tree: a79b0cbd5119edfc4819d61a4cd9b5c778842845 [path history] [tgz]
  1. a32_sgemm_8x6.hpp
  2. a32_sgemm_8x6/
  3. a64_ffhybrid_bf16fp32_mmla_6x16.hpp
  4. a64_ffhybrid_bf16fp32_mmla_6x16/
  5. a64_ffhybrid_fp16_mla_6x32.hpp
  6. a64_ffhybrid_fp16_mla_6x32/
  7. a64_ffhybrid_fp32_mla_6x16.hpp
  8. a64_ffhybrid_fp32_mla_6x16/
  9. a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
  10. a64_ffhybrid_fp32bf16fp32_mmla_4x24/
  11. a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
  12. a64_ffhybrid_fp32bf16fp32_mmla_6x16/
  13. a64_ffinterleaved_bf16fp32_dot_8x12.hpp
  14. a64_ffinterleaved_bf16fp32_dot_8x12/
  15. a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
  16. a64_ffinterleaved_bf16fp32_mmla_8x12/
  17. a64_ffinterleaved_fp16_mla_8x24.hpp
  18. a64_ffinterleaved_fp16_mla_8x24/
  19. a64_ffinterleaved_fp32_mla_8x12.hpp
  20. a64_ffinterleaved_fp32_mla_8x12/
  21. a64_gemm_s16_8x12.hpp
  22. a64_gemm_s16_8x12/
  23. a64_gemm_s8_4x4.hpp
  24. a64_gemm_s8_4x4/
  25. a64_gemm_s8_8x12.hpp
  26. a64_gemm_s8_8x12/
  27. a64_gemm_u16_8x12.hpp
  28. a64_gemm_u16_8x12/
  29. a64_gemm_u8_4x4.hpp
  30. a64_gemm_u8_4x4/
  31. a64_gemm_u8_8x12.hpp
  32. a64_gemm_u8_8x12/
  33. a64_hgemm_8x24.hpp
  34. a64_hgemm_8x24/
  35. a64_hybrid_bf16fp32_dot_6x16.hpp
  36. a64_hybrid_bf16fp32_dot_6x16/
  37. a64_hybrid_bf16fp32_mmla_6x16.hpp
  38. a64_hybrid_bf16fp32_mmla_6x16/
  39. a64_hybrid_fp16_mla_6x32.hpp
  40. a64_hybrid_fp16_mla_6x32/
  41. a64_hybrid_fp32_mla_4x24.hpp
  42. a64_hybrid_fp32_mla_4x24/
  43. a64_hybrid_fp32_mla_6x16.hpp
  44. a64_hybrid_fp32_mla_6x16/
  45. a64_hybrid_fp32_mla_8x4.hpp
  46. a64_hybrid_fp32_mla_8x4/
  47. a64_hybrid_fp32bf16fp32_mmla_4x24.hpp
  48. a64_hybrid_fp32bf16fp32_mmla_4x24/
  49. a64_hybrid_fp32bf16fp32_mmla_6x16.hpp
  50. a64_hybrid_fp32bf16fp32_mmla_6x16/
  51. a64_hybrid_s8qa_dot_4x16.hpp
  52. a64_hybrid_s8qa_dot_4x16/
  53. a64_hybrid_s8qa_mmla_4x16.hpp
  54. a64_hybrid_s8qa_mmla_4x16/
  55. a64_hybrid_s8qs_dot_6x16.hpp
  56. a64_hybrid_s8qs_dot_6x16/
  57. a64_hybrid_s8qs_mmla_6x16.hpp
  58. a64_hybrid_s8qs_mmla_6x16/
  59. a64_hybrid_s8s32_dot_6x16.hpp
  60. a64_hybrid_s8s32_dot_6x16/
  61. a64_hybrid_s8s32_mmla_6x16.hpp
  62. a64_hybrid_s8s32_mmla_6x16/
  63. a64_hybrid_u8qa_dot_4x16.hpp
  64. a64_hybrid_u8qa_dot_4x16/
  65. a64_hybrid_u8qa_mmla_4x16.hpp
  66. a64_hybrid_u8qa_mmla_4x16/
  67. a64_hybrid_u8u32_dot_6x16.hpp
  68. a64_hybrid_u8u32_dot_6x16/
  69. a64_hybrid_u8u32_mmla_6x16.hpp
  70. a64_hybrid_u8u32_mmla_6x16/
  71. a64_interleaved_bf16fp32_dot_8x12.hpp
  72. a64_interleaved_bf16fp32_dot_8x12/
  73. a64_interleaved_bf16fp32_mmla_8x12.hpp
  74. a64_interleaved_bf16fp32_mmla_8x12/
  75. a64_interleaved_s8s32_mmla_8x12.hpp
  76. a64_interleaved_s8s32_mmla_8x12/
  77. a64_interleaved_u8u32_mmla_8x12.hpp
  78. a64_interleaved_u8u32_mmla_8x12/
  79. a64_sgemm_8x12.hpp
  80. a64_sgemm_8x12/
  81. a64_sgemm_8x6.hpp
  82. a64_sgemm_8x6/
  83. a64_sgemv_pretransposed.hpp
  84. a64_sgemv_pretransposed/
  85. a64_smallK_hybrid_fp32_mla_6x4.hpp
  86. a64_smallK_hybrid_fp32_mla_6x4/
