src/mesa/drivers/dri/i965/brw_oa.py

   1 # Copyright (c) 2015-2017 Intel Corporation
   2 #
   3 # Permission is hereby granted, free of charge, to any person obtaining a
   4 # copy of this software and associated documentation files (the "Software"),
   5 # to deal in the Software without restriction, including without limitation
   6 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
   7 # and/or sell copies of the Software, and to permit persons to whom the
   8 # Software is furnished to do so, subject to the following conditions:
   9 #
  10 # The above copyright notice and this permission notice (including the next
  11 # paragraph) shall be included in all copies or substantial portions of the
  12 # Software.
  13 #
  14 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  18 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  19 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  20 # IN THE SOFTWARE.
  21
  22 import argparse
  23 import os
  24 import sys
  25 import textwrap
  26
  27 import xml.etree.cElementTree as et
  28
  29 max_values = {}
  30 read_funcs = {}
  31
  32 c_file = None
  33 _c_indent = 0
  34
  35 def c(*args):
  36     if c_file:
  37         code = ' '.join(map(str,args))
  38         for line in code.splitlines():
  39             text = ''.rjust(_c_indent) + line
  40             c_file.write(text.rstrip() + "\n")
  41
  42 # indented, but no trailing newline...
  43 def c_line_start(code):
  44     if c_file:
  45         c_file.write(''.rjust(_c_indent) + code)
  46 def c_raw(code):
  47     if c_file:
  48         c_file.write(code)
  49
  50 def c_indent(n):
  51     global _c_indent
  52     _c_indent = _c_indent + n
  53 def c_outdent(n):
  54     global _c_indent
  55     _c_indent = _c_indent - n
  56
  57 header_file = None
  58 _h_indent = 0
  59
  60 def h(*args):
  61     if header_file:
  62         code = ' '.join(map(str,args))
  63         for line in code.splitlines():
  64             text = ''.rjust(_h_indent) + line
  65             header_file.write(text.rstrip() + "\n")
  66
  67 def h_indent(n):
  68     global _c_indent
  69     _h_indent = _h_indent + n
  70 def h_outdent(n):
  71     global _c_indent
  72     _h_indent = _h_indent - n
  73
  74
  75 def emit_fadd(tmp_id, args):
  76     c("double tmp{0} = {1} + {2};".format(tmp_id, args[1], args[0]))
  77     return tmp_id + 1
  78
  79 # Be careful to check for divide by zero...
  80 def emit_fdiv(tmp_id, args):
  81     c("double tmp{0} = {1};".format(tmp_id, args[1]))
  82     c("double tmp{0} = {1};".format(tmp_id + 1, args[0]))
  83     c("double tmp{0} = tmp{1} ? tmp{2} / tmp{1} : 0;".format(tmp_id + 2, tmp_id + 1, tmp_id))
  84     return tmp_id + 3
  85
  86 def emit_fmax(tmp_id, args):
  87     c("double tmp{0} = {1};".format(tmp_id, args[1]))
  88     c("double tmp{0} = {1};".format(tmp_id + 1, args[0]))
  89     c("double tmp{0} = MAX(tmp{1}, tmp{2});".format(tmp_id + 2, tmp_id, tmp_id + 1))
  90     return tmp_id + 3
  91
  92 def emit_fmul(tmp_id, args):
  93     c("double tmp{0} = {1} * {2};".format(tmp_id, args[1], args[0]))
  94     return tmp_id + 1
  95
  96 def emit_fsub(tmp_id, args):
  97     c("double tmp{0} = {1} - {2};".format(tmp_id, args[1], args[0]))
  98     return tmp_id + 1
  99
 100 def emit_read(tmp_id, args):
 101     type = args[1].lower()
 102     c("uint64_t tmp{0} = accumulator[query->{1}_offset + {2}];".format(tmp_id, type, args[0]))
 103     return tmp_id + 1
 104
 105 def emit_uadd(tmp_id, args):
 106     c("uint64_t tmp{0} = {1} + {2};".format(tmp_id, args[1], args[0]))
 107     return tmp_id + 1
 108
 109 # Be careful to check for divide by zero...
