i965: XML description of Haswell OA metric set

In preparation for exposing Gen Observation Architecture performance
counters via INTEL_performance_query this adds an XML description for an
initial 'Render Metrics Basic Gen7.5' query and corresponding counters.

The intention is to auto generate code for building a query from these
counters as well as the code for normalizing the individual counters.

Note that the upstream for this XML data is currently GPU Top:

  https://github.com/rib/gputop

The files are maintained under gputop-data/ and they are themselves
derived from files in an internal 'MDAPI XML' schema. There are scripts
under gputop-scripts/ and make rules in gputop-data/Makefile.xml for
maintaining these files.

Signed-off-by: Robert Bragg <[email protected]>
Reviewed-by: Lionel Landwerlin <[email protected]>
Acked-by: Kenneth Graunke <[email protected]>

1 files changed, 998 insertions, 0 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_oa_hsw.xml b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
new file mode 100644
index 00000000000..49476712633
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
@@ -0,0 +1,998 @@
+<?xml version="1.0"?>
+<metrics version="1486657040" merge_md5="">
+  <set name="Render Metrics Basic Gen7.5"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="403d8832-1a27-4aa6-a64e-f5389ce7b212"
+       chipset="HSW"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES EU Stall"
+             description="The percentage of time in which evaluation shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_stall"
+             units="percent"
+             symbol_name="DsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Alpha Test Fails"
+             description="The total number of pixels dropped on post-FS alpha test."
+             data_type="uint64"
+             equation="A 37 READ"
+             underscore_name="alpha_test_fails"
+             units="pixels"
+             symbol_name="AlphaTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which sampler 1 was bottlenecks."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="TES Threads Dispatched"
+             description="The total number of evaluation shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 15 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="TES AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 12 READ $DsThreads UDIV"
+             underscore_name="ds_eu_active_per_thread"
+             units="cycles"
+             symbol_name="DsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 25 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GS EU Stall"
+             description="The percentage of time in which geometry shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 23 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_stall"
+             units="percent"
+             symbol_name="GsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Active"
+             description="The percentage of time in which compute shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_active"
+             units="percent"
+             symbol_name="CsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS EU Active"
+             description="The percentage of time in which vertex shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 2 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_active"
+             units="percent"
+             symbol_name="VsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS EU Active"
+             description="The percentage of time in which control shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_active"
+             units="percent"
+             symbol_name="HsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES EU Active"
+             description="The percentage of time in which evaluation shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_eu_active"
+             units="percent"
+             symbol_name="DsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Evaluation Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS EU Active"
+             description="The percentage of time in which geometry shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 22 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_eu_active"
+             units="percent"
+             symbol_name="GsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 27 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_active"
+             units="percent"
+             symbol_name="PsEuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS EU Stall"
+             description="The percentage of time in which compute shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cs_eu_stall"
+             units="percent"
+             symbol_name="CsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 1 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS EU Stall"
+             description="The percentage of time in which vertex shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 3 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_eu_stall"
+             units="percent"
+             symbol_name="VsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS EU Stall"
+             description="The percentage of time in which control shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_eu_stall"
+             units="percent"
+             symbol_name="HsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS EU Stall"
+             description="The percentage of time in which fragment shaders were stalled on the EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 28 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_stall"
+             units="percent"
+             symbol_name="PsEuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="CS Duration"
+             description="Total Compute Shader GPU duration."
+             data_type="uint64"
+             equation="A 17 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 18 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="cs_duration"
+             units="us"
+             symbol_name="CsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 30 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which sampler 0 was busy."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which sampler 1 was busy."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers were busy."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ  B 1 READ UADD $GpuCoreClocks FDIV 2 FDIV 100 FMUL"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TES Duration"
+             description="Total Evaluation Shader GPU duration."
+             data_type="uint64"