  87. a64_smallK_hybrid_fp32_mla_8x4.hpp
  88. a64_smallK_hybrid_fp32_mla_8x4/
  89. a64_smallK_hybrid_s8s32_dot_6x4.hpp
  90. a64_smallK_hybrid_s8s32_dot_6x4/
  91. a64_smallK_hybrid_s8s32_dot_8x4.hpp
  92. a64_smallK_hybrid_s8s32_dot_8x4/
  93. a64_smallK_hybrid_u8u32_dot_6x4.hpp
  94. a64_smallK_hybrid_u8u32_dot_6x4/
  95. a64_smallK_hybrid_u8u32_dot_8x4.hpp
  96. a64_smallK_hybrid_u8u32_dot_8x4/
  97. sme2_gemv_bf16fp32_dot_16VL.hpp
  98. sme2_gemv_bf16fp32_dot_16VL/
  99. sme2_gemv_fp16fp32fp16_dot_16VL.hpp
  100. sme2_gemv_fp16fp32fp16_dot_16VL/
  101. sme2_gemv_fp32_mla_16VL.hpp
  102. sme2_gemv_fp32_mla_16VL/
  103. sme2_gemv_fp32bf16fp32_dot_16VL.hpp
  104. sme2_gemv_fp32bf16fp32_dot_16VL/
  105. sme2_gemv_s8qa_dot_16VL.hpp
  106. sme2_gemv_s8qa_dot_16VL/
  107. sme2_gemv_u8qa_dot_16VL.hpp
  108. sme2_gemv_u8qa_dot_16VL/
  109. sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp
  110. sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/
  111. sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp
  112. sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/
  113. sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp
  114. sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/
  115. sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp
  116. sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/
  117. sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp
  118. sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/
  119. sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp
  120. sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/
  121. sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp
  122. sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/
  123. sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp
  124. sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/
  125. sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp
  126. sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/
  127. sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp
  128. sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/
  129. sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp
  130. sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/
  131. sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp
  132. sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/
  133. sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL.hpp
  134. sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/
  135. sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL.hpp
  136. sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/
  137. sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL.hpp
  138. sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/
  139. sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp
  140. sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/
  141. sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp
  142. sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/
  143. sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp
  144. sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/
