i965: Add script to gen code for OA counter queries
authorRobert Bragg <robert@sixbynine.org>
Thu, 4 Jun 2015 12:55:37 +0000 (13:55 +0100)
committerRobert Bragg <robert@sixbynine.org>
Thu, 9 Mar 2017 13:45:44 +0000 (13:45 +0000)
Avoiding lots of error prone boilerplate and easing our ability to add +
maintain support for multiple OA performance counter queries for each
generation:

This adds a python script to generate code for building up
performance_queries from the metric sets and counters described in
brw_oa_hsw.xml as well as functions to normalize each counter based on
the RPN expressions given.

Although the XML file currently only includes a single metric set, the
code generated assumes there could be many sets.

The metrics as described in XML get translated into C structures
which are registered in a brw->perfquery.oa_metrics_table hash table
keyed by the GUID of the metric set in XML.

v2: numerous python style improvements (Dylan)
v3: Makefile.am fixups (Emil)
v4: Pattern rule for codegen + orthogonal .c and .h rules (Robert)

Signed-off-by: Robert Bragg <robert@sixbynine.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
src/mesa/drivers/dri/i965/Makefile.am
src/mesa/drivers/dri/i965/Makefile.sources
src/mesa/drivers/dri/i965/brw_oa.py [new file with mode: 0644]

index 4ff7e70b51f649dd23c8cba67bc13bdfdd124357..9e922ac41fdf2783bf108abe0deebc8622a7334b 100644 (file)
@@ -76,7 +76,9 @@ noinst_LTLIBRARIES = \
        libi965_compiler.la \
        $(I965_PERGEN_LIBS)
 
-libi965_dri_la_SOURCES = $(i965_FILES)
+libi965_dri_la_SOURCES = \
+       $(i965_FILES) \
+       $(i965_oa_GENERATED_FILES)
 libi965_dri_la_LIBADD = \
        $(top_builddir)/src/intel/common/libintel_common.la \
        $(top_builddir)/src/intel/isl/libisl.la \
@@ -93,7 +95,9 @@ BUILT_SOURCES = $(i965_compiler_GENERATED_FILES)
 CLEANFILES = $(BUILT_SOURCES)
 
 EXTRA_DIST = \
-       brw_nir_trig_workarounds.py
+       brw_nir_trig_workarounds.py \
+       brw_oa_hsw.xml \
+       brw_oa.py
 
 TEST_LIBS = \
        libi965_compiler.la \
@@ -170,3 +174,14 @@ test_eu_validate_SOURCES = \
 test_eu_validate_LDADD = \
        $(top_builddir)/src/gtest/libgtest.la \
        $(TEST_LIBS)
+
+BUILT_SOURCES += $(i965_oa_GENERATED_FILES)
+
+# Note: we avoid using a multi target rule here and outputting both the
+# .c and .h files in one go so we don't hit problems with parallel
+# make and multiple invocations of the same script trying to write
+# to the same files.
+brw_oa_%.h: brw_oa_%.xml brw_oa.py Makefile.am
+       $(PYTHON2) $(PYTHON_FLAGS) $(srcdir)/brw_oa.py --header=$(builddir)/brw_oa_$(*).h --chipset="$(*)" $(srcdir)/brw_oa_$(*).xml
+brw_oa_%.c: brw_oa_%.xml brw_oa.py Makefile.am
+       $(PYTHON2) $(PYTHON_FLAGS) $(srcdir)/brw_oa.py --code=$(builddir)/brw_oa_$(*).c --chipset="$(*)" $(srcdir)/brw_oa_$(*).xml
index b1776a8513ae0f5b5c76fea9a80d33e974dae741..f9e5f6e594fa992ca365fb106ea7ec10b7156ba4 100644 (file)
@@ -262,3 +262,7 @@ i965_gen8_FILES = \
 
 i965_gen9_FILES = \
        genX_blorp_exec.c
+
+i965_oa_GENERATED_FILES = \
+       brw_oa_hsw.h \
+       brw_oa_hsw.c
diff --git a/src/mesa/drivers/dri/i965/brw_oa.py b/src/mesa/drivers/dri/i965/brw_oa.py
new file mode 100644 (file)
index 0000000..bf950b1
--- /dev/null
@@ -0,0 +1,554 @@
+# Copyright (c) 2015-2017 Intel Corporation
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice (including the next
+# paragraph) shall be included in all copies or substantial portions of the
+# Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+# IN THE SOFTWARE.