+             equation="A 12 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 13 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="ds_duration"
+             units="us"
+             symbol_name="DsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 1 READ 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 128 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 20 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="CS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 17 READ $CsThreads UDIV"
+             underscore_name="cs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="CsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="EU Array/Compute Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which sampler 0 was bottlenecks."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 23 READ $GsThreads UDIV"
+             underscore_name="gs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="GsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Sampler 0 Texels LOD0"
+             description="The total number of texels lookups in LOD0 in sampler 0 unit."
+             data_type="uint64"
+             equation="B 4 READ 4 UMUL"
+             underscore_name="sampler0_texels"
+             units="texels"
+             symbol_name="Sampler0Texels"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler"
+             />
+    <counter name="Sampler 1 Texels LOD0"
+             description="The total number of texels lookups in LOD0 in sampler 1 unit."
+             data_type="uint64"
+             equation="B 5 READ 4 UMUL"
+             underscore_name="sampler1_texels"
+             units="texels"
+             symbol_name="Sampler1Texels"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="Sampler"
+             />
+    <counter name="Sampler Texels LOD0"
+             description="The total number of texels lookups in LOD0 in all sampler units."
+             data_type="uint64"
+             equation="$Sampler0Texels $Sampler1Texels UADD $EuSlicesTotalCount UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="GS Duration"
+             description="Total Geometry Shader GPU duration."
+             data_type="uint64"
+             equation="A 22 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 23 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="gs_duration"
+             units="us"
+             symbol_name="GsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Idle"
+             description="The percentage of time in which the Execution Units were idle."
+             data_type="float"
+             max_equation="100"
+             equation="100 $EuActive $EuStall FADD FSUB"
+             underscore_name="eu_idle"
+             units="percent"
+             symbol_name="EuIdle"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 28 READ $PsThreads UDIV"
+             underscore_name="ps_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="PsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 3 READ $VsThreads UDIV"
+             underscore_name="vs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="VsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="C 5 READ $EuSlicesTotalCount 4 UMUL UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has being processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 41 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which fragment shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 27 READ $PsThreads UDIV"
+             underscore_name="ps_eu_active_per_thread"
+             units="cycles"
+             symbol_name="PsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="TCS Duration"
+             description="Total Control Shader GPU duration."
+             data_type="uint64"
+             equation="A 7 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 8 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="hs_duration"
+             units="us"
+             symbol_name="HsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 OCL"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TES AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which evaluation shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 13 READ $DsThreads UDIV"
+             underscore_name="ds_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="DsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Evaluation Shader"
+             />
+    <counter name="GS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which geometry shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 22 READ $GsThreads UDIV"
+             underscore_name="gs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="GsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS Threads Dispatched"
+             description="The total number of control shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 10 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="TCS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 8 READ $HsThreads UDIV"
+             underscore_name="hs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="HsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Control Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 36 READ"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Late Depth Test Fails"
+             description="The total number of pixels dropped on post-FS depth test."
+             data_type="uint64"
+             equation="A 39 READ $SamplesKilledInPs USUB"
+             underscore_name="post_ps_depth_test_fails"
+             units="pixels"
+             symbol_name="PostPsDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Sampler Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers were bottlenecks."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TCS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which control shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 7 READ $HsThreads UDIV"
+             underscore_name="hs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="HsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Control Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Duration"
+             description="Total Fragment Shader GPU duration."
+             data_type="uint64"
+             equation="A 27 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 28 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="ps_duration"
+             units="us"
+             symbol_name="PsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 33 READ"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="CS AVG Stall per Thread"
+             description="The average number of cycles per hardware thread run in which compute shaders were stalled on the EUs."
