llama-cpp-sys-4 0.7.0

Low Level Bindings to llama.cpp
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383

#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#pragma OPENCL EXTENSION cl_khr_subgroups : enable

#ifdef cl_qcom_reqd_sub_group_size
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
#endif

// assume
#define QK4_0 32
#define N_SIMDGROUP 4

#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
    float shared_y; \
    shared_y = sub_group_broadcast(y.s0, 0); \
    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s1, 0); \
    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s2, 0); \
    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s3, 0); \
    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s4, 0); \
    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s5, 0); \
    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s6, 0); \
    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s7, 0); \
    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s0, 1); \
    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s1, 1); \
    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s2, 1); \
    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s3, 1); \
    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s4, 1); \
    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s5, 1); \
    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s6, 1); \
    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s7, 1); \
    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \


#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
    shared_y = sub_group_broadcast(y.s0, 2); \
    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s1, 2); \
    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s2, 2); \
    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s3, 2); \
    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s4, 2); \
    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s5, 2); \
    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s6, 2); \
    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s7, 2); \
    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s0, 3); \
    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s1, 3); \
    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s2, 3); \
    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s3, 3); \
    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s4, 3); \
    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s5, 3); \
    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s6, 3); \
    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
    shared_y = sub_group_broadcast(y.s7, 3); \
    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \


#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
    float8 shared_y; \
    shared_y = sub_group_broadcast(y, 0); \
    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
    shared_y = sub_group_broadcast(y, 1); \
    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \


#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
    shared_y = sub_group_broadcast(y, 2); \
    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
    shared_y = sub_group_broadcast(y, 3); \
    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \

#ifdef ADRENO_GPU
REQD_SUBGROUP_SIZE_64
#endif
__kernel void kernel_gemv_noshuffle_q4_0_f32(
        __read_only  image1d_buffer_t src0_q,  // quantized A
        global half2  * src0_d,  // A scales
        __read_only  image1d_buffer_t src1,    // B
        ulong offset1,            // offset to B (0)
        global float * dst,     // C
        ulong offsetd,            // offset to C (0)
        int ne00,               // K
        int ne01,               // M
        int ne02,               // 1
        int ne10,               // K
        int ne12,               // 1
        int ne0,                // M
        int ne1,                // N
        int r2,                 // 1
        int r3)
{
    uint groupId = get_local_id(1);
    uint gid     = get_global_id(0);
    ushort slid    = get_sub_group_local_id();

    uint K = ne00;
    uint M = ne01;

    uint LINE_STRIDE_A = M / 2;
    uint BLOCK_STRIDE_A = N_SIMDGROUP * M;

    __private uint4     regA;
    __private half2     regS;
    __private float8    regB;

    __private float2 totalSum = (float2)(0.0f);

    // loop along K in block granularity, skip 4 blocks every iter
    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
        // first 4 fibers in each wave load 8 B values to its private scope
        if (slid < 4) {
            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
        }

        // load half weights for two blocks in consecutive rows
        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
#ifdef VECTOR_SUB_GROUP_BROADCAST
        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
#else
        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
#endif // VECTOR_SUB_GROUP_BROADCAST

        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
#ifdef VECTOR_SUB_GROUP_BROADCAST
        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
#else
        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
#endif // VECTOR_SUB_GROUP_BROADCAST
    }

    // reduction in local memory, assumes #wave=4
    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
    barrier(CLK_LOCAL_MEM_FENCE);
    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];

    // 2 outputs per fiber in wave 0
    if (groupId == 0) {
        dst = (global float*)((global char*)dst + offsetd);
        // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
        // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
        // and would overrun dst into the adjacent tensor. No-op / byte-identical when
        // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
        if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
        if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
    }

