swr/rast: Add general SWTag statistics

Update Archrast parser to use stats, used with an internal tool

Reviewed-by: Bruce Cherniak <[email protected]>

3 files changed, 191 insertions, 161 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/archrast/events.proto b/src/gallium/drivers/swr/rasterizer/archrast/events.proto
index aedc37b7440..1057a9437be 100644
--- a/src/gallium/drivers/swr/rasterizer/archrast/events.proto
+++ b/src/gallium/drivers/swr/rasterizer/archrast/events.proto
@@ -96,355 +96,371 @@ event SwrStoreTilesEvent
 event FrontendStatsEvent
 {
     uint32_t drawId;
-    uint64_t IaVertices;
-    uint64_t IaPrimitives;
-    uint64_t VsInvocations;
-    uint64_t HsInvocations;
-    uint64_t DsInvocations;
-    uint64_t GsInvocations;
-    uint64_t GsPrimitives;
-    uint64_t CInvocations;
-    uint64_t CPrimitives;
-    uint64_t SoPrimStorageNeeded0;
-    uint64_t SoPrimStorageNeeded1;
-    uint64_t SoPrimStorageNeeded2;
-    uint64_t SoPrimStorageNeeded3;
-    uint64_t SoNumPrimsWritten0;
-    uint64_t SoNumPrimsWritten1;
-    uint64_t SoNumPrimsWritten2;
-    uint64_t SoNumPrimsWritten3;
+    uint64_t counter IaVertices;
+    uint64_t counter IaPrimitives;
+    uint64_t counter VsInvocations;
+    uint64_t counter HsInvocations;
+    uint64_t counter DsInvocations;
+    uint64_t counter GsInvocations;
+    uint64_t counter GsPrimitives;
+    uint64_t counter CInvocations;
+    uint64_t counter CPrimitives;
+    uint64_t counter SoPrimStorageNeeded0;
+    uint64_t counter SoPrimStorageNeeded1;
+    uint64_t counter SoPrimStorageNeeded2;
+    uint64_t counter SoPrimStorageNeeded3;
+    uint64_t counter SoNumPrimsWritten0;
+    uint64_t counter SoNumPrimsWritten1;
+    uint64_t counter SoNumPrimsWritten2;
+    uint64_t counter SoNumPrimsWritten3;
 };
 
 event BackendStatsEvent
 {
     uint32_t drawId;
-    uint64_t DepthPassCount;
-    uint64_t PsInvocations;
-    uint64_t CsInvocations;
+    uint64_t counter DepthPassCount;
+    uint64_t counter PsInvocations;
+    uint64_t counter CsInvocations;
 
 };
 
 event EarlyZSingleSample
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateZSingleSample
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyStencilSingleSample
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateStencilSingleSample
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyZSampleRate
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateZSampleRate
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyStencilSampleRate
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateStencilSampleRate
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 // Total Early-Z counts, SingleSample and SampleRate
 event EarlyZ
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 // Total LateZ counts, SingleSample and SampleRate
 event LateZ
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 // Total EarlyStencil counts, SingleSample and SampleRate
 event EarlyStencil
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 // Total LateStencil counts, SingleSample and SampleRate
 event LateStencil
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyZNullPS
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyStencilNullPS
 {
     uint32_t drawId; 
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyZPixelRate
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateZPixelRate
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 
 event EarlyOmZ
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event EarlyOmStencil
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateOmZ
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event LateOmStencil
 {
     uint32_t drawId;
-    uint64_t passCount;
-    uint64_t failCount;
+    uint64_t counter passCount;
+    uint64_t counter failCount;
 };
 
 event GSInputPrims
 {
     uint32_t drawId;
-    uint64_t inputPrimCount;
+    uint64_t counter inputPrimCount;
 };
 
 event GSPrimsGen
 {
     uint32_t drawId;
-    uint64_t primGeneratedCount;
+    uint64_t counter primGeneratedCount;
 };
 
 event GSVertsInput
 {
     uint32_t drawId;
-    uint64_t vertsInput;
+    uint64_t counter vertsInput;
 };
 
 event TessPrims
 {
     uint32_t drawId;
-    uint64_t primCount;
+    uint64_t counter primCount;
 };
 
