|
36 | 36 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
37 | 37 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
38 | 38 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
39 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
40 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
41 | 39 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
42 | 40 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
43 | 41 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
|
49 | 47 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
50 | 48 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
51 | 49 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
52 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
53 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
54 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
55 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
56 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
57 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
58 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
59 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 50 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 51 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
60 | 52 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
61 | 53 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
62 | 54 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
|
68 | 60 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
69 | 61 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
70 | 62 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
71 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
72 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
73 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
74 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
75 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
76 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
77 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
78 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 63 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 64 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
79 | 65 | #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) |
80 | 66 | // repack.cpp |
81 | 67 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
82 | 68 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
83 | 69 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
84 | 70 | #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K |
85 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
86 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
87 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
88 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
89 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
90 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
91 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
92 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
93 | 71 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
94 | 72 | #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K |
95 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
96 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
97 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
98 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
99 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
100 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
101 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
102 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
103 | 73 | #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64) |
104 | 74 | // repack.cpp |
105 | 75 | #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 |
106 | 76 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
107 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
108 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
109 | 77 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
110 | 78 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
111 | 79 | #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K |
112 | 80 | #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K |
113 | 81 | #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K |
114 | 82 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
115 | 83 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
116 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
117 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
118 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
119 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
120 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
121 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
122 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
123 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 84 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 85 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
124 | 86 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
125 | 87 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
126 | 88 | #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K |
127 | 89 | #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K |
128 | 90 | #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K |
129 | 91 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
130 | 92 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
131 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
132 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
133 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
134 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
135 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
136 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
137 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
138 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 93 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 94 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
139 | 95 | #elif defined(__POWERPC__) || defined(__powerpc__) |
140 | 96 | // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679 |
141 | 97 | // quants.c |
|
148 | 104 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
149 | 105 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
150 | 106 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
151 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
152 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
153 | 107 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
154 | 108 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
155 | 109 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
|
161 | 115 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
162 | 116 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
163 | 117 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
164 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
165 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
166 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
167 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
168 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
169 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
170 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
171 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 118 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 119 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
172 | 120 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
173 | 121 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
174 | 122 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
|
180 | 128 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
181 | 129 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
182 | 130 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
183 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
184 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
185 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
186 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
187 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
188 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
189 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
190 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 131 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 132 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
191 | 133 | #elif defined(__loongarch64) |
192 | 134 | // quants.c |
193 | 135 | #define quantize_row_q8_K_generic quantize_row_q8_K |
|
200 | 142 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
201 | 143 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
202 | 144 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
203 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
204 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
205 | 145 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
206 | 146 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
207 | 147 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
|
213 | 153 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
214 | 154 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
215 | 155 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
216 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
217 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
218 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
219 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
220 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
221 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
222 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
223 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 156 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 157 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
224 | 158 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
225 | 159 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
226 | 160 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
|
232 | 166 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
233 | 167 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
234 | 168 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
235 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
236 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
237 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
238 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
239 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
240 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
241 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
242 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 169 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 170 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
243 | 171 | #elif defined(__riscv) |
244 | 172 | // quants.c |
245 | 173 | #define quantize_row_q8_K_generic quantize_row_q8_K |
|
304 | 232 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
305 | 233 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
306 | 234 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
307 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
308 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
309 | 235 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
310 | 236 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
311 | 237 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
|
317 | 243 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
318 | 244 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
319 | 245 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
320 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
321 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
322 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
323 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
324 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
325 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
326 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
327 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 246 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 247 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
328 | 248 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
329 | 249 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
330 | 250 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
|
336 | 256 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
337 | 257 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
338 | 258 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
339 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
340 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
341 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
342 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
343 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
344 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
345 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
346 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 259 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 260 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
347 | 261 | #elif defined(__wasm__) |
348 | 262 | // quants.c |
349 | 263 | #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1 |
|
364 | 278 | #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8 |
365 | 279 | #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 |
366 | 280 | #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 |
367 | | -#define ggml_repack_mat_f16_7x1_generic ggml_repack_mat_f16_7x1 |
368 | | -#define ggml_repack_mat_f32_7x1_generic ggml_repack_mat_f32_7x1 |
369 | 281 | #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0 |
370 | 282 | #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 |
371 | 283 | #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 |
|
377 | 289 | #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K |
378 | 290 | #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 |
379 | 291 | #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 |
380 | | -#define ggml_gemv_f16_1x16_f16_generic ggml_gemv_f16_1x16_f16 |
381 | | -#define ggml_gemv_f16_1x32_f16_generic ggml_gemv_f16_1x32_f16 |
382 | | -#define ggml_gemv_f16_1x64_f16_generic ggml_gemv_f16_1x64_f16 |
383 | | -#define ggml_gemv_f16_1x128_f16_generic ggml_gemv_f16_1x128_f16 |
384 | | -#define ggml_gemv_f32_1x16_f32_generic ggml_gemv_f32_1x16_f32 |
385 | | -#define ggml_gemv_f32_1x32_f32_generic ggml_gemv_f32_1x32_f32 |
386 | | -#define ggml_gemv_f32_1x64_f32_generic ggml_gemv_f32_1x64_f32 |
387 | | -#define ggml_gemv_f32_1x128_f32_generic ggml_gemv_f32_1x128_f32 |
| 292 | +#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 |
| 293 | +#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 |
388 | 294 | #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 |
389 | 295 | #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 |
390 | 296 | #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 |
|
396 | 302 | #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K |
397 | 303 | #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 |
398 | 304 | #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 |
399 | | -#define ggml_gemm_f16_7x1x16_f16_generic ggml_gemm_f16_7x1x16_f16 |
400 | | -#define ggml_gemm_f16_7x1x32_f16_generic ggml_gemm_f16_7x1x32_f16 |
401 | | -#define ggml_gemm_f16_7x1x64_f16_generic ggml_gemm_f16_7x1x64_f16 |
402 | | -#define ggml_gemm_f16_7x1x128_f16_generic ggml_gemm_f16_7x1x128_f16 |
403 | | -#define ggml_gemm_f32_7x1x16_f32_generic ggml_gemm_f32_7x1x16_f32 |
404 | | -#define ggml_gemm_f32_7x1x32_f32_generic ggml_gemm_f32_7x1x32_f32 |
405 | | -#define ggml_gemm_f32_7x1x64_f32_generic ggml_gemm_f32_7x1x64_f32 |
406 | | -#define ggml_gemm_f32_7x1x128_f32_generic ggml_gemm_f32_7x1x128_f32 |
| 305 | +#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 |
| 306 | +#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 |
407 | 307 | #endif |
0 commit comments