llama-cpp-sys-4 0.10.0

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

#pragma clang diagnostic ignored "-Wunused-but-set-variable"

#include <HAP_farf.h>
#include <HAP_perf.h>

#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "hex-common.h"
#include "hex-profile.h"
#include "htp-ctx.h"
#include "htp-ops.h"
#include "htp-tensor.h"
#include "htp-vtcm.h"
#include "hvx-types.h"
#include "hex-utils.h"
#include "hvx-copy.h"
#include "dma-queue.h"
#include "diag-ops.h"

#define htp_diag_tensors_preamble                           \
    const struct htp_tensor * restrict src0 = octx->src[0]; \
    const struct htp_tensor * restrict dst  = octx->dst;    \
                                                            \
    const uint32_t ne02 = src0->ne[2];                      \
                                                            \
    const uint32_t ne0 = dst->ne[0];                        \
    const uint32_t ne1 = dst->ne[1];                        \
                                                            \
    const uint32_t nb02 = src0->nb[2];                      \
    const uint32_t nb03 = src0->nb[3];                      \
                                                            \
    const uint32_t nb1 = dst->nb[1];                        \
    const uint32_t nb2 = dst->nb[2];                        \
    const uint32_t nb3 = dst->nb[3];

struct htp_diag_context {
    struct htp_ops_context * octx;
    const struct htp_diag_kernel_params * kparams;
    size_t          src_batch_size;
    size_t          dst_row_size;
    size_t          src_batch_size_aligned;
    size_t          dst_row_size_aligned;
    uint32_t        batches_per_thread;
    uint32_t        total_batches;
    uint32_t        batch_start;
};

#define htp_diag_preamble                                              \
    struct htp_diag_context * dctx = (struct htp_diag_context *) data; \
    struct htp_ops_context *  octx = dctx->octx;                       \
    htp_diag_tensors_preamble;

static inline void hvx_diag_row_f32(const float * restrict src, float * restrict dst,
                                    uint32_t row_idx, uint32_t n) {
    hvx_splat_f32_a((uint8_t *) dst, 0.0f, n);
    dst[row_idx] = src[row_idx];
}

// ---------------------------------------------------------------------------
// Per thread worker: DMA src fetch, compute in VTCM, DMA dst writeback
// ---------------------------------------------------------------------------

static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data) {
    htp_diag_preamble;
    dma_queue * dma_q = octx->ctx->dma[ith];

    const uint32_t ib0 = dctx->batch_start + dctx->batches_per_thread * ith;
    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->batch_start + dctx->total_batches);

    if (ib0 >= ib1) {
        return;
    }

    const size_t src_batch_size         = dctx->src_batch_size;
    const size_t dst_row_size           = dctx->dst_row_size;
    const size_t src_batch_size_aligned = dctx->src_batch_size_aligned;
    const size_t dst_row_size_aligned   = dctx->dst_row_size_aligned;

    const dma_addr_t src_data = src0->data;
    const dma_addr_t dst_data = dst->data;

    // 1 src buffer + 1 dst row buffer per thread in VTCM
    uint8_t * src_spad = octx->src0_spad.data + (ith * src_batch_size_aligned);
    uint8_t * dst_spad = octx->dst_spad.data  + (ith * dst_row_size_aligned);

    struct htp_thread_trace * tr = &octx->ctx->trace[ith];

    for (uint32_t ib = ib0; ib < ib1; ib++) {
        const uint32_t i3 = fastdiv(ib, &dctx->kparams->div_ne02);
        const uint32_t i2 = ib - i3 * ne02;

        const dma_addr_t src_batch = src_data + i3 * nb03 + i2 * nb02;

        // Fetch source vector into VTCM
        dma_queue_push(dma_q,
                       dma_make_data(src_spad, src_batch),
                       src_batch_size_aligned, src_batch_size, src_batch_size, 1);
        dma_queue_flush(dma_q);

        const float * src_spad_f32 = (const float *) src_spad;
        float       * dst_spad_f32 = (float *) dst_spad;

        for (uint32_t i1 = 0; i1 < ne1; i1++) {
            // Compute row in VTCM
            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (ib * ne1 + i1));
            hvx_diag_row_f32(src_spad_f32, dst_spad_f32, i1, ne0);
            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (ib * ne1 + i1));

            // Write completed row back to DDR
            const dma_addr_t dst_row = dst_data + i3 * nb3 + i2 * nb2 + i1 * nb1;
            dma_queue_push(dma_q,
                           dma_make_data(dst_row, dst_spad),
                           dst_row_size, dst_row_size_aligned, dst_row_size, 1);
            dma_queue_flush(dma_q);
        }
    }