+
+import argparse
+import os
+import sys
+import textwrap
+
+import xml.etree.cElementTree as et
+
+max_values = {}
+read_funcs = {}
+
+c_file = None
+_c_indent = 0
+
+def c(*args):
+    if c_file:
+        code = ' '.join(map(str,args))
+        for line in code.splitlines():
+            text = ''.rjust(_c_indent) + line
+            c_file.write(text.rstrip() + "\n")
+
+# indented, but no trailing newline...
+def c_line_start(code):
+    if c_file:
+        c_file.write(''.rjust(_c_indent) + code)
+def c_raw(code):
+    if c_file:
+        c_file.write(code)
+
+def c_indent(n):
+    global _c_indent
+    _c_indent = _c_indent + n
+def c_outdent(n):
+    global _c_indent
+    _c_indent = _c_indent - n
+
+header_file = None
+_h_indent = 0
+
+def h(*args):
+    if header_file:
+        code = ' '.join(map(str,args))
+        for line in code.splitlines():
+            text = ''.rjust(_h_indent) + line
+            header_file.write(text.rstrip() + "\n")
+
+def h_indent(n):
+    global _c_indent
+    _h_indent = _h_indent + n
+def h_outdent(n):
+    global _c_indent
+    _h_indent = _h_indent - n
+
+
+def emit_fadd(tmp_id, args):
+    c("double tmp{0} = {1} + {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+# Be careful to check for divide by zero...
+def emit_fdiv(tmp_id, args):
+    c("double tmp{0} = {1};".format(tmp_id, args[1]))
+    c("double tmp{0} = {1};".format(tmp_id + 1, args[0]))
+    c("double tmp{0} = tmp{1} ? tmp{2} / tmp{1} : 0;".format(tmp_id + 2, tmp_id + 1, tmp_id))
+    return tmp_id + 3
+
+def emit_fmax(tmp_id, args):
+    c("double tmp{0} = {1};".format(tmp_id, args[1]))
+    c("double tmp{0} = {1};".format(tmp_id + 1, args[0]))
+    c("double tmp{0} = MAX(tmp{1}, tmp{2});".format(tmp_id + 2, tmp_id, tmp_id + 1))
+    return tmp_id + 3
+
+def emit_fmul(tmp_id, args):
+    c("double tmp{0} = {1} * {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+def emit_fsub(tmp_id, args):
+    c("double tmp{0} = {1} - {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+def emit_read(tmp_id, args):
+    type = args[1].lower()
+    c("uint64_t tmp{0} = accumulator[query->{1}_offset + {2}];".format(tmp_id, type, args[0]))
+    return tmp_id + 1
+
+def emit_uadd(tmp_id, args):
+    c("uint64_t tmp{0} = {1} + {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+# Be careful to check for divide by zero...
