Debian Codegraph @

origami/python/origami_module.cpp

raw
1 // Copyright Advanced Micro Devices, Inc., or its affiliates.
2 // SPDX-License-Identifier: MIT
3
4 #include "origami/hardware.hpp"
5 #include "origami/streamk.hpp"
6 #include "origami/utils.hpp"
7
8 #include <nanobind/nanobind.h>
9 #include <nanobind/stl/string.h>
10 #include <nanobind/stl/vector.h>
11 #include <nanobind/stl/map.h>
12 #include <nanobind/stl/unordered_map.h>
13 #include <nanobind/stl/pair.h>
14 #include <nanobind/stl/tuple.h>
15
16 using hardware_t = origami::hardware_t;
17
18 NB_MODULE(origami, m)
19 {
20 nanobind::enum_<hardware_t::architecture_t>(m, "architecture_t")
21 .value("gfx942", hardware_t::architecture_t::gfx942)
22 .value("gfx950", hardware_t::architecture_t::gfx950)
23 .export_values();
24
25 nanobind::enum_<origami::data_type_t>(m, "data_type_t")
26 .value("Float", origami::data_type_t::Float)
27 .value("ComplexFloat", origami::data_type_t::ComplexFloat)
28 .value("ComplexDouble", origami::data_type_t::ComplexDouble)
29 .value("Double", origami::data_type_t::Double)
30 .value("Half", origami::data_type_t::Half)
31 .value("Int8x4", origami::data_type_t::Int8x4)
32 .value("Int32", origami::data_type_t::Int32)
33 .value("BFloat16", origami::data_type_t::BFloat16)
34 .value("Int8", origami::data_type_t::Int8)
35 .value("Int64", origami::data_type_t::Int64)
36 .value("XFloat32", origami::data_type_t::XFloat32)
37 .value("Float8_fnuz", origami::data_type_t::Float8_fnuz)
38 .value("BFloat8_fnuz", origami::data_type_t::BFloat8_fnuz)
39 .value("Float8BFloat8_fnuz", origami::data_type_t::Float8BFloat8_fnuz)
40 .value("BFloat8Float8_fnuz", origami::data_type_t::BFloat8Float8_fnuz)
41 .value("Float8", origami::data_type_t::Float8)
42 .value("BFloat8", origami::data_type_t::BFloat8)
43 .value("Float8BFloat8", origami::data_type_t::Float8BFloat8)
44 .value("BFloat8Float8", origami::data_type_t::BFloat8Float8)
45 .value("Float6", origami::data_type_t::Float6)
46 .value("BFloat6", origami::data_type_t::BFloat6)
47 .value("Float4", origami::data_type_t::Float4)
48 .export_values();
49
50 m.def("int_to_data_type",
51 &origami::int_to_data_type,
52 "Convert int to data_type_t.");
53
54 nanobind::enum_<origami::streamk::reduction_type>(m, "reduction_type")
55 .value("Tree", origami::streamk::reduction_type::Tree)
56 .value("Parallel", origami::streamk::reduction_type::Parallel)
57 .export_values();
58
59 m.def("int_to_reduction_type",
60 &origami::streamk::int_to_reduction_type,
61 "Convert int to reduction_type.");
62
63 nanobind::class_<hardware_t>(m, "hardware_t")
64 .def(nanobind::init<hardware_t::architecture_t,
65 size_t,
66 size_t,
67 size_t,
68 double,
69 double,
70 double,
71 size_t,
72 double,
73 size_t,
74 std::tuple<double, double, double>>())
75 .def("print", &hardware_t::print)
76 .def("print_debug_info", &hardware_t::print_debug_info)
77 .def_rw("N_CU", &hardware_t::N_CU)
78 .def_rw("lds_capacity", &hardware_t::lds_capacity)
79 .def_rw("mem1_perf_ratio", &hardware_t::mem1_perf_ratio)
80 .def_rw("mem2_perf_ratio", &hardware_t::mem2_perf_ratio)
81 .def_rw("mem3_perf_ratio", &hardware_t::mem3_perf_ratio)
82 .def_rw("L2_capacity", &hardware_t::L2_capacity)
83 .def_rw("CU_per_L2", &hardware_t::CU_per_L2)
84 .def_rw("compute_clock_ghz", &hardware_t::compute_clock_ghz)
85 .def_rw("parallel_mi_cu", &hardware_t::parallel_mi_cu)
86 .def_rw("mem_bw_per_wg_coefficients", &hardware_t::mem_bw_per_wg_coefficients)
87 .def_rw("NUM_XCD", &hardware_t::NUM_XCD);
88
89 m.def("get_hardware_for_device",
90 &hardware_t::get_hardware_for_device,
91 "This gets a hardware object for a device.");
92
93 m.def("datatype_to_bits", &origami::data_type_to_bits, "Return the number of bits in a datatype");
94 m.def("string_to_datatype", &origami::string_to_data_type, "Convert a string representation of a datatype into data_type_t enum");
95 m.def("select_best_macro_tile_size",
96 &origami::select_best_macro_tile_size,
97 "Get best macro tile sizes.");
98 m.def("select_reduction", &origami::streamk::select_reduction, "Select best StreamK reduction strategy");
99 m.def("select_grid", &origami::streamk::select_grid, "Select Best StreamK Grid Size");
100 m.def("compute_total_latency", &origami::compute_total_latency, "compute_total_latency");
101 m.def("select_best_wgm", &origami::select_best_wgm, "Get best workgroup mapping.");
102 }