  145. sme2_interleaved_nomerge_u8q_mopa_1VLx4VL.hpp
  146. sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/
  147. sme2_interleaved_nomerge_u8q_mopa_2VLx2VL.hpp
  148. sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/
  149. sme2_interleaved_nomerge_u8q_mopa_4VLx1VL.hpp
  150. sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/
  151. sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp
  152. sve_ffhybrid_bf16fp32_mmla_6x4VL/
  153. sve_ffhybrid_fp16_mla_6x4VL.hpp
  154. sve_ffhybrid_fp16_mla_6x4VL/
  155. sve_ffhybrid_fp32_mla_6x4VL.hpp
  156. sve_ffhybrid_fp32_mla_6x4VL/
  157. sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
  158. sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/
  159. sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
  160. sve_ffinterleaved_bf16fp32_mmla_8x3VL/
  161. sve_ffinterleaved_fp16_mla_8x3VL.hpp
  162. sve_ffinterleaved_fp16_mla_8x3VL/
  163. sve_ffinterleaved_fp32_mla_8x3VL.hpp
  164. sve_ffinterleaved_fp32_mla_8x3VL/
  165. sve_hybrid_bf16fp32_dot_6x4VL.hpp
  166. sve_hybrid_bf16fp32_dot_6x4VL/
  167. sve_hybrid_bf16fp32_mmla_6x4VL.hpp
  168. sve_hybrid_bf16fp32_mmla_6x4VL/
  169. sve_hybrid_fp16_mla_6x4VL.hpp
  170. sve_hybrid_fp16_mla_6x4VL/
  171. sve_hybrid_fp32_mla_6x4VL.hpp
  172. sve_hybrid_fp32_mla_6x4VL/
  173. sve_hybrid_fp32_mla_8x1VL.hpp
  174. sve_hybrid_fp32_mla_8x1VL/
  175. sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp
  176. sve_hybrid_fp32bf16fp32_mmla_4x6VL/
  177. sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp
  178. sve_hybrid_fp32bf16fp32_mmla_6x4VL/
  179. sve_hybrid_s8qa_dot_4x4VL.hpp
  180. sve_hybrid_s8qa_dot_4x4VL/
  181. sve_hybrid_s8qa_mmla_4x4VL.hpp
  182. sve_hybrid_s8qa_mmla_4x4VL/
  183. sve_hybrid_s8qs_dot_6x4VL.hpp
  184. sve_hybrid_s8qs_dot_6x4VL/
  185. sve_hybrid_s8qs_mmla_6x4VL.hpp
  186. sve_hybrid_s8qs_mmla_6x4VL/
  187. sve_hybrid_s8s32_dot_6x4VL.hpp
  188. sve_hybrid_s8s32_dot_6x4VL/
  189. sve_hybrid_s8s32_mmla_6x4VL.hpp
  190. sve_hybrid_s8s32_mmla_6x4VL/
  191. sve_hybrid_u8qa_dot_4x4VL.hpp
  192. sve_hybrid_u8qa_dot_4x4VL/
  193. sve_hybrid_u8qa_mmla_4x4VL.hpp
  194. sve_hybrid_u8qa_mmla_4x4VL/
  195. sve_hybrid_u8u32_dot_6x4VL.hpp
  196. sve_hybrid_u8u32_dot_6x4VL/
  197. sve_hybrid_u8u32_mmla_6x4VL.hpp
  198. sve_hybrid_u8u32_mmla_6x4VL/
  199. sve_interleaved_bf16fp32_dot_8x3VL.hpp
  200. sve_interleaved_bf16fp32_dot_8x3VL/
  201. sve_interleaved_bf16fp32_mmla_8x3VL.hpp
  202. sve_interleaved_bf16fp32_mmla_8x3VL/
  203. sve_interleaved_fp16_mla_8x3VL.hpp
  204. sve_interleaved_fp16_mla_8x3VL/
  205. sve_interleaved_fp32_mla_8x3VL.hpp
  206. sve_interleaved_fp32_mla_8x3VL/
  207. sve_interleaved_fp32_mmla_8x3VL.hpp
  208. sve_interleaved_fp32_mmla_8x3VL/
  209. sve_interleaved_s8s32_dot_8x3VL.hpp
  210. sve_interleaved_s8s32_dot_8x3VL/
  211. sve_interleaved_s8s32_mmla_8x3VL.hpp
  212. sve_interleaved_s8s32_mmla_8x3VL/
  213. sve_interleaved_u8u32_dot_8x3VL.hpp
  214. sve_interleaved_u8u32_dot_8x3VL/
  215. sve_interleaved_u8u32_mmla_8x3VL.hpp
  216. sve_interleaved_u8u32_mmla_8x3VL/