+def emit_udiv(tmp_id, args):
+    c("uint64_t tmp{0} = {1};".format(tmp_id, args[1]))
+    c("uint64_t tmp{0} = {1};".format(tmp_id + 1, args[0]))
+    c("uint64_t tmp{0} = tmp{1} ? tmp{2} / tmp{1} : 0;".format(tmp_id + 2, tmp_id + 1, tmp_id))
+    return tmp_id + 3
+
+def emit_umul(tmp_id, args):
+    c("uint64_t tmp{0} = {1} * {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+def emit_usub(tmp_id, args):
+    c("uint64_t tmp{0} = {1} - {2};".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+def emit_umin(tmp_id, args):
+    c("uint64_t tmp{0} = MIN({1}, {2});".format(tmp_id, args[1], args[0]))
+    return tmp_id + 1
+
+ops = {}
+#             (n operands, emitter)
+ops["FADD"] = (2, emit_fadd)
+ops["FDIV"] = (2, emit_fdiv)
+ops["FMAX"] = (2, emit_fmax)
+ops["FMUL"] = (2, emit_fmul)
+ops["FSUB"] = (2, emit_fsub)
+ops["READ"] = (2, emit_read)
+ops["UADD"] = (2, emit_uadd)
+ops["UDIV"] = (2, emit_udiv)
+ops["UMUL"] = (2, emit_umul)
+ops["USUB"] = (2, emit_usub)
+ops["UMIN"] = (2, emit_umin)
+
+def brkt(subexp):
+    if " " in subexp:
+        return "(" + subexp + ")"
+    else:
+        return subexp
+
+def splice_bitwise_and(args):
+    return brkt(args[1]) + " & " + brkt(args[0])
+
+def splice_logical_and(args):
+    return brkt(args[1]) + " && " + brkt(args[0])
+
+def splice_ult(args):
+    return brkt(args[1]) + " < " + brkt(args[0])
+
+def splice_ugte(args):
+    return brkt(args[1]) + " >= " + brkt(args[0])
+
+exp_ops = {}
+#                 (n operands, splicer)
+exp_ops["AND"]  = (2, splice_bitwise_and)
+exp_ops["UGTE"] = (2, splice_ugte)
+exp_ops["ULT"]  = (2, splice_ult)
+exp_ops["&&"]   = (2, splice_logical_and)
+
+
+hw_vars = {}
+hw_vars["$EuCoresTotalCount"] = "brw->perfquery.sys_vars.n_eus"
+hw_vars["$EuSlicesTotalCount"] = "brw->perfquery.sys_vars.n_eu_slices"
+hw_vars["$EuSubslicesTotalCount"] = "brw->perfquery.sys_vars.n_eu_sub_slices"
+hw_vars["$EuThreadsCount"] = "brw->perfquery.sys_vars.eu_threads_count"
+hw_vars["$SliceMask"] = "brw->perfquery.sys_vars.slice_mask"
+hw_vars["$SubsliceMask"] = "brw->perfquery.sys_vars.subslice_mask"
+hw_vars["$GpuTimestampFrequency"] = "brw->perfquery.sys_vars.timestamp_frequency"
+hw_vars["$GpuMinFrequency"] = "brw->perfquery.sys_vars.gt_min_freq"
+hw_vars["$GpuMaxFrequency"] = "brw->perfquery.sys_vars.gt_max_freq"
+
+def output_rpn_equation_code(set, counter, equation, counter_vars):
+    c("/* RPN equation: " + equation + " */")
+    tokens = equation.split()
+    stack = []
+    tmp_id = 0
+    tmp = None
+
+    for token in tokens:
+        stack.append(token)
+        while stack and stack[-1] in ops:
+            op = stack.pop()
+            argc, callback = ops[op]
+            args = []
+            for i in range(0, argc):
+                operand = stack.pop()
+                if operand[0] == "$":
+                    if operand in hw_vars:
+                        operand = hw_vars[operand]
+                    elif operand in counter_vars:
+                        reference = counter_vars[operand]
+                        operand = read_funcs[operand[1:]] + "(brw, query, accumulator)"
+                    else:
+                        raise Exception("Failed to resolve variable " + operand + " in equation " + equation + " for " + set.get('name') + " :: " + counter.get('name'));
+                args.append(operand)
+
+            tmp_id = callback(tmp_id, args)
+
+            tmp = "tmp{0}".format(tmp_id - 1)
+            stack.append(tmp)
+
+    if len(stack) != 1:
+        raise Exception("Spurious empty rpn code for " + set.get('name') + " :: " +