 110 def emit_udiv(tmp_id, args):
 111     c("uint64_t tmp{0} = {1};".format(tmp_id, args[1]))
 112     c("uint64_t tmp{0} = {1};".format(tmp_id + 1, args[0]))
 113     c("uint64_t tmp{0} = tmp{1} ? tmp{2} / tmp{1} : 0;".format(tmp_id + 2, tmp_id + 1, tmp_id))
 114     return tmp_id + 3
 115
 116 def emit_umul(tmp_id, args):
 117     c("uint64_t tmp{0} = {1} * {2};".format(tmp_id, args[1], args[0]))
 118     return tmp_id + 1
 119
 120 def emit_usub(tmp_id, args):
 121     c("uint64_t tmp{0} = {1} - {2};".format(tmp_id, args[1], args[0]))
 122     return tmp_id + 1
 123
 124 def emit_umin(tmp_id, args):
 125     c("uint64_t tmp{0} = MIN({1}, {2});".format(tmp_id, args[1], args[0]))
 126     return tmp_id + 1
 127
 128 ops = {}
 129 #             (n operands, emitter)
 130 ops["FADD"] = (2, emit_fadd)
 131 ops["FDIV"] = (2, emit_fdiv)
 132 ops["FMAX"] = (2, emit_fmax)
 133 ops["FMUL"] = (2, emit_fmul)
 134 ops["FSUB"] = (2, emit_fsub)
 135 ops["READ"] = (2, emit_read)
 136 ops["UADD"] = (2, emit_uadd)
 137 ops["UDIV"] = (2, emit_udiv)
 138 ops["UMUL"] = (2, emit_umul)
 139 ops["USUB"] = (2, emit_usub)
 140 ops["UMIN"] = (2, emit_umin)
 141
 142 def brkt(subexp):
 143     if " " in subexp:
 144         return "(" + subexp + ")"
 145     else:
 146         return subexp
 147
 148 def splice_bitwise_and(args):
 149     return brkt(args[1]) + " & " + brkt(args[0])
 150
 151 def splice_logical_and(args):
 152     return brkt(args[1]) + " && " + brkt(args[0])
 153
 154 def splice_ult(args):
 155     return brkt(args[1]) + " < " + brkt(args[0])
 156
 157 def splice_ugte(args):
 158     return brkt(args[1]) + " >= " + brkt(args[0])
 159
 160 exp_ops = {}
 161 #                 (n operands, splicer)
 162 exp_ops["AND"]  = (2, splice_bitwise_and)
 163 exp_ops["UGTE"] = (2, splice_ugte)
 164 exp_ops["ULT"]  = (2, splice_ult)
 165 exp_ops["&&"]   = (2, splice_logical_and)
 166
 167
 168 hw_vars = {}
 169 hw_vars["$EuCoresTotalCount"] = "brw->perfquery.sys_vars.n_eus"
 170 hw_vars["$EuSlicesTotalCount"] = "brw->perfquery.sys_vars.n_eu_slices"
 171 hw_vars["$EuSubslicesTotalCount"] = "brw->perfquery.sys_vars.n_eu_sub_slices"
 172 hw_vars["$EuThreadsCount"] = "brw->perfquery.sys_vars.eu_threads_count"
 173 hw_vars["$SliceMask"] = "brw->perfquery.sys_vars.slice_mask"
 174 hw_vars["$SubsliceMask"] = "brw->perfquery.sys_vars.subslice_mask"
 175 hw_vars["$GpuTimestampFrequency"] = "brw->perfquery.sys_vars.timestamp_frequency"
 176 hw_vars["$GpuMinFrequency"] = "brw->perfquery.sys_vars.gt_min_freq"
 177 hw_vars["$GpuMaxFrequency"] = "brw->perfquery.sys_vars.gt_max_freq"
 178
 179 def output_rpn_equation_code(set, counter, equation, counter_vars):
 180     c("/* RPN equation: " + equation + " */")
 181     tokens = equation.split()
 182     stack = []
 183     tmp_id = 0
 184     tmp = None
 185
 186     for token in tokens:
 187         stack.append(token)
 188         while stack and stack[-1] in ops:
 189             op = stack.pop()
 190             argc, callback = ops[op]
 191             args = []
 192             for i in range(0, argc):
 193                 operand = stack.pop()
 194                 if operand[0] == "$":
 195                     if operand in hw_vars:
 196                         operand = hw_vars[operand]
 197                     elif operand in counter_vars:
 198                         reference = counter_vars[operand]
 199                         operand = read_funcs[operand[1:]] + "(brw, query, accumulator)"
 200                     else:
 201                         raise Exception("Failed to resolve variable " + operand + " in equation " + equation + " for " + set.get('name') + " :: " + counter.get('name'));
 202                 args.append(operand)
 203
 204             tmp_id = callback(tmp_id, args)
 205
 206             tmp = "tmp{0}".format(tmp_id - 1)
 207             stack.append(tmp)
 208
 209     if len(stack) != 1:
 210         raise Exception("Spurious empty rpn code for " + set.get('name') + " :: " +
