175
+ const uint i13 = im/ne12;
5176
+
5177
+ const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
5178
+ device const block_iq4_xs * x = (device const block_iq4_xs *) src0 + ib_row + offset0;
5179
+ device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
5180
+
5181
+ const int ix = tiisg/16; // 0 or 1
5182
+ const int it = tiisg%16; // 0...15
5183
+ const int ib = it/2;
5184
+ const int il = it%2;
5185
+
5186
+ shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
5187
+ threadgroup_barrier(mem_flags::mem_threadgroup);
5188
+
5189
+ float4 yl[4];
5190
+ float sumf[2]={0.f}, all_sum;
5191
+
5192
+ device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
5193
+
5194
+ uint32_t aux32[2];
5195
+ thread const uint8_t * q8 = (thread const uint8_t *)aux32;
5196
+
5197
+ float4 qf1, qf2;
5198
+
5199
+ for (int ibl = ix; ibl < nb; ibl += 2) {
5200
+
5201
+ device const float4 * y4 = (device const float4 *)yb;
5202
+ yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
5203
+
5204
+ for (int row = 0; row < 2; ++row) {
5205
+
5206
+ device const block_iq4_xs & xb = x[row*nb + ibl];
5207
+ device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
5208
+
5209
+ float4 acc1 = {0.f}, acc2 = {0.f};
5210
+
5211
+ aux32[0] = q4[0] & 0x0f0f0f0f;
5212
+ aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
5213
+ qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
5214
+ qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
5215
+ acc1 += yl[0] * qf1;
5216
+ acc2 += yl[1] * qf2;
5217
+
5218
+ aux32[0] = q4[1] & 0x0f0f0f0f;
5219
+ aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
5220
+ qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
5221
+ qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
5222
+ acc1 += yl[2] * qf1;
5223
+ acc2 += yl[3] * qf2;
5224
+
5225
+ acc1 += acc2;
5226
+
5227
+ const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32;
5228
+ sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
5229
+
5230
+ }
5231
+
5232
+ yb += 2 * QK_K;
5233
+ }
5234
+
5235
+ for (int row = 0; row < 2; ++row) {
5236
+ all_sum = simd_sum(sumf[row]);
5237
+ if (tiisg == 0) {
5238
+ dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
5239
+ }
5240
+ }
5241
+ }
5242
+ #endif
5243
+
4763
5244
  [[host_name("kernel_mul_mv_iq1_s_f32")]]
4764
5245
  kernel void kernel_mul_mv_iq1_s_f32(
4765
5246
  device const void * src0,
@@ -4817,6 +5298,39 @@ kernel void kernel_mul_mv_iq4_nl_f32(
4817
5298
  kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
4818
5299
  }
4819
5300
 
5301
+ [[host_name("kernel_mul_mv_iq4_xs_f32")]]
5302
+ kernel void kernel_mul_mv_iq4_xs_f32(
5303
+ device const void * src0,
5304
+ device const float * src1,
5305
+ device float * dst,
5306
+ constant int64_t & ne00,
5307
+ constant int64_t & ne01,
5308
+ constant int64_t & ne02,
5309
+ constant uint64_t & nb00,
5310
+ constant uint64_t & nb01,
5311
+ constant uint64_t & nb02,
5312
+ constant int64_t & ne10,
5313
+ constant int64_t & ne11,
5314
+ constant int64_t & ne12,
5315
+ constant uint64_t & nb10,
5316
+ constant uint64_t & nb11,
5317
+ constant uint64_t & nb12,
5318
+ constant int64_t & ne0,
5319
+ constant int64_t & ne1,
5320
+ constant uint & r2,
5321
+ constant uint & r3,
5322
+ threadgroup float * shared_values [[threadgroup(0)]],
5323
+ uint3 tgpig[[threadgroup_position_in_grid]],
5324
+ uint tiisg[[thread_index_in_simdgroup]],
5325
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
5326
+
5327
+ #if QK_K == 64
5328
+ kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
5329
+ #else
5330
+ kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
5331
+ #endif
5332
+ }
5333
+
4820
5334
  //============================= templates and their specializations =============================
4821
5335
 
4822
5336
  // NOTE: this is not dequantizing - we are simply fitting the template
@@ -5173,21 +5687,40 @@ void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 &
5173
5687
  device const uint8_t * qs = xb->qs + 8*ib32;
5174
5688
  device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
5175
5689
  const uint8_t qh = xb->qh[ib32] >> 4*il;
5176
- const float dl = d * (0.5f + ((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * 0.5f;
5177
- constant uint8_t * grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+0] | ((qh << 8) & 256)));
5178
- constant uint8_t * grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+1] | ((qh << 7) & 256)));
5690
+ const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
5691
+ constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
5692
+ constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
5179
5693
  for (int i = 0; i < 4; ++i) {
5180
5694
  reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
5181
5695
  reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
5182
5696
  }
5183
- grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+2] | ((qh << 6) & 256)));
5184
- grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+3] | ((qh << 5) & 256)));
5697
+ grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
5698
+ grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
5185
5699
  for (int i = 0; i < 4; ++i) {
5186
5700
  reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
5187
5701
  reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
5188
5702
  }
5189
5703
  }
5190
5704
 