+             data_type="uint64"
+             equation="A 18 READ $CsThreads UDIV"
+             underscore_name="cs_eu_stall_per_thread"
+             units="cycles"
+             symbol_name="CsEuStallPerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Late Stencil Test Fails"
+             description="The total number of pixels dropped on post-FS stencil test."
+             data_type="uint64"
+             equation="A 38 READ"
+             underscore_name="post_ps_stencil_test_fails"
+             units="pixels"
+             symbol_name="PostPsStencilTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 3 READ 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ UADD 2 UMUL $EuSlicesTotalCount UMUL 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS AVG Active per Thread"
+             description="The average number of cycles per hardware thread run in which vertex shaders were processed actively on the EUs."
+             data_type="uint64"
+             max_equation="100"
+             equation="A 2 READ $VsThreads UDIV"
+             underscore_name="vs_eu_active_per_thread"
+             units="cycles"
+             symbol_name="VsEuActivePerThread"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Duration"
+             description="Total Vertex Shader GPU duration."
+             data_type="uint64"
+             equation="A 2 READ  A 0 READ UMUL  A 2 READ  A 7 READ UADD  A 12 READ UADD  A 17 READ UADD  A 22 READ UADD  A 27 READ UADD UDIV  A 3 READ  A 1 READ UMUL  A 3 READ  A 8 READ UADD  A 13 READ UADD  A 18 READ UADD  A 23 READ UADD  A 28 READ UADD UDIV UADD $GpuTime UMUL $GpuCoreClocks $EuCoresTotalCount UMUL 1000 UMUL UDIV"
+             underscore_name="vs_duration"
+             units="us"
+             symbol_name="VsDuration"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 40 READ"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <register_config id="0">
+        <register type="NOA" address="0x000253A4" value="0x01600000" />
+        <register type="NOA" address="0x00025440" value="0x00100000" />
+        <register type="NOA" address="0x00025128" value="0x00000000" />
+        <register type="NOA" address="0x0002691C" value="0x00000800" />
+        <register type="NOA" address="0x00026AA0" value="0x01500000" />
+        <register type="NOA" address="0x00026B9C" value="0x00006000" />
+        <register type="NOA" address="0x0002791C" value="0x00000800" />
+        <register type="NOA" address="0x00027AA0" value="0x01500000" />
+        <register type="NOA" address="0x00027B9C" value="0x00006000" />
+        <register type="NOA" address="0x0002641C" value="0x00000400" />
+        <register type="NOA" address="0x00025380" value="0x00000010" />
+        <register type="NOA" address="0x0002538C" value="0x00000000" />
+        <register type="NOA" address="0x00025384" value="0x0800AAAA" />
+        <register type="NOA" address="0x00025400" value="0x00000004" />
+        <register type="NOA" address="0x0002540C" value="0x06029000" />
+        <register type="NOA" address="0x00025410" value="0x00000002" />
+        <register type="NOA" address="0x00025404" value="0x5C30FFFF" />
+        <register type="NOA" address="0x00025100" value="0x00000016" />
+        <register type="NOA" address="0x00025110" value="0x00000400" />
+        <register type="NOA" address="0x00025104" value="0x00000000" />
+        <register type="NOA" address="0x00026804" value="0x00001211" />
+        <register type="NOA" address="0x00026884" value="0x00000100" />
+        <register type="NOA" address="0x00026900" value="0x00000002" />
+        <register type="NOA" address="0x00026908" value="0x00700000" />
+        <register type="NOA" address="0x00026904" value="0x00000000" />
+        <register type="NOA" address="0x00026984" value="0x00001022" />
+        <register type="NOA" address="0x00026A04" value="0x00000011" />
+        <register type="NOA" address="0x00026A80" value="0x00000006" />
+        <register type="NOA" address="0x00026A88" value="0x00000C02" />
+        <register type="NOA" address="0x00026A84" value="0x00000000" />
+        <register type="NOA" address="0x00026B04" value="0x00001000" />
+        <register type="NOA" address="0x00026B80" value="0x00000002" />
+        <register type="NOA" address="0x00026B8C" value="0x00000007" />
+        <register type="NOA" address="0x00026B84" value="0x00000000" />
+        <register type="NOA" address="0x00027804" value="0x00004844" />
+        <register type="NOA" address="0x00027884" value="0x00000400" />
+        <register type="NOA" address="0x00027900" value="0x00000002" />
+        <register type="NOA" address="0x00027908" value="0x0E000000" />
+        <register type="NOA" address="0x00027904" value="0x00000000" />
+        <register type="NOA" address="0x00027984" value="0x00004088" />
+        <register type="NOA" address="0x00027A04" value="0x00000044" />
+        <register type="NOA" address="0x00027A80" value="0x00000006" />
+        <register type="NOA" address="0x00027A88" value="0x00018040" />
+        <register type="NOA" address="0x00027A84" value="0x00000000" />
+        <register type="NOA" address="0x00027B04" value="0x00004000" />
+        <register type="NOA" address="0x00027B80" value="0x00000002" />
+        <register type="NOA" address="0x00027B8C" value="0x000000E0" />
+        <register type="NOA" address="0x00027B84" value="0x00000000" />
+        <register type="NOA" address="0x00026104" value="0x00002222" />
+        <register type="NOA" address="0x00026184" value="0x0C006666" />
+        <register type="NOA" address="0x00026284" value="0x04000000" />
+        <register type="NOA" address="0x00026304" value="0x04000000" />
+        <register type="NOA" address="0x00026400" value="0x00000002" />
+        <register type="NOA" address="0x00026410" value="0x000000A0" />
+        <register type="NOA" address="0x00026404" value="0x00000000" />
+        <register type="NOA" address="0x00025420" value="0x04108020" />
+        <register type="NOA" address="0x00025424" value="0x1284A420" />
+        <register type="NOA" address="0x0002541C" value="0x00000000" />
+        <register type="NOA" address="0x00025428" value="0x00042049" />
+    </register_config>
+    <register_config id="1">
+        <register type="OA" address="0x00002724" value="0x00800000" /> <!--OASTARTTRIG6 -->
+        <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002714" value="0x00800000" /> <!--OASTARTTRIG1 -->
+        <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
+    </register_config>
+  </set>
+
+</metrics>