 211                 counter.get('name') + ".\nThis is probably due to some unhandled RPN function, in the equation \"" +
 212                 equation + "\"")
 213
 214     value = stack[-1]
 215
 216     if value in hw_vars:
 217         value = hw_vars[value];
 218
 219     c("\nreturn " + value + ";")
 220
 221 def splice_rpn_expression(set, counter, expression):
 222     tokens = expression.split()
 223     stack = []
 224
 225     for token in tokens:
 226         stack.append(token)
 227         while stack and stack[-1] in exp_ops:
 228             op = stack.pop()
 229             argc, callback = exp_ops[op]
 230             args = []
 231             for i in range(0, argc):
 232                 operand = stack.pop()
 233                 if operand[0] == "$":
 234                     if operand in hw_vars:
 235                         operand = hw_vars[operand]
 236                     else:
 237                         raise Exception("Failed to resolve variable " + operand + " in expression " + expression + " for " + set.get('name') + " :: " + counter.get('name'));
 238                 args.append(operand)
 239
 240             subexp = callback(args)
 241
 242             stack.append(subexp)
 243
 244     if len(stack) != 1:
 245         raise Exception("Spurious empty rpn expression for " + set.get('name') + " :: " +
 246                 counter.get('name') + ".\nThis is probably due to some unhandled RPN operation, in the expression \"" +
 247                 expression + "\"")
 248
 249     return stack[-1]
 250
 251 def output_counter_read(set, counter, counter_vars):
 252     c("\n")
 253     c("/* {0} :: {1} */".format(set.get('name'), counter.get('name')))
 254     ret_type = counter.get('data_type')
 255     if ret_type == "uint64":
 256         ret_type = "uint64_t"
 257
 258     c("static " + ret_type)
 259     read_sym = "{0}__{1}__{2}__read".format(set.get('chipset').lower(), set.get('underscore_name'), counter.get('underscore_name'))
 260     c(read_sym + "(struct brw_context *brw,\n")
 261     c_indent(len(read_sym) + 1)
 262     c("const struct brw_perf_query_info *query,\n")
 263     c("uint64_t *accumulator)\n")
 264     c_outdent(len(read_sym) + 1)
 265
 266     c("{")
 267     c_indent(3)
 268
 269     output_rpn_equation_code(set, counter, counter.get('equation'), counter_vars)
 270
 271     c_outdent(3)
 272     c("}")
 273
 274     return read_sym
 275
 276 def output_counter_max(set, counter, counter_vars):
 277     max_eq = counter.get('max_equation')
 278
 279     if not max_eq:
 280         return "0; /* undefined */"
 281
 282     try:
 283         val = float(max_eq)
 284         return max_eq + ";"
 285     except ValueError:
 286         pass
 287
 288     # We can only report constant maximum values via INTEL_performance_query
 289     for token in max_eq.split():
 290         if token[0] == '$' and token not in hw_vars:
 291             return "0; /* unsupported (varies over time) */"
 292
 293     c("\n")
 294     c("/* {0} :: {1} */".format(set.get('name'), counter.get('name')))
 295     ret_type = counter.get('data_type')
 296     if ret_type == "uint64":
 297         ret_type = "uint64_t"
 298
 299     c("static " + ret_type)
 300     max_sym = "{0}__{1}__{2}__max".format(set.get('chipset').lower(), set.get('underscore_name'), counter.get('underscore_name'))
 301     c(max_sym + "(struct brw_context *brw)\n")
 302
 303     c("{")
 304     c_indent(3)
 305
 306     output_rpn_equation_code(set, counter, max_eq, counter_vars)
 307
 308     c_outdent(3)
 309     c("}")
 310
 311     return max_sym + "(brw);"
 312
 313 c_type_sizes = { "uint32_t": 4, "uint64_t": 8, "float": 4, "double": 8, "bool": 4 }
 314 def sizeof(c_type):
 315     return c_type_sizes[c_type]
 316
 317 def pot_align(base, pot_alignment):
 318     return (base + pot_alignment - 1) & ~(pot_alignment - 1);
 319
 320 semantic_type_map = {
 321     "duration": "raw",
 322     "ratio": "event"
 323     }
 324
 325 def output_counter_report(set, counter, current_offset):
 326     data_type = counter.get('data_type')