 event RasterTiles
 {
     uint32_t drawId;
-    uint32_t rastTileCount;
+    uint32_t counter rastTileCount;
 };
 
 event ClipperEvent
 {
     uint32_t drawId;
-    uint32_t trivialRejectCount;
-    uint32_t trivialAcceptCount;
-    uint32_t mustClipCount;
+    uint32_t counter trivialRejectCount;
+    uint32_t counter trivialAcceptCount;
+    uint32_t counter mustClipCount;
 };
 
 event CullEvent
 {
     uint32_t drawId;
-    uint64_t backfacePrimCount;
-    uint64_t degeneratePrimCount;
+    uint64_t counter backfacePrimCount;
+    uint64_t counter degeneratePrimCount;
 };
 
 event AlphaEvent
 {
     uint32_t drawId;
-    uint32_t alphaTestCount;
-    uint32_t alphaBlendCount;
+    uint32_t counter alphaTestCount;
+    uint32_t counter alphaBlendCount;
 };
 
 event VSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 };
 
 event HSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 };
 
 event DSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 };
 
 event GSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 
 };
 
 event PSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 };
 
 event CSInfo
 {
     uint32_t drawId;
-    uint32_t numInstExecuted;
-    uint32_t numSampleExecuted;
-    uint32_t numSampleLExecuted;
-    uint32_t numSampleBExecuted;
-    uint32_t numSampleCExecuted;
-    uint32_t numSampleCLZExecuted;
-    uint32_t numSampleCDExecuted;
-    uint32_t numGather4Executed;
-    uint32_t numGather4CExecuted;
-    uint32_t numGather4CPOExecuted;
-    uint32_t numGather4CPOCExecuted;
-    uint32_t numLodExecuted;
+    uint32_t counter numInstExecuted;
+    uint32_t counter numSampleExecuted;
+    uint32_t counter numSampleLExecuted;
+    uint32_t counter numSampleBExecuted;
+    uint32_t counter numSampleCExecuted;
+    uint32_t counter numSampleCLZExecuted;
+    uint32_t counter numSampleCDExecuted;
+    uint32_t counter numGather4Executed;
+    uint32_t counter numGather4CExecuted;
+    uint32_t counter numGather4CPOExecuted;
+    uint32_t counter numGather4CPOCExecuted;
+    uint32_t counter numLodExecuted;
 };
 
 event SWTagFrameEvent
 {
+	uint64_t swTagFrame;
 };
 
 event SWTagRenderpassEvent
 {
+	uint64_t swTagFrame;
+	uint32_t swTagDrawOrDispatch;
+	uint32_t swTagDraw;
+	uint32_t swTagDispatch;
+	uint32_t swTagRenderpassCount;
 };
 
 event SWTagDrawEvent
 {
-    uint32_t drawId;
+	uint64_t swTagFrame;
+	uint32_t swTagDrawOrDispatch;
+	uint32_t swTagDraw;
+	uint32_t swTagDispatch;
 };
 
 event SWTagDispatchEvent
 {
-    uint32_t drawId;
+	uint64_t swTagFrame;
+	uint32_t swTagDrawOrDispatch;
+	uint32_t swTagDraw;
+	uint32_t swTagDispatch;
 };
 
 event SWTagFlushEvent
 {
-    uint32_t drawId;
-    char flushReason[256];
-    uint32_t flushType;
+	uint64_t swTagFrame;
+	uint32_t swTagDrawOrDispatch;
+	uint32_t swTagDraw;
+	uint32_t swTagDispatch;
+	uint32_t swTagFlushCounter;
+    char swTagFlushReason[256];
+    uint32_t swTagFlushType;
 };
\ No newline at end of file
diff --git a/src/gallium/drivers/swr/rasterizer/codegen/gen_archrast.py b/src/gallium/drivers/swr/rasterizer/codegen/gen_archrast.py
index 6182504567a..b4b6127c8cc 100644
--- a/src/gallium/drivers/swr/rasterizer/codegen/gen_archrast.py
+++ b/src/gallium/drivers/swr/rasterizer/codegen/gen_archrast.py
@@ -27,6 +27,9 @@ import re
 from gen_common import *
 
 def parse_event_fields(lines, idx, event_dict):
+    """
+        Parses lines from a proto file that contain an event definition and stores it in event_dict
+    """
     fields = []
     end_of_event = False
 