+                counter.get('name') + ".\nThis is probably due to some unhandled RPN function, in the equation \"" +
+                equation + "\"")
+
+    value = stack[-1]
+
+    if value in hw_vars:
+        value = hw_vars[value];
+
+    c("\nreturn " + value + ";")
+
+def splice_rpn_expression(set, counter, expression):
+    tokens = expression.split()
+    stack = []
+
+    for token in tokens:
+        stack.append(token)
+        while stack and stack[-1] in exp_ops:
+            op = stack.pop()
+            argc, callback = exp_ops[op]
+            args = []
+            for i in range(0, argc):
+                operand = stack.pop()
+                if operand[0] == "$":
+                    if operand in hw_vars:
+                        operand = hw_vars[operand]
+                    else:
+                        raise Exception("Failed to resolve variable " + operand + " in expression " + expression + " for " + set.get('name') + " :: " + counter.get('name'));
+                args.append(operand)
+
+            subexp = callback(args)
+
+            stack.append(subexp)
+
+    if len(stack) != 1:
+        raise Exception("Spurious empty rpn expression for " + set.get('name') + " :: " +
+                counter.get('name') + ".\nThis is probably due to some unhandled RPN operation, in the expression \"" +
+                expression + "\"")
+
+    return stack[-1]
+
+def output_counter_read(set, counter, counter_vars):
+    c("\n")
+    c("/* {0} :: {1} */".format(set.get('name'), counter.get('name')))
+    ret_type = counter.get('data_type')
+    if ret_type == "uint64":
+        ret_type = "uint64_t"
+
+    c("static " + ret_type)
+    read_sym = "{0}__{1}__{2}__read".format(set.get('chipset').lower(), set.get('underscore_name'), counter.get('underscore_name'))
+    c(read_sym + "(struct brw_context *brw,\n")
+    c_indent(len(read_sym) + 1)
+    c("const struct brw_perf_query_info *query,\n")
+    c("uint64_t *accumulator)\n")
+    c_outdent(len(read_sym) + 1)
+
+    c("{")
+    c_indent(3)
+
+    output_rpn_equation_code(set, counter, counter.get('equation'), counter_vars)
+
+    c_outdent(3)
+    c("}")
+
+    return read_sym
+
+def output_counter_max(set, counter, counter_vars):
+    max_eq = counter.get('max_equation')
+
+    if not max_eq:
+        return "0; /* undefined */"
+
+    try:
+        val = float(max_eq)
+        return max_eq + ";"
+    except ValueError:
+        pass
+
+    # We can only report constant maximum values via INTEL_performance_query
+    for token in max_eq.split():
+        if token[0] == '$' and token not in hw_vars:
+            return "0; /* unsupported (varies over time) */"
+
+    c("\n")
+    c("/* {0} :: {1} */".format(set.get('name'), counter.get('name')))
+    ret_type = counter.get('data_type')
+    if ret_type == "uint64":
+        ret_type = "uint64_t"
+
+    c("static " + ret_type)
+    max_sym = "{0}__{1}__{2}__max".format(set.get('chipset').lower(), set.get('underscore_name'), counter.get('underscore_name'))
+    c(max_sym + "(struct brw_context *brw)\n")
+
+    c("{")
+    c_indent(3)
+
+    output_rpn_equation_code(set, counter, max_eq, counter_vars)
+
+    c_outdent(3)
+    c("}")
+
+    return max_sym + "(brw);"
+
+c_type_sizes = { "uint32_t": 4, "uint64_t": 8, "float": 4, "double": 8, "bool": 4 }
+def sizeof(c_type):
+    return c_type_sizes[c_type]
+
+def pot_align(base, pot_alignment):
+    return (base + pot_alignment - 1) & ~(pot_alignment - 1);
+
+semantic_type_map = {
+    "duration": "raw",
+    "ratio": "event"
+    }
+
+def output_counter_report(set, counter, current_offset):
+    data_type = counter.get('data_type')
+    data_type_uc = data_type.upper()