 327     data_type_uc = data_type.upper()
 328     c_type = data_type
 329
 330     if "uint" in c_type:
 331         c_type = c_type + "_t"
 332
 333     semantic_type = counter.get('semantic_type')
 334     if semantic_type in semantic_type_map:
 335         semantic_type = semantic_type_map[semantic_type]
 336
 337     semantic_type_uc = semantic_type.upper()
 338
 339     c("\n")
 340
 341     availability = counter.get('availability')
 342     if availability:
 343         expression = splice_rpn_expression(set, counter, availability)
 344         lines = expression.split(' && ')
 345         n_lines = len(lines)
 346         if n_lines == 1:
 347             c("if (" + lines[0] + ") {")
 348         else:
 349             c("if (" + lines[0] + " &&")
 350             c_indent(4)
 351             for i in range(1, (n_lines - 1)):
 352                 c(lines[i] + " &&")
 353             c(lines[(n_lines - 1)] + ") {")
 354             c_outdent(4)
 355         c_indent(3)
 356
 357     c("counter = &query->counters[query->n_counters++];\n")
 358     c("counter->oa_counter_read_" + data_type + " = " + read_funcs[counter.get('symbol_name')] + ";\n")
 359     c("counter->name = \"" + counter.get('name') + "\";\n")
 360     c("counter->desc = \"" + counter.get('description') + "\";\n")
 361     c("counter->type = GL_PERFQUERY_COUNTER_" + semantic_type_uc + "_INTEL;\n")
 362     c("counter->data_type = GL_PERFQUERY_COUNTER_DATA_" + data_type_uc + "_INTEL;\n")
 363     c("counter->raw_max = " + max_values[counter.get('symbol_name')] + "\n")
 364
 365     current_offset = pot_align(current_offset, sizeof(c_type))
 366     c("counter->offset = " + str(current_offset) + ";\n")
 367     c("counter->size = sizeof(" + c_type + ");\n")
 368
 369     if availability:
 370         c_outdent(3);
 371         c("}")
 372
 373     return current_offset + sizeof(c_type)
 374
 375
 376 def main():
 377     global c_file
 378     global header_file
 379     global max_values
 380     global read_funcs
 381
 382     parser = argparse.ArgumentParser()
 383     parser.add_argument("xml", help="XML description of metrics")
 384     parser.add_argument("--header", help="Header file to write")
 385     parser.add_argument("--code", help="C file to write")
 386     parser.add_argument("--chipset", help="Chipset to generate code for", required=True)
 387
 388     args = parser.parse_args()
 389
 390     chipset = args.chipset.lower()
 391
 392     if args.header:
 393         header_file = open(args.header, 'w')
 394
 395     if args.code:
 396         c_file = open(args.code, 'w')
 397
 398     tree = et.parse(args.xml)
 399
 400
 401     copyright = textwrap.dedent("""\
 402         /* Autogenerated file, DO NOT EDIT manually! generated by {}
 403          *
 404          * Copyright (c) 2015 Intel Corporation
 405          *
 406          * Permission is hereby granted, free of charge, to any person obtaining a
 407          * copy of this software and associated documentation files (the "Software"),
 408          * to deal in the Software without restriction, including without limitation
 409          * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 410          * and/or sell copies of the Software, and to permit persons to whom the
 411          * Software is furnished to do so, subject to the following conditions:
 412          *
 413          * The above copyright notice and this permission notice (including the next
 414          * paragraph) shall be included in all copies or substantial portions of the
 415          * Software.
 416          *
 417          * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 418          * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 419          * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 420          * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 421          * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 422          * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 423          * DEALINGS IN THE SOFTWARE.