@@ -36,13 +39,14 @@ def parse_event_fields(lines, idx, event_dict):
         line = lines[idx].rstrip()
         idx += 1
 
-        match = re.match(r'(\s*)([\w\*]+)(\s*)([\w]+)(\[\d+\])*', line)
+        match = re.match(r'(\s*)([\w\*]+)(\s+)(counter\s+)*([\w]+)(\[\d+\])*', line)
 
         if match:
             field = {
                 "type": match.group(2),
-                "name": match.group(4),
-                "size": int(match.group(5)[1:-1]) if match.group(5) else 1
+                "name": match.group(5),
+                "size": int(match.group(6)[1:-1]) if match.group(6) else 1,
+                "counter": True if match.group(4) else False
             }
             fields.append(field)
 
@@ -54,6 +58,9 @@ def parse_event_fields(lines, idx, event_dict):
     return idx
 
 def parse_enums(lines, idx, event_dict):
+    """
+        Parses lines from a proto file that contain an enum definition and stores it in event_dict
+    """
     enum_names = []
     end_of_enum = False
 
@@ -77,15 +84,21 @@ def parse_enums(lines, idx, event_dict):
     return idx
 
 def parse_protos(files, verbose=False):
-
+    """
+        Parses a proto file and returns a dictionary of event definitions
+    """
     protos = {}
     protos['events'] = {}       # event dictionary containing events with their fields
     protos['event_names'] = []  # needed to keep events in order parsed. dict is not ordered.
+    protos['event_map'] = {}    # dictionary to map event ids to event names
     protos['enums'] = {}
     protos['enum_names'] = []
 
     eventId = 0
 
+    if type(files) is not list:
+        files = [files]
+
     for filename in files:
         if verbose:
             print("Parsing proto file: %s" % os.path.normpath(filename))
@@ -106,21 +119,22 @@ def parse_protos(files, verbose=False):
                 if match:
                     eventId += 1
                     event_name = match.group(3)
-                    protos['event_names'].append(event_name)
+                    protos["event_names"].append(event_name)
 
-                    protos['events'][event_name] = {}
-                    protos['events'][event_name]['event_id'] = eventId
-                    idx = parse_event_fields(lines, idx, protos['events'][event_name])
+                    protos["events"][event_name] = {}
+                    protos["events"][event_name]["event_id"] = eventId
+                    protos["event_map"][eventId] = event_name
+                    idx = parse_event_fields(lines, idx, protos["events"][event_name])
 
                 # search for enums.
                 match = re.match(r'(\s*)enum(\s*)(\w+)', line)
 
                 if match:
                     enum_name = match.group(3)
-                    protos['enum_names'].append(enum_name)
+                    protos["enum_names"].append(enum_name)
 
-                    protos['enums'][enum_name] = {}
-                    idx = parse_enums(lines, idx, protos['enums'][enum_name])
+                    protos["enums"][enum_name] = {}
+                    idx = parse_enums(lines, idx, protos["enums"][enum_name])
     return protos
 
 
diff --git a/src/gallium/drivers/swr/rasterizer/codegen/templates/gen_ar_event.hpp b/src/gallium/drivers/swr/rasterizer/codegen/templates/gen_ar_event.hpp
index 884a3210039..6c592f74461 100644
--- a/src/gallium/drivers/swr/rasterizer/codegen/templates/gen_ar_event.hpp
+++ b/src/gallium/drivers/swr/rasterizer/codegen/templates/gen_ar_event.hpp
@@ -113,7 +113,7 @@ namespace ArchRast
         % for i in range(len(fields)):
             % if fields[i]['size'] > 1:
                 % if fields[i]['type'] == 'char':
-            // Copy size of string (null-terminated) followed by string buffer info entire buffer
+            // Copy size of string (null-terminated) followed by string into entire buffer
             SWR_ASSERT(${fields[i]['name']}_size + 1 < ${fields[i]['size']} - sizeof(uint32_t), "String length must be less than size of char buffer - size(uint32_t)!");
             memcpy(data.${fields[i]['name']}, &${fields[i]['name']}_size, sizeof(uint32_t));
             strcpy_s(data.${fields[i]['name']} + sizeof(uint32_t), ${fields[i]['name']}_size + 1, ${fields[i]['name']});