llama-cpp-sys-4 0.3.2

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

//
// MIT license
// Copyright (C) 2026 Intel Corporation
// SPDX-License-Identifier: MIT
//

//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//

#include "pool.hpp"
#include <float.h>

template <typename Ti, typename To>
static void pool2d_nchw_kernel(
        const int ih, const int iw, const int oh, const int ow,
        const int kh, const int kw, const int sh, const int sw,
        const int ph, const int pw, const int parallel_elements,
        const Ti* src, To* dst, const enum ggml_op_pool op,
        const sycl::nd_item<3> &item_ct1) {
        int idx = item_ct1.get_local_id(2) +
                  item_ct1.get_group(2) * item_ct1.get_local_range(2);
        if (idx >= parallel_elements) {
            return;
        }

        const int I_HW = ih * iw;
        const int O_HW = oh * ow;
        const int nc = idx / O_HW;
        const int cur_oh = idx % O_HW / ow;
        const int cur_ow = idx % O_HW % ow;
        const Ti* i_ptr = src + nc * I_HW;
        To* o_ptr = dst + nc * O_HW;
        const int start_h = cur_oh * sh - ph;
        const int bh = sycl::max(0, start_h);
        const int eh = sycl::min(ih, start_h + kh);
        const int start_w = cur_ow * sw - pw;
        const int bw = sycl::max(0, start_w);
        const int ew = sycl::min(iw, start_w + kw);

        To res = 0;

        switch (op) {
            case GGML_OP_POOL_AVG: res = 0; break;
            case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
            default:
                res      = (To) sycl::nan(uint32_t(0));
                break;
        }

        for (int i = bh; i < eh; i += 1) {
            for (int j = bw; j < ew; j += 1) {
                Ti cur = i_ptr[i * iw + j];
                switch (op) {
                    case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
                    case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
                    default:
                        res = (To) sycl::nan(uint32_t(0));
                        break;
                }
            }
        }
        o_ptr[cur_oh * ow + cur_ow] = res;
}

template <typename Ti, typename To>
static void pool1d_ncw_kernel(
        const int iw, const int ow,
        const int k, const int s,
        const int p, const int parallel_elements,
        const Ti * src, To * dst, const enum ggml_op_pool op,
        const sycl::nd_item<3> & item_ct1) {
    int idx = item_ct1.get_local_id(2) +
              item_ct1.get_group(2) * item_ct1.get_local_range(2);
    if (idx >= parallel_elements) {
        return;
    }

    const int nc     = idx / ow;
    const int cur_ow = idx % ow;
    const Ti * i_ptr = src + nc * iw;
    To *       o_ptr = dst + nc * ow;
    const int  start = cur_ow * s - p;
    const int  b     = sycl::max(0, start);
    const int  e     = sycl::min(iw, start + k);

    To res = 0;
    switch (op) {
        case GGML_OP_POOL_AVG: res = 0;        break;
        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
        default:
            res = (To) sycl::nan(uint32_t(0));
            break;
    }

    for (int j = b; j < e; j += 1) {
        Ti cur = i_ptr[j];
        switch (op) {
            case GGML_OP_POOL_AVG: res += cur; break;
            case GGML_OP_POOL_MAX: res = sycl::max(res, (To) cur); break;
            default:
                res = (To) sycl::nan(uint32_t(0));
                break;
        }
    }

    const int count = e - b;
    if (op == GGML_OP_POOL_AVG) {
        res = (count > 0) ? (res / count) : (To) 0;
    }
    o_ptr[cur_ow] = res;
}

void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
    GGML_ASSERT( dst->type == GGML_TYPE_F32);
    dpct::queue_ptr main_stream = ctx.stream();
    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
    float *       dst_dd  = static_cast<float *>(dst->data);

    const int32_t * opts = (const int32_t *)dst->op_params;
    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
    const int k0 = opts[1];
    const int k1 = opts[2];
    const int s0 = opts[3];
    const int s1 = opts[4];
    const int p0 = opts[5];
    const int p1 = opts[6];

    const int64_t IH = dst->src[0]->ne[1];
    const int64_t IW = dst->src[0]->ne[0];

    const int64_t N = dst->ne[3];
    const int64_t OC = dst->ne[2];
    const int64_t OH = dst->ne[1];
    const int64_t OW = dst->ne[0];

    const int parallel_elements = N * OC * OH * OW;
    const int num_blocks = (parallel_elements + SYCL_POOL2D_BLOCK_SIZE - 1) / SYCL_POOL2D_BLOCK_SIZE;
    sycl::range<3> block_nums(1, 1, num_blocks);
    main_stream->parallel_for(
        sycl::nd_range<3>(block_nums *
                              sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE),
                          sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)),
        [=](sycl::nd_item<3> item_ct1) {
            pool2d_nchw_kernel(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0,
                               parallel_elements, src0_dd, dst_dd, op,
                               item_ct1);
        });
}

void ggml_sycl_op_pool1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
    GGML_ASSERT( dst->type == GGML_TYPE_F32);
    dpct::queue_ptr main_stream = ctx.stream();
    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
    float *       dst_dd  = static_cast<float *>(dst->data);

    const int32_t * opts = (const int32_t *)dst->op_params;
    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
    const int k0 = opts[1];
    const int s0 = opts[2];
    const int p0 = opts[3];

    const int64_t IW = dst->src[0]->ne[0];
    const int64_t OW = dst->ne[0];
    const int64_t NC = dst->ne[3] * dst->ne[2] * dst->ne[1];

    const int parallel_elements = NC * OW;
    const int num_blocks = (parallel_elements + SYCL_POOL1D_BLOCK_SIZE - 1) / SYCL_POOL1D_BLOCK_SIZE;
    sycl::range<3> block_nums(1, 1, num_blocks);
    main_stream->parallel_for(
        sycl::nd_range<3>(block_nums *
                              sycl::range<3>(1, 1, SYCL_POOL1D_BLOCK_SIZE),
                          sycl::range<3>(1, 1, SYCL_POOL1D_BLOCK_SIZE)),
        [=](sycl::nd_item<3> item_ct1) {
            pool1d_ncw_kernel(IW, OW, k0, s0, p0,
                              parallel_elements, src0_dd, dst_dd, op,
                              item_ct1);
        });
}