+    c_type = data_type
+
+    if "uint" in c_type:
+        c_type = c_type + "_t"
+
+    semantic_type = counter.get('semantic_type')
+    if semantic_type in semantic_type_map:
+        semantic_type = semantic_type_map[semantic_type]
+
+    semantic_type_uc = semantic_type.upper()
+
+    c("\n")
+
+    availability = counter.get('availability')
+    if availability:
+        expression = splice_rpn_expression(set, counter, availability)
+        lines = expression.split(' && ')
+        n_lines = len(lines)
+        if n_lines == 1:
+            c("if (" + lines[0] + ") {")
+        else:
+            c("if (" + lines[0] + " &&")
+            c_indent(4)
+            for i in range(1, (n_lines - 1)):
+                c(lines[i] + " &&")
+            c(lines[(n_lines - 1)] + ") {")
+            c_outdent(4)
+        c_indent(3)
+
+    c("counter = &query->counters[query->n_counters++];\n")
+    c("counter->oa_counter_read_" + data_type + " = " + read_funcs[counter.get('symbol_name')] + ";\n")
+    c("counter->name = \"" + counter.get('name') + "\";\n")
+    c("counter->desc = \"" + counter.get('description') + "\";\n")
+    c("counter->type = GL_PERFQUERY_COUNTER_" + semantic_type_uc + "_INTEL;\n")
+    c("counter->data_type = GL_PERFQUERY_COUNTER_DATA_" + data_type_uc + "_INTEL;\n")
+    c("counter->raw_max = " + max_values[counter.get('symbol_name')] + "\n")
+
+    current_offset = pot_align(current_offset, sizeof(c_type))
+    c("counter->offset = " + str(current_offset) + ";\n")
+    c("counter->size = sizeof(" + c_type + ");\n")
+
+    if availability:
+        c_outdent(3);
+        c("}")
+
+    return current_offset + sizeof(c_type)
+
+
+def main():
+    global c_file
+    global header_file
+    global max_values
+    global read_funcs
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("xml", help="XML description of metrics")
+    parser.add_argument("--header", help="Header file to write")
+    parser.add_argument("--code", help="C file to write")
+    parser.add_argument("--chipset", help="Chipset to generate code for", required=True)
+
+    args = parser.parse_args()
+
+    chipset = args.chipset.lower()
+
+    if args.header:
+        header_file = open(args.header, 'w')
+
+    if args.code:
+        c_file = open(args.code, 'w')
+
+    tree = et.parse(args.xml)
+
+
+    copyright = textwrap.dedent("""\
+        /* Autogenerated file, DO NOT EDIT manually! generated by {}
+         *
+         * Copyright (c) 2015 Intel Corporation
+         *
+         * Permission is hereby granted, free of charge, to any person obtaining a
+         * copy of this software and associated documentation files (the "Software"),
+         * to deal in the Software without restriction, including without limitation
+         * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+         * and/or sell copies of the Software, and to permit persons to whom the
+         * Software is furnished to do so, subject to the following conditions:
+         *
+         * The above copyright notice and this permission notice (including the next
+         * paragraph) shall be included in all copies or substantial portions of the
+         * Software.
+         *
+         * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+         * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+         * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+         * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+         * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+         * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+         * DEALINGS IN THE SOFTWARE.