 424          */
 425
 426         """).format(os.path.basename(__file__))
 427
 428     h(copyright)
 429     h(textwrap.dedent("""\
 430         #pragma once
 431
 432         struct brw_context;
 433
 434         """))
 435
 436     c(copyright)
 437     c(textwrap.dedent("""\
 438         #include <stdint.h>
 439         #include <stdbool.h>
 440
 441         #include "util/hash_table.h"
 442
 443         """))
 444
 445     c("#include \"brw_oa_" + chipset + ".h\"")
 446
 447     c(textwrap.dedent("""\
 448         #include "brw_context.h"
 449         #include "brw_performance_query.h"
 450
 451
 452         #define MIN(a, b) ((a < b) ? (a) : (b))
 453         #define MAX(a, b) ((a > b) ? (a) : (b))
 454
 455         """))
 456
 457     for set in tree.findall(".//set"):
 458         max_values = {}
 459         read_funcs = {}
 460         counter_vars = {}
 461         counters = set.findall("counter")
 462
 463         assert set.get('chipset').lower() == chipset
 464
 465         for counter in counters:
 466             empty_vars = {}
 467             read_funcs[counter.get('symbol_name')] = output_counter_read(set, counter, counter_vars)
 468             max_values[counter.get('symbol_name')] = output_counter_max(set, counter, empty_vars)
 469             counter_vars["$" + counter.get('symbol_name')] = counter
 470
 471
 472         c("\nstatic struct brw_perf_query_counter {0}_{1}_query_counters[{2}];\n".format(chipset, set.get('underscore_name'), len(counters)))
 473         c("static struct brw_perf_query_info " + chipset + "_" + set.get('underscore_name') + "_query = {\n")
 474         c_indent(3)
 475
 476         c(".kind = OA_COUNTERS,\n")
 477         c(".name = \"" + set.get('name') + "\",\n")
 478         c(".guid = \"" + set.get('hw_config_guid') + "\",\n")
 479
 480         c(".counters = {0}_{1}_query_counters,".format(chipset, set.get('underscore_name')))
 481         c(".n_counters = 0,")
 482         c(".oa_metrics_set_id = 0, /* determined at runtime, via sysfs */")
 483
 484         if chipset == "hsw":
 485             c(textwrap.dedent("""\
 486                 .oa_format = I915_OA_FORMAT_A45_B8_C8,
 487
 488                 /* Accumulation buffer offsets... */
 489                 .gpu_time_offset = 0,
 490                 .a_offset = 1,
 491                 .b_offset = 46,
 492                 .c_offset = 54,
 493                 """))
 494         else:
 495             c(textwrap.dedent("""\
 496                 .oa_format = I915_OA_FORMAT_A32u40_A4u32_B8_C8,
 497
 498                 /* Accumulation buffer offsets... */
 499                 .gpu_time_offset = 0,
 500                 .gpu_clock_offset = 1,
 501                 .a_offset = 2,
 502                 .b_offset = 38,
 503                 .c_offset = 46,
 504                 """))
 505
 506         c_outdent(3)
 507         c("};\n")
 508
 509         c("\nstatic void\n")
 510         c("register_" + set.get('underscore_name') + "_counter_query(struct brw_context *brw)\n")
 511         c("{\n")
 512         c_indent(3)
 513
 514         c("static struct brw_perf_query_info *query = &" + chipset + "_" + set.get('underscore_name') + "_query;\n")
 515         c("struct brw_perf_query_counter *counter;\n")
 516
 517         c("\n")
 518         c("/* Note: we're assuming there can't be any variation in the definition ")
 519         c(" * of a query between contexts so it's ok to describe a query within a ")
 520         c(" * global variable which only needs to be initialized once... */")
 521         c("\nif (!query->data_size) {")
 522         c_indent(3)
 523
 524         offset = 0
 525         for counter in counters:
 526             offset = output_counter_report(set, counter, offset)
 527
 528
 529         c("\nquery->data_size = counter->offset + counter->size;\n")
 530
 531         c_outdent(3)
 532         c("}");
 533
 534         c("\n_mesa_hash_table_insert(brw->perfquery.oa_metrics_table, query->guid, query);")
 535
 536         c_outdent(3)
 537         c("}\n")
 538
 539     h("void brw_oa_register_queries_" + chipset + "(struct brw_context *brw);\n")
 540
 541     c("\nvoid")
 542     c("brw_oa_register_queries_" + chipset + "(struct brw_context *brw)")
 543     c("{")
 544     c_indent(3)
 545
 546     for set in tree.findall(".//set"):
 547         c("register_" + set.get('underscore_name') + "_counter_query(brw);")
 548
 549     c_outdent(3)
 550     c("}")
 551
 552
 553 if __name__ == '__main__':
 554     main()