summaryrefslogtreecommitdiff
path: root/xmake/modules/lib/detect/find_cudadevices.lua
blob: 71ee12e3b71aafbf551386fa4929915b81a2054d (plain)
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
--!A cross-platform build utility based on Lua
--
-- Licensed under the Apache License, Version 2.0 (the "License");
-- you may not use this file except in compliance with the License.
-- You may obtain a copy of the License at
--
--     http://www.apache.org/licenses/LICENSE-2.0
--
-- Unless required by applicable law or agreed to in writing, software
-- distributed under the License is distributed on an "AS IS" BASIS,
-- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-- See the License for the specific language governing permissions and
-- limitations under the License.
--
-- Copyright (C) 2015-present, TBOOX Open Source Group.
--
-- @author      OpportunityLiu
-- @file        find_cudadevices.lua
--

-- imports
import("core.base.option")
import("core.platform.platform")
import("core.project.config")
import("core.cache.detectcache")
import("lib.detect.find_tool")

-- a magic string to filter output
local _PRINT_SUFFIX = "<find_cudadevices>"

-- filter stdout and stderr with _PRINT_SUFFIX
function _get_lines(str)
    local result = {}
    for _, l in ipairs(str:split("\n")) do
        if l:startswith(_PRINT_SUFFIX) then
            table.insert(result, l:sub(#_PRINT_SUFFIX + 1))
        end
    end
    return result
end

-- parse a single value
--
-- format:
-- 1. a number:     `2048`
-- 2. an array:     `(65536, 2048, 2048)`
-- 3. bool value:   `true` or `false`
-- 4. string:       `"string"`
--
function _parse_value(value)
    local num = tonumber(value)
    if num then return num end

    if value:lower() == "true" then return true end
    if value:lower() == "false" then return false end
    if value:startswith('"') and value:endswith('"') then
        return value:sub(2, -2)
    end

    if value:startswith("(") and value:endswith(")") then
        local values = value:sub(2, -2):split(",")
        local result = {}
        for _, v in ipairs(values) do
            table.insert(result, _parse_value(v:trim()))
        end
        return result
    end

    raise("don't know how to parse value: %s", value)
end


-- parse single line
--
-- format:
--     key = value
--
function _parse_line(line, device)
    local key = line:match("%s+(%g+) = .+")
    local value = line:match("%s+%g+ = (.+)")
    if key and value then
        key = key:trim()
        value = value:trim()
        assert(not device[key], "duplicate key: " .. key)
        device[key] = _parse_value(value)
    end
end

-- parse filtered lines
function _parse_result(lines, verbose)

    if #lines == 0 then
        -- not a failure, returns {} rather than nil
        utils.warning("no cuda devices was found")
        return {}
    end

    local devices = {}
    local currentDevice = nil
    for _, l in ipairs(lines) do
        if verbose then
            cprint("${dim}> %s", l)
        end
        local devId = tonumber(l:match("%s*DEVICE #(%d+)"))
        if devId then
            currentDevice = { ["$id"] = devId }
            table.insert(devices, currentDevice)
        elseif currentDevice then
            _parse_line(l, currentDevice)
        end
    end
    return devices
end

-- find devices
function _find_devices(verbose, opt)

    -- find nvcc
    local nvcc = assert(find_tool("nvcc"), "nvcc not found")

    -- trace
    if verbose then
        cprint("${dim}checking for cuda devices")
    end

    -- get cuda devices
    local sourcefile = path.join(os.programdir(), "scripts", "find_cudadevices.cpp")
    local outfile = os.tmpfile({ramdisk = false}) -- no execution permision in docker's /shm
    local compile_errors = nil
    local results, errors = try
    {
        function ()
            local args = { sourcefile, "-run", "-o", outfile , '-DPRINT_SUFFIX="' .. _PRINT_SUFFIX .. '"' }
            if opt.arch == "x86" then
                table.insert(args, "-m32")
            elseif opt.arch == "x64" or opt.arch == "x86_64" then
                table.insert(args, "-m64")
            end
            return os.iorunv(nvcc.program, args, {envs = opt.envs})
        end,
        catch
        {
            function (errs)
                compile_errors = tostring(errs)
            end
        }
    }

    if compile_errors then
        if not option.get("diagnosis") then
            compile_errors = compile_errors:split('\n')[1]
        end
        utils.warning("failed to find cuda devices: " .. compile_errors)
        return nil
    end

