|
| 1 | + |
| 2 | +#include "convert.hpp" |
| 3 | + |
| 4 | +#include "logger.hpp" |
| 5 | + |
| 6 | +namespace { |
| 7 | + |
| 8 | +size_t get_convert_buffer_size(const qnn::ggml_dimension_array_t & dimensions, ggml_type dst_type) { |
| 9 | + GGML_ASSERT(ggml_blck_size(dst_type) == 1); |
| 10 | + size_t nbytes = ggml_type_size(dst_type); |
| 11 | + for (size_t i = 0; i < GGML_MAX_DIMS; ++i) { |
| 12 | + nbytes *= dimensions[i]; // tight packing |
| 13 | + } |
| 14 | + |
| 15 | + return nbytes; |
| 16 | +} |
| 17 | + |
| 18 | +// from ggml_backend_blas_mul_mat, when omp available, use it otherwise will fall back to standard lib solution |
| 19 | +// TODO: remove this when we can fall back the convert to blas backend |
| 20 | +#ifdef GGML_USE_OPENMP |
| 21 | + |
| 22 | +void convert_tensor_impl(const ggml_tensor * src, int max_threads, |
| 23 | + std::shared_ptr<qnn::qnn_mem_buffer_slice> & output_buffer) { |
| 24 | + const auto ne03 = src->ne[3]; |
| 25 | + const auto ne02 = src->ne[2]; |
| 26 | + const auto ne01 = src->ne[1]; |
| 27 | + const auto ne00 = src->ne[0]; |
| 28 | + const auto ne_plane = ne01 * ne00; |
| 29 | + const auto nb03 = src->nb[3]; |
| 30 | + const auto nb02 = src->nb[2]; |
| 31 | + const auto nb01 = src->nb[1]; |
| 32 | + const int min_cols_per_thread = 4096; |
| 33 | + void * wdata = output_buffer->get_buffer(); |
| 34 | + const auto to_float = ggml_get_type_traits(src->type)->to_float; |
| 35 | + GGML_ASSERT(to_float); |
| 36 | + |
| 37 | + for (int64_t i03 = 0; i03 < ne03; i03++) { |
| 38 | + for (int64_t i02 = 0; i02 < ne02; i02++) { |
| 39 | + const void * x = (char *) src->data + i02 * nb02 + i03 * nb03; |
| 40 | + float * const wplane = (float *) wdata + i02 * ne_plane + i03 * ne02 * ne_plane; |
| 41 | + |
| 42 | + const int min_rows_per_thread = std::max((int) (min_cols_per_thread / ne00), 1); |
| 43 | + const int n_threads = std::max(std::min(max_threads, (int) (ne01 / min_rows_per_thread)), 1); |
| 44 | + |
| 45 | +# pragma omp parallel for num_threads(n_threads) |
| 46 | + for (int64_t i01 = 0; i01 < ne01; i01++) { |
| 47 | + to_float((const char *) x + i01 * nb01, wplane + i01 * ne00, ne00); |
| 48 | + } |
| 49 | + } |
| 50 | + } |
| 51 | + |
| 52 | + return output_buffer; |
| 53 | +} |
| 54 | + |
| 55 | +#else |
| 56 | + |
| 57 | +void convert_tensor_impl(const ggml_tensor * src, int max_threads, std::vector<std::future<void>> & tasks, |
| 58 | + std::shared_ptr<qnn::qnn_mem_buffer_slice> & output_buffer) { |
| 59 | + const auto ne03 = src->ne[3]; |
| 60 | + const auto ne02 = src->ne[2]; |
| 61 | + const auto ne01 = src->ne[1]; |
| 62 | + const auto ne00 = src->ne[0]; |
| 63 | + const auto ne_plane = ne01 * ne00; |
| 64 | + const auto nb03 = src->nb[3]; |
| 65 | + const auto nb02 = src->nb[2]; |
| 66 | + const auto nb01 = src->nb[1]; |
| 67 | + const int min_cols_per_thread = 4096; |
| 68 | + void * wdata = output_buffer->get_buffer(); |
| 69 | + const auto to_float = ggml_get_type_traits(src->type)->to_float; |
| 70 | + GGML_ASSERT(to_float); |
| 71 | + |
| 72 | + for (int64_t i03 = 0; i03 < ne03; i03++) { |
| 73 | + for (int64_t i02 = 0; i02 < ne02; i02++) { |
| 74 | + const void * x = (char *) src->data + i02 * nb02 + i03 * nb03; |
| 75 | + float * const wplane = (float *) wdata + i02 * ne_plane + i03 * ne02 * ne_plane; |
| 76 | + |