    FARF(HIGH, "diag-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
}

// ---------------------------------------------------------------------------
// Per thread worker: Direct HVX (no DMA)
// ---------------------------------------------------------------------------

static void diag_thread_f32(unsigned int nth, unsigned int ith, void * data) {
    htp_diag_preamble;

    const uint8_t * src_data = (const uint8_t *) src0->data;
    uint8_t *       dst_data = (uint8_t *) dst->data;

    const uint32_t ib0 = dctx->batch_start + dctx->batches_per_thread * ith;
    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->batch_start + dctx->total_batches);

    struct htp_thread_trace * tr = &octx->ctx->trace[ith];
    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ib0);

    for (uint32_t ib = ib0; ib < ib1; ib++) {
        const uint32_t i3 = fastdiv(ib, &dctx->kparams->div_ne02);
        const uint32_t i2 = ib - i3 * ne02;

        const float * restrict src_batch = (const float *)(src_data + i3 * nb03 + i2 * nb02);

        for (uint32_t i1 = 0; i1 < ne1; i1++) {
            float * restrict dst_row = (float *)(dst_data + i3 * nb3 + i2 * nb2 + i1 * nb1);
            hvx_diag_row_f32(src_batch, dst_row, i1, ne0);
        }
    }

    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ib0);

    FARF(HIGH, "diag-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u\n",
         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
}

int op_diag_f32(struct htp_ops_context * octx) {
    const struct htp_tensor * src0 = octx->src[0];
    const struct htp_tensor * dst  = octx->dst;
    const struct htp_diag_kernel_params * kparams =
        (const struct htp_diag_kernel_params *) octx->kernel_params;

    const uint32_t total_batches = kparams->total_batches;

    uint32_t batch_start = 0;
    uint32_t nbatches    = total_batches;

    if (octx->ctx->mdev.count > 1) {
        const bool can_split = (kparams->can_split != 0) && htp_tensor_mdev_data_aligned(dst);
        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(total_batches, can_split ? kparams->batches_per_chunk : 0, octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
        batch_start = range.start;
        nbatches    = range.count;
    }

    if (nbatches == 0) {
        return HTP_STATUS_OK;
    }

    if (!htp_ops_context_set_n_threads(octx, MIN(kparams->n_threads, nbatches))) {
        return HTP_STATUS_INVAL_PARAMS;
    }

    const uint32_t n_threads = octx->n_threads;

    if (!kparams->use_dma && (htp_tensor_is_extended(src0) || htp_tensor_is_extended(dst))) {
        return HTP_STATUS_NO_SUPPORT;
    }

    const size_t src_batch_size = src0->ne[0] * sizeof(float);
    const size_t dst_row_size   = dst->ne[0] * sizeof(float);

    struct htp_diag_vtcm_layout layout = { 0 };
    if (kparams->use_dma) {
        htp_diag_vtcm_layout_build(&layout, src_batch_size, dst_row_size, n_threads);

        octx->src0_spad.size_per_thread = layout.src0_bytes_per_thread;
        octx->dst_spad.size_per_thread  = layout.dst_bytes_per_thread;
        octx->src0_spad.size = n_threads * layout.src0_bytes_per_thread;
        octx->dst_spad.size  = n_threads * layout.dst_bytes_per_thread;
        octx->src0_spad.data = VTCM_LAYOUT_PTR(uint8_t, octx->ctx->vtcm_base, layout.off_src0);
        octx->dst_spad.data  = VTCM_LAYOUT_PTR(uint8_t, octx->ctx->vtcm_base, layout.off_dst);
        octx->src0_spad.src  = NULL;
        octx->dst_spad.src   = NULL;
    }

    struct htp_diag_context dctx = {
        .octx                   = octx,
        .kparams                = kparams,
        .src_batch_size         = src_batch_size,
        .dst_row_size           = dst_row_size,
        .src_batch_size_aligned = layout.src_batch_size_aligned,
        .dst_row_size_aligned   = layout.dst_row_size_aligned,
        .batches_per_thread     = fastdiv(nbatches + n_threads - 1, &octx->n_threads_div),
        .total_batches          = nbatches,
        .batch_start            = batch_start,
    };

    if (kparams->use_dma) {
        work_queue_run(octx->ctx->work_queue, diag_thread_f32_dma, &dctx, n_threads);
    } else {
        work_queue_run(octx->ctx->work_queue, diag_thread_f32, &dctx, n_threads);
    }

    return HTP_STATUS_OK;
}

int op_diag(struct htp_ops_context * octx) {
    const struct htp_tensor * dst = octx->dst;

    int err = HTP_STATUS_OK;

    switch (dst->type) {
        case HTP_TYPE_F32:
            err = op_diag_f32(octx);
            break;
        default:
            err = HTP_STATUS_NO_SUPPORT;
            break;
    }

    return err;
}