From 3490d5bf189eb2600a35572341a8073fdbd0d333 Mon Sep 17 00:00:00 2001
From: Matthew Poremba <matthew.poremba@amd.com>
Date: Wed, 24 Apr 2024 10:42:58 -0700
Subject: [PATCH] gpu-compute: Add DebugFlag for LDS

This prints what values are read/written to LDS and the previous value
on write. This is useful for debugging problems with LDS instructions.

Change-Id: I30063327bec1a1a808914a018467d5d78d5d58b4
---
 src/gpu-compute/SConscript   |  3 ++-
 src/gpu-compute/lds_state.hh | 50 ++++++++++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+), 1 deletion(-)

diff --git a/src/gpu-compute/SConscript b/src/gpu-compute/SConscript
index e4536ba2a5..23e3377f50 100644
--- a/src/gpu-compute/SConscript
+++ b/src/gpu-compute/SConscript
@@ -84,6 +84,7 @@ DebugFlag('GPUExec')
 DebugFlag('GPUFetch')
 DebugFlag('GPUInst')
 DebugFlag('GPUKernelInfo')
+DebugFlag('GPULDS')
 DebugFlag('GPUMem')
 DebugFlag('GPUPort')
 DebugFlag('GPUPrefetch')
@@ -106,4 +107,4 @@ DebugFlag('WavefrontStack')
 CompoundFlag('GPUALL', ['GPUCoalescer', 'GPUDisp', 'GPUExec', 'GPUFetch',
                         'GPUMem', 'GPUPort', 'GPUSched', 'GPUSRF', 'GPUSync',
                         'GPUTLB', 'GPUVRF', 'GPURFC', 'GPUWgLatency',
-                        'GPUKernelInfo', 'GPUInitAbi'])
+                        'GPUKernelInfo', 'GPUInitAbi', 'GPULDS'])
diff --git a/src/gpu-compute/lds_state.hh b/src/gpu-compute/lds_state.hh
index 3228b7822c..d336d35079 100644
--- a/src/gpu-compute/lds_state.hh
+++ b/src/gpu-compute/lds_state.hh
@@ -39,6 +39,7 @@
 #include <utility>
 #include <vector>
 
+#include "debug/GPULDS.hh"
 #include "gpu-compute/misc.hh"
 #include "mem/port.hh"
 #include "params/LdsState.hh"
@@ -75,10 +76,30 @@ class LdsChunk
          * chunk allocated to this WG we return 0.
          */
         if (index >= chunk.size()) {
+            DPRINTF(GPULDS, "LDS[%d][%d]: Read 0 beyond size (%ld)\n",
+                    dispatchId, wgId, chunk.size());
             return (T)0;
         }
 
         T *p0 = (T *) (&(chunk.at(index)));
+
+        if (sizeof(T) <= 4) {
+            [[maybe_unused]] uint32_t int_val =
+                *reinterpret_cast<uint32_t*>(p0);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Read %08x from index %d\n",
+                    dispatchId, wgId, int_val, index);
+        } else if (sizeof(T) <= 8) {
+            [[maybe_unused]] uint64_t int_val =
+                *reinterpret_cast<uint64_t*>(p0);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Read %016lx from index %d\n",
+                    dispatchId, wgId, int_val, index);
+        } else if (sizeof(T) <= 16) {
+            [[maybe_unused]] uint64_t *int_vals =
+                reinterpret_cast<uint64_t*>(p0);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Read %016lx%016lx from index %d\n",
+                    dispatchId, wgId, int_vals[1], int_vals[0], index);
+        }
+
         return *p0;
     }
 
@@ -94,10 +115,33 @@ class LdsChunk
          * chunk allocated to this WG are dropped.
          */
         if (index >= chunk.size()) {
+            DPRINTF(GPULDS, "LDS[%d][%d]: Ignoring write beyond size (%ld)\n",
+                    dispatchId, wgId, chunk.size());
             return;
         }
 
         T *p0 = (T *) (&(chunk.at(index)));
+
+        if (sizeof(T) <= 4) {
+            [[maybe_unused]] uint32_t prev_val =
+                *reinterpret_cast<uint32_t*>(p0);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Write %08lx to index %d (was "
+                    "%08lx)\n", dispatchId, wgId, value, index, prev_val);
+        } else if (sizeof(T) <= 8) {
+            [[maybe_unused]] uint64_t prev_val =
+                *reinterpret_cast<uint64_t*>(p0);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Write %016lx to index %d (was "
+                    "%016lx)\n", dispatchId, wgId, value, index, prev_val);
+        } else if (sizeof(T) <= 16) {
+            [[maybe_unused]] uint64_t *prev_vals =
+                reinterpret_cast<uint64_t*>(p0);
+            [[maybe_unused]] const uint64_t *next_vals =
+                reinterpret_cast<const uint64_t*>(&value);
+            DPRINTF(GPULDS, "LDS[%d][%d]: Write %016lx%016lx to index %d "
+                    "(was %016lx%016lx)\n", dispatchId, wgId, next_vals[1],
+                    next_vals[0], index, prev_vals[1], prev_vals[0]);
+        }
+
         *p0 = value;
     }
 
@@ -131,6 +175,9 @@ class LdsChunk
         return chunk.size();
     }
 
+    uint32_t dispatchId;
+    uint32_t wgId;
+
   protected:
     // the actual data store for this slice of the LDS
     std::vector<uint8_t> chunk;
@@ -402,6 +449,9 @@ class LdsState: public ClockedObject
             // make an entry for this workgroup
             refCounter[dispatchId][wgId] = 0;
 
+            chunkMap[dispatchId][wgId].dispatchId = dispatchId;
+            chunkMap[dispatchId][wgId].wgId = wgId;
+
             return &chunkMap[dispatchId][wgId];
         }
     }