5705
+ template <typename type4x4>
5706
+ void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
5707
+ // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
5708
+ const float d = xb->d;
5709
+ const int ib32 = il/2;
5710
+ il = il%2;
5711
+ // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
5712
+ device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
5713
+ device const uint8_t * signs = qs + QK_K/8;
5714
+ const uint8_t qh = xb->qh[ib32] >> 4*il;
5715
+ const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
5716
+ constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
5717
+ constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
5718
+ for (int i = 0; i < 8; ++i) {
5719
+ reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
5720
+ reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
5721
+ }
5722
+ }
5723
+
5191
5724
  template <typename type4x4>
5192
5725
  void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
5193
5726
  // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
@@ -5219,6 +5752,30 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4
5219
5752
  }
5220
5753
  }
5221
5754
 
5755
+ template <typename type4x4>
5756
+ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
5757
+ #if QK_K == 64
5758
+ dequantize_iq4_nl(xb, il, reg);
5759
+ #else
5760
+ // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
5761
+ const int ib32 = il/2;
5762
+ il = il%2;
5763
+ // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
5764
+ device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
5765
+ const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
5766
+ const float d = (float)xb->d * (ls - 32);
5767
+ uint32_t aux32;
5768
+ thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
5769
+ for (int i = 0; i < 4; ++i) {
5770
+ aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
5771
+ reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
5772
+ reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
5773
+ reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
5774
+ reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
5775
+ }
5776
+ #endif
5777
+ }
5778
+
5222
5779
  template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
5223
5780
  kernel void kernel_get_rows(
5224
5781
  device const void * src0,
@@ -5762,8 +6319,14 @@ template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_r
5762
6319
  template [[host_name("kernel_get_rows_iq2_xs")]] kernel get_rows_t kernel_get_rows<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
5763
6320
  template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
5764
6321
  template [[host_name("kernel_get_rows_iq3_s")]] kernel get_rows_t kernel_get_rows<block_iq3_s, QK_NL, dequantize_iq3_s>;
6322
+ template [[host_name("kernel_get_rows_iq2_s")]] kernel get_rows_t kernel_get_rows<block_iq2_s, QK_NL, dequantize_iq2_s>;
5765
6323
  template [[host_name("kernel_get_rows_iq1_s")]] kernel get_rows_t kernel_get_rows<block_iq1_s, QK_NL, dequantize_iq1_s>;
5766
- template [[host_name("kernel_get_rows_iq4_nl")]] kernel get_rows_t kernel_get_rows<block_iq4_nl, 2, dequantize_iq4_nl>;
6324
+ template [[host_name("kernel_get_rows_iq4_nl")]] kernel get_rows_t kernel_get_rows<block_iq4_nl, 2, dequantize_iq4_nl>;
6325
+ #if QK_K == 64
6326
+ template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, 2, dequantize_iq4_xs>;
6327
+ #else
6328
+ template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
6329
+ #endif
5767
6330
 
5768
6331
  //
5769
6332
  // matrix-matrix multiplication
@@ -5804,8 +6367,14 @@ template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_m
5804
6367
  template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
5805
6368
  template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
5806
6369
  template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_s, QK_NL, dequantize_iq3_s>;
6370
+ template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_s, QK_NL, dequantize_iq2_s>;
5807
6371
  template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_s, QK_NL, dequantize_iq1_s>;
5808
- template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_nl>;
6372
+ template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_nl>;
6373
+ #if QK_K == 64
6374
+ template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_xs>;
6375
+ #else
6376
+ template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
6377
+ #endif
5809
6378
 
5810
6379
  //
5811
6380
  // indirect matrix-matrix multiplication
@@ -5858,8 +6427,14 @@ template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel
5858
6427
  template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
5859
6428
  template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
5860
6429
  template [[host_name("kernel_mul_mm_id_iq3_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s, QK_NL, dequantize_iq3_s>;
6430
+ template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s, QK_NL, dequantize_iq2_s>;
5861
6431
  template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s, QK_NL, dequantize_iq1_s>;
5862
- template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl, 2, dequantize_iq4_nl>;
6432
+ template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl, 2, dequantize_iq4_nl>;
6433
+ #if QK_K == 64
6434
+ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, 2, dequantize_iq4_xs>;
6435
+ #else
6436
+ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
6437
+ #endif
5863
6438
 
5864
6439
  //
5865
6440
  // matrix-vector multiplication
@@ -6893,6 +7468,71 @@ kernel void kernel_mul_mv_id_iq3_s_f32(
6893
7468
  sgitg);
6894
7469
  }
6895
7470
 