| 77 | + const int min_rows_per_thread = std::max((int) (min_cols_per_thread / ne00), 1); |
| 78 | + const int n_threads = std::max(std::min(max_threads, (int) (ne01 / min_rows_per_thread)), 1); |
| 79 | + |
| 80 | + for (int i = 1; i < n_threads; i++) { |
| 81 | + const int64_t start = i * ne01 / n_threads; |
| 82 | + const int64_t end = (i + 1) * ne01 / n_threads; |
| 83 | + if (start < end) { |
| 84 | + tasks.push_back(std::async(std::launch::async, [=]() { |
| 85 | + for (int64_t i01 = start; i01 < end; i01++) { |
| 86 | + to_float((const char *) x + i01 * nb01, wplane + i01 * ne00, ne00); |
| 87 | + } |
| 88 | + })); |
| 89 | + } |
| 90 | + } |
| 91 | + { |
| 92 | + // reuse the current thread for the first task |
| 93 | + const int64_t start = 0; |
| 94 | + const int64_t end = ne01 / n_threads; |
| 95 | + for (int64_t i01 = start; i01 < end; i01++) { |
| 96 | + to_float((const char *) x + i01 * nb01, wplane + i01 * ne00, ne00); |
| 97 | + } |
| 98 | + } |
| 99 | + } |
| 100 | + } |
| 101 | + |
| 102 | + // wait for all tasks to finish |
| 103 | + for (auto & task : tasks) { |
| 104 | + task.get(); |
| 105 | + } |
| 106 | + tasks.clear(); |
| 107 | +} |
| 108 | + |
| 109 | +#endif |
| 110 | + |
| 111 | +} // namespace |
| 112 | + |
| 113 | +namespace qnn { |
| 114 | + |
| 115 | +std::vector<qnn::qnn_buffer_ptr> convert(std::shared_ptr<qnn_convert_context_t> convert_context, |
| 116 | + const ggml_tensor_array_t & tensors, ggml_type target_data_type) { |
| 117 | + convert_context->buffers.resize(tensors.size()); |
| 118 | + std::vector<qnn::qnn_buffer_ptr> output_buffers(tensors.size()); |
| 119 | + for (size_t i = 0; i < tensors.size(); ++i) { |
| 120 | + const ggml_tensor * src = tensors[i]; |
| 121 | + if (src->type == target_data_type) { |
| 122 | + continue; |
| 123 | + } |
| 124 | + |
| 125 | + auto & data_buffer = convert_context->buffers[i]; |
| 126 | + const auto dst_size = get_convert_buffer_size(src->ne, target_data_type); |
| 127 | + if (!data_buffer || data_buffer->get_size() < dst_size) { |
| 128 | +#ifndef NDEBUG |
| 129 | + auto old_size = data_buffer ? data_buffer->get_size() : 0; |
| 130 | + QNN_LOG_DEBUG("create buffer[%d] for tensor %s(%s), old_size: %d, new_size: %d\n", (int) i, |
| 131 | + ggml_get_name(src), ggml_type_name(src->type), (int) old_size, (int) dst_size); |
| 132 | +#endif |
| 133 | + data_buffer = std::make_shared<qnn::qnn_mem_buffer>(dst_size); |
| 134 | + } |
| 135 | + |
| 136 | + // TODO: add more restrictions to the buffer slice here |
| 137 | + std::shared_ptr<qnn::qnn_mem_buffer_slice> output_buffer = |
| 138 | + std::make_shared<qnn::qnn_mem_buffer_slice>(data_buffer->get_buffer(), dst_size); |
| 139 | + |
| 140 | + QNN_LOG_DEBUG("convert tensor(%s) from %s to %s, size: %d, n_threads: %d\n", ggml_get_name(src), |
| 141 | + ggml_type_name(src->type), ggml_type_name(target_data_type), (int) dst_size, |
| 142 | + convert_context->n_threads); |
| 143 | + |
| 144 | +#ifdef GGML_USE_OPENMP |
| 145 | + convert_tensor_impl(src, convert_context->n_threads, output_buffer); |
| 146 | +#else |
| 147 | + convert_tensor_impl(src, convert_context->n_threads, convert_context->tasks, output_buffer); |
| 148 | +#endif |
| 149 | + output_buffers[i] = output_buffer; |
| 150 | + } |
| 151 | + |
| 152 | + return output_buffers; |
| 153 | +} |
| 154 | + |
| 155 | +} // namespace qnn |
0 commit comments