    -- clean up
    os.tryrm(outfile)
    os.tryrm(outfile .. ".*")

    -- get results
    local results_lines = _get_lines(results)
    local errors_lines = _get_lines(errors)
    if #errors_lines ~= 0 then
        utils.warning("failed to find cuda devices: " .. table.concat(errors_lines, "\n"))
        return nil
    end

    -- print raw result only with -D flags
    local devices = _parse_result(results_lines, option.get("diagnosis"))
    if verbose then
        for _, v in ipairs(devices) do
            cprint("${dim}> found device #%d: ${green bright}%s${reset dim} with compute ${bright}%d.%d${reset dim} capability", v["$id"], v.name, v.major, v.minor)
        end
    end
    return devices
end

-- get devices array form cache or via _find_devices
function _get_devices(opt)

    -- init cachekey
    local cachekey = "find_cudadevices"
    if opt.cachekey then
        cachekey = cachekey .. "_" .. opt.cachekey
    end

    -- check cache
    local cachedata = detectcache:get(cachekey) or {}
    if cachedata.succeed and not opt.force then
        return cachedata.data
    end

    local verbose = opt.verbose or option.get("verbose") or option.get("diagnosis")
    local devices = _find_devices(verbose, opt)
    if devices then
        cachedata = { succeed = true, data = devices }
    else
        cachedata = { succeed = false }
        devices = {}
    end

    -- fill cache
    detectcache:set(cachekey, cachedata)
    detectcache:save()
    return devices
end

function _skip_compute_mode_prohibited(devices)
    local results = {}
    local cuda_compute_mode_prohibited = 2
    for _, dev in ipairs(devices) do
        if dev.computeMode ~= cuda_compute_mode_prohibited then
            table.insert(results, dev)
        end
    end
    return results
end

function _min_sm_arch(devices, min_sm_arch)
    local results = {}
    for _, dev in ipairs(devices) do
        if dev.major * 10 + dev.minor >= min_sm_arch then
            table.insert(results, dev)
        end
    end
    return results
end

function _order_by_flops(devices)

    local ngpu_arch_cores_per_sm =
    {
        [30] =    192
    ,   [32] =    192
    ,   [35] =    192
    ,   [37] =    192
    ,   [50] =    128
    ,   [52] =    128
    ,   [53] =    128
    ,   [60] =     64
    ,   [61] =    128
    ,   [62] =    128
    ,   [70] =     64
    ,   [72] =     64
    ,   [75] =     64
    ,   [80] =     64
    ,   [86] =    128
    ,   [87] =    128
    }

    for _, dev in ipairs(devices) do
        local sm_per_multiproc = 0
        if dev.major == 9999 and dev.minor == 9999 then
            sm_per_multiproc = 1
        else
            sm_per_multiproc = ngpu_arch_cores_per_sm[dev.major * 10 + dev.minor] or 64;
        end
        dev["$flops"] = dev.multiProcessorCount * sm_per_multiproc * dev.clockRate
    end

    table.sort(devices, function (a,b) return a["$flops"] > b["$flops"] end)
    return devices
end

-- find cuda devices of the host
--
-- @param opt   the options
--              e.g. { verbose = false, force = false, cachekey = "xxxx", min_sm_arch = 35, skip_compute_mode_prohibited = false, order_by_flops = true }
--
-- @return      { { ["$id"] = 0, name = "GeForce GTX 960M", major = 5, minor = 0, ... }, ... }
--              for all keys, see https://docs.nvidia.com/cuda/cuda-runtime-api/structcudaDeviceProp.html#structcudaDeviceProp
--              keys might be differ as your cuda version varies
--
function main(opt)

    -- init options
    opt = opt or {}

    -- get devices
    local devices = _get_devices(opt)

    -- apply filters
    if opt.min_sm_arch then
        devices = _min_sm_arch(devices, opt.min_sm_arch)
    end
    if opt.skip_compute_mode_prohibited then
        devices = _skip_compute_mode_prohibited(devices)
    end
    if opt.order_by_flops then
        devices = _order_by_flops(devices)
    end

    return devices
end