+         */
+
+        """).format(os.path.basename(__file__))
+
+    h(copyright)
+    h(textwrap.dedent("""\
+        #pragma once
+
+        struct brw_context;
+
+        """))
+
+    c(copyright)
+    c(textwrap.dedent("""\
+        #include <stdint.h>
+        #include <stdbool.h>
+
+        #include "util/hash_table.h"
+
+        """))
+
+    c("#include \"brw_oa_" + chipset + ".h\"")
+
+    c(textwrap.dedent("""\
+        #include "brw_context.h"
+        #include "brw_performance_query.h"
+
+
+        #define MIN(a, b) ((a < b) ? (a) : (b))
+        #define MAX(a, b) ((a > b) ? (a) : (b))
+
+        """))
+
+    for set in tree.findall(".//set"):
+        max_values = {}
+        read_funcs = {}
+        counter_vars = {}
+        counters = set.findall("counter")
+
+        assert set.get('chipset').lower() == chipset
+
+        for counter in counters:
+            empty_vars = {}
+            read_funcs[counter.get('symbol_name')] = output_counter_read(set, counter, counter_vars)
+            max_values[counter.get('symbol_name')] = output_counter_max(set, counter, empty_vars)
+            counter_vars["$" + counter.get('symbol_name')] = counter
+
+
+        c("\nstatic struct brw_perf_query_counter {0}_{1}_query_counters[{2}];\n".format(chipset, set.get('underscore_name'), len(counters)))
+        c("static struct brw_perf_query_info " + chipset + "_" + set.get('underscore_name') + "_query = {\n")
+        c_indent(3)
+
+        c(".kind = OA_COUNTERS,\n")
+        c(".name = \"" + set.get('name') + "\",\n")
+        c(".guid = \"" + set.get('hw_config_guid') + "\",\n")
+
+        c(".counters = {0}_{1}_query_counters,".format(chipset, set.get('underscore_name')))
+        c(".n_counters = 0,")
+        c(".oa_metrics_set_id = 0, /* determined at runtime, via sysfs */")
+
+        if chipset == "hsw":
+            c(textwrap.dedent("""\
+                .oa_format = I915_OA_FORMAT_A45_B8_C8,
+
+                /* Accumulation buffer offsets... */
+                .gpu_time_offset = 0,
+                .a_offset = 1,
+                .b_offset = 46,
+                .c_offset = 54,
+                """))
+        else:
+            c(textwrap.dedent("""\
+                .oa_format = I915_OA_FORMAT_A32u40_A4u32_B8_C8,
+
+                /* Accumulation buffer offsets... */
+                .gpu_time_offset = 0,
+                .gpu_clock_offset = 1,
+                .a_offset = 2,
+                .b_offset = 38,
+                .c_offset = 46,
+                """))
+
+        c_outdent(3)
+        c("};\n")
+
+        c("\nstatic void\n")
+        c("register_" + set.get('underscore_name') + "_counter_query(struct brw_context *brw)\n")
+        c("{\n")
+        c_indent(3)
+
+        c("static struct brw_perf_query_info *query = &" + chipset + "_" + set.get('underscore_name') + "_query;\n")
+        c("struct brw_perf_query_counter *counter;\n")
+
+        c("\n")
+        c("/* Note: we're assuming there can't be any variation in the definition ")
+        c(" * of a query between contexts so it's ok to describe a query within a ")
+        c(" * global variable which only needs to be initialized once... */")
+        c("\nif (!query->data_size) {")
+        c_indent(3)
+
+        offset = 0
+        for counter in counters:
+            offset = output_counter_report(set, counter, offset)
+
+
+        c("\nquery->data_size = counter->offset + counter->size;\n")
+
+        c_outdent(3)
+        c("}");
+
+        c("\n_mesa_hash_table_insert(brw->perfquery.oa_metrics_table, query->guid, query);")
+
+        c_outdent(3)
+        c("}\n")
+
+    h("void brw_oa_register_queries_" + chipset + "(struct brw_context *brw);\n")
+
+    c("\nvoid")
+    c("brw_oa_register_queries_" + chipset + "(struct brw_context *brw)")
+    c("{")
+    c_indent(3)
+
+    for set in tree.findall(".//set"):
+        c("register_" + set.get('underscore_name') + "_counter_query(brw);")
+
+    c_outdent(3)
+    c("}")
+
+
+if __name__ == '__main__':
+    main()