}

// Multi-column (N in [2..4]) variant of the q4_0 decode GEMV, for the speculative
// / MTP verify batch (n_cols = 2..4 = drafted + bonus positions). Routes the small-
// batch verify OFF the transposed-GEMM dead-zone (gemm_noshuffle_q4_0) onto the
// efficient GEMV path. Each K-block's weights (regA hi+lo) are loaded ONCE and
// reused across the n_cols activation columns. Per-column accumulation is
// independent and identical to n_cols standalone GEMVs. n_cols==3 is byte-identical
// to the original mc3 (col3 disabled, slots 6/7 stay zero). Kept the _mc3 name.
#ifdef VECTOR_SUB_GROUP_BROADCAST
#define MC_DQ_HI dequantizeBlockAccum_ns_sgbroadcast_8_hi
#define MC_DQ_LO dequantizeBlockAccum_ns_sgbroadcast_8_lo
#else
#define MC_DQ_HI dequantizeBlockAccum_ns_sgbroadcast_1_hi
#define MC_DQ_LO dequantizeBlockAccum_ns_sgbroadcast_1_lo
#endif
// One column c: load this column's activation (own brace scope so the macros'
// `shared_y` decl is re-scoped), then dequant (hi+lo) against the shared weights.
#define MC_COL_Q40(ts, c) \
    { if (slid < 4) { regB.s0123 = read_imagef(src1, (c)*COL_STRIDE     + slid*2 + k*8); \
                      regB.s4567 = read_imagef(src1, (c)*COL_STRIDE + 1 + slid*2 + k*8); } \
      MC_DQ_HI(ts, as_ushort8(regA_hi), regS, regB); \
      MC_DQ_LO(ts, as_ushort8(regA_lo), regS, regB); }

#ifdef ADRENO_GPU
REQD_SUBGROUP_SIZE_64
#endif
__kernel void kernel_gemv_noshuffle_q4_0_f32_mc3(
        __read_only  image1d_buffer_t src0_q,  // quantized A
        global half2  * src0_d,                // A scales
        __read_only  image1d_buffer_t src1,    // B (n_cols columns, col-major image)
        global float * dst,                    // C (column-major [M x n_cols])
        ulong offsetd,
        int ne00,                              // K
        int ne01,                              // M
        int n_cols)                            // N (2..4)
{
    uint groupId = get_local_id(1);
    uint gid     = get_global_id(0);
    ushort slid  = get_sub_group_local_id();

    uint K = ne00;
    uint M = ne01;

    uint LINE_STRIDE_A  = M / 2;
    // BLOCK_STRIDE_A is the LAYOUT stride between consecutive K-blocks = 4 uints
    // per q4_0 block * M (set by the trans4_ns convert). The "4" is uints/block, NOT
    // the subgroup count — keep it fixed so the K-split count (nsg) can vary.
    uint BLOCK_STRIDE_A = N_SIMDGROUP * M;   // = 4 * M (N_SIMDGROUP is the #define 4)
    uint COL_STRIDE     = K / 4;   // float4 pixels per activation column
    uint nsg            = get_local_size(1);  // runtime K-split (4 default, 8 small-M)

    __private uint4  regA_hi, regA_lo;
    __private half2  regS;
    __private float8 regB;

    __private float2 ts0 = (float2)(0.0f);
    __private float2 ts1 = (float2)(0.0f);
    __private float2 ts2 = (float2)(0.0f);
    __private float2 ts3 = (float2)(0.0f);

    for (uint k = groupId; k < (K / QK4_0); k += nsg) {
        regS = src0_d[gid + k * LINE_STRIDE_A];

        // weights loaded ONCE, reused across the columns
        regA_hi.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
        regA_hi.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
        regA_hi.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
        regA_hi.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
        regA_lo.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
        regA_lo.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
        regA_lo.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
        regA_lo.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;

        MC_COL_Q40(ts0, 0);
        MC_COL_Q40(ts1, 1);
        if (n_cols > 2) MC_COL_Q40(ts2, 2);
        if (n_cols > 3) MC_COL_Q40(ts3, 3);
    }

    // cross-subgroup reduce over nsg subgroups: pack the (up to 4) columns' float2
    // into a float8. Generalized to runtime nsg (4 default, 8 for small-M). Each
    // subgroup writes its partial; subgroup 0 sums the rest into its own acc. At
    // nsg==4 this is byte-identical to the original (sums subgroups 1,2,3 in order).
    __local float8 reduceLM[SIMDGROUP_WIDTH * 8];
    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, ts3.s0, ts3.s1);
    reduceLM[groupId * SIMDGROUP_WIDTH + slid] = acc;

    barrier(CLK_LOCAL_MEM_FENCE);

    if (groupId == 0) {
        for (uint g = 1; g < nsg; g++) {
            acc += reduceLM[g * SIMDGROUP_WIDTH + slid];
        }
        dst = (global float*)((global char*)dst + offsetd);
        // dst is column-major [M rows x n_cols cols]: (row, col) at col*M + row
        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0 * M + gid * 2]));
        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1 * M + gid * 2]));
        if (n_cols > 2) vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2 * M + gid * 2]));
        if (n_cols > 3) vstore2((float2)(acc.s6, acc.s7), 0, &(dst[3 * M + gid * 2]));
    }
}
#undef MC_COL_Q40
#undef MC_DQ_HI
#undef MC_DQ_LO