7471
+ [[host_name("kernel_mul_mv_id_iq2_s_f32")]]
7472
+ kernel void kernel_mul_mv_id_iq2_s_f32(
7473
+ device const char * ids,
7474
+ device const char * src1,
7475
+ device float * dst,
7476
+ constant uint64_t & nbi1,
7477
+ constant int64_t & ne00,
7478
+ constant int64_t & ne01,
7479
+ constant int64_t & ne02,
7480
+ constant uint64_t & nb00,
7481
+ constant uint64_t & nb01,
7482
+ constant uint64_t & nb02,
7483
+ constant int64_t & ne10,
7484
+ constant int64_t & ne11,
7485
+ constant int64_t & ne12,
7486
+ constant int64_t & ne13,
7487
+ constant uint64_t & nb10,
7488
+ constant uint64_t & nb11,
7489
+ constant uint64_t & nb12,
7490
+ constant int64_t & ne0,
7491
+ constant int64_t & ne1,
7492
+ constant uint64_t & nb1,
7493
+ constant uint & r2,
7494
+ constant uint & r3,
7495
+ constant int & idx,
7496
+ device const char * src00,
7497
+ device const char * src01,
7498
+ device const char * src02,
7499
+ device const char * src03,
7500
+ device const char * src04,
7501
+ device const char * src05,
7502
+ device const char * src06,
7503
+ device const char * src07,
7504
+ threadgroup int8_t * shared_values [[threadgroup(0)]],
7505
+ uint3 tgpig[[threadgroup_position_in_grid]],
7506
+ uint tiitg[[thread_index_in_threadgroup]],
7507
+ uint tiisg[[thread_index_in_simdgroup]],
7508
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
7509
+ device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
7510
+
7511
+ const int64_t bid = tgpig.z/(ne12*ne13);
7512
+
7513
+ tgpig.z = tgpig.z%(ne12*ne13);
7514
+
7515
+ const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
7516
+
7517
+ kernel_mul_mv_iq2_s_f32_impl(
7518
+ src0[id],
7519
+ (device const float *) (src1 + bid*nb11),
7520
+ dst + bid*ne0,
7521
+ ne00,
7522
+ ne01,
7523
+ ne02,
7524
+ ne10,
7525
+ ne12,
7526
+ ne0,
7527
+ ne1,
7528
+ r2,
7529
+ r3,
7530
+ shared_values,
7531
+ tgpig,
7532
+ tiisg,
7533
+ sgitg);
7534
+ }
7535
+
6896
7536
  [[host_name("kernel_mul_mv_id_iq1_s_f32")]]
6897
7537
  kernel void kernel_mul_mv_id_iq1_s_f32(
6898
7538
  device const char * ids,
@@ -7020,3 +7660,72 @@ kernel void kernel_mul_mv_id_iq4_nl_f32(
7020
7660
  tiisg,
7021
7661
  sgitg);
7022
7662
  }
7663
+
7664
+ [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]
7665
+ kernel void kernel_mul_mv_id_iq4_xs_f32(
7666
+ device const char * ids,
7667
+ device const char * src1,
7668
+ device float * dst,
7669
+ constant uint64_t & nbi1,
7670
+ constant int64_t & ne00,
7671
+ constant int64_t & ne01,
7672
+ constant int64_t & ne02,
7673
+ constant uint64_t & nb00,
7674
+ constant uint64_t & nb01,
7675
+ constant uint64_t & nb02,
7676
+ constant int64_t & ne10,
7677
+ constant int64_t & ne11,
7678
+ constant int64_t & ne12,
7679
+ constant int64_t & ne13,
7680
+ constant uint64_t & nb10,
7681
+ constant uint64_t & nb11,
7682
+ constant uint64_t & nb12,
7683
+ constant int64_t & ne0,
7684
+ constant int64_t & ne1,
7685
+ constant uint64_t & nb1,
7686
+ constant uint & r2,
7687
+ constant uint & r3,
7688
+ constant int & idx,
7689
+ device const char * src00,
7690
+ device const char * src01,
7691
+ device const char * src02,
7692
+ device const char * src03,
7693
+ device const char * src04,
7694
+ device const char * src05,
7695
+ device const char * src06,
7696
+ device const char * src07,
7697
+ threadgroup float * shared_values [[threadgroup(0)]],
7698
+ uint3 tgpig[[threadgroup_position_in_grid]],
7699
+ uint tiitg[[thread_index_in_threadgroup]],
7700
+ uint tiisg[[thread_index_in_simdgroup]],
7701
+ uint sgitg[[simdgroup_index_in_threadgroup]]) {
7702
+ device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
7703
+
7704
+ const int64_t bid = tgpig.z/(ne12*ne13);
7705
+
7706
+ tgpig.z = tgpig.z%(ne12*ne13);
7707
+
7708
+ const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
7709
+
7710
+ #if QK_K == 64
7711
+ kernel_mul_mv_iq4_nl_f32_impl(
7712
+ #else
7713
+ kernel_mul_mv_iq4_xs_f32_impl(
7714
+ #endif
7715
+ src0[id],
7716
+ (device const float *) (src1 + bid*nb11),
7717
+ dst + bid*ne0,
7718
+ ne00,
7719
+ ne01,
7720
+ ne02,
7721
+ ne10,
7722
+ ne12,
7723
+ ne0,
7724
+ ne1,
7725
+ r2,
7726
+ r3,
7727
+ shared_values,
7728
+ tgpig,
7729
+ tiisg,
7730
+ sgitg);
7731
+ }