Harden GPU pipeline selection and recall validation

.github/workflows/ci.yml

@@ -52,6 +52,16 @@ jobs:
             -o tests/tari_miner_reliability_test
           tests/tari_miner_reliability_test
 
+      - name: Build and run pipeline selection regressions
+        run: |
+          g++ -std=c++17 -O2 -I. \
+            tests/tari_miner_pipeline_test.cpp \
+            -o tests/tari_miner_pipeline_test
+          tests/tari_miner_pipeline_test
+
+      - name: Run GPU recall verifier fixtures
+        run: python3 tests/tari_c29_gpu_recall.py --self-test
+
       - name: Build and run Tari wrapper self-test
         run: |
           g++ -std=c++17 -O2 \
@@ -92,6 +102,8 @@ jobs:
             'tests\tari_pool_protocol_test.exe',
             'cl /nologo /std:c++17 /EHsc /O2 /I. tests\tari_miner_reliability_test.cpp /Fe:tests\tari_miner_reliability_test.exe',
             'tests\tari_miner_reliability_test.exe',
+            'cl /nologo /std:c++17 /EHsc /O2 /I. tests\tari_miner_pipeline_test.cpp /Fe:tests\tari_miner_pipeline_test.exe',
+            'tests\tari_miner_pipeline_test.exe',
             'cl /nologo /std:c++17 /EHsc /O2 tari_c29.cpp tari_c29_selftest.cpp /Fe:tari_c29_selftest.exe',
             'tari_c29_selftest.exe'
           ) -join ' && '
@@ -120,3 +132,7 @@ jobs:
           if ($manifest -contains 13) {
             throw 'HiveOS manifest must use LF line endings.'
           }
+
+      - name: Run GPU recall verifier fixtures
+        shell: powershell
+        run: python tests\tari_c29_gpu_recall.py --self-test

README.md

@@ -22,7 +22,11 @@ NVIDIA GPU and selects the correct backend for each card:
 
 Mixed rigs are supported. Each GPU runs in its own process with a worker name
 such as `RIG01-gpu0` or `RIG01-gpu1`. Pipeline depth is also selected inside
-each process from that card's available VRAM.
+each process from the free VRAM remaining after its first solver context starts.
+Automatic selection can use up to five contexts on Linux and Windows TCC, and
+up to four on Windows WDDM. An explicit `--pipeline` value from 1 through 5
+overrides the automatic platform cap; allocation still stops safely if VRAM
+runs out.
 
 The starter does not change GPU clocks, voltage, fans, or power limits.
 
@@ -198,6 +202,35 @@ bin/tari_c29_solver_sm_89 --device 0 --count 320 --pipeline 2
 
 A correct run ends with `verify failures: 0`.
 
+### Exact GPU recall regression
+
+The GPU recall test compares exact `(header nonce, proof hash)` sets, not just
+cycle counts. It independently packs every 42-edge proof, recalculates its
+BLAKE2b-256 hash and difficulty, repeats both builds, and checks pipeline
+parity. Build the shipped and conservative reference profiles for the GPU
+architecture, then run:
+
+```bat
+build_solver.bat sm_120
+build_solver.bat sm_120 reference
+python tests\tari_c29_gpu_recall.py run --candidate bin\tari_c29_solver_sm_120.exe --reference bin\validation\tari_c29_solver_sm_120_reference.exe --arch sm_120 --output-dir validation --parity-pipeline 4
+```
+
+```bash
+./build_solver.sh sm_89
+./build_solver.sh sm_89 reference
+python3 tests/tari_c29_gpu_recall.py run --candidate bin/tari_c29_solver_sm_89 --reference bin/validation/tari_c29_solver_sm_89_reference --arch sm_89 --output-dir validation --parity-pipeline 2
+```
+
+The default sequence tests 4,200 fixed graphs and saves the raw logs, JSONL
+proof records, invoked commands, and comparison report. The reference binaries
+remain under `bin/validation` and are not included in release packages.
+The runner also rejects a binary whose embedded build target, runtime GPU
+architecture, or compiled trim default does not match `--arch`. Throughput
+measurements are a separate performance check and do not replace the exact
+recall gate. Hosted CI compiles both profiles and exercises the verifier
+fixtures; the full sequence runs only on a trusted host with the matching GPU.
+
 ## Build From Source
 
 The repository includes its required third-party source under
@@ -218,8 +251,10 @@ three target architectures:
 ./build_all.sh
 ```
 
-Individual backends can be built with `build_solver` or `build_pool_miner` and
-one of `sm_86`, `sm_89`, or `sm_120`.
+`build_all` also compiles the non-release reference solvers used by the recall
+test. Individual backends can be built with `build_solver` or
+`build_pool_miner` and one of `sm_86`, `sm_89`, or `sm_120`; pass `reference`
+as the second `build_solver` argument to build only that validation profile.
 
 ## License
 

build_all.bat

@@ -7,6 +7,8 @@ for %%A in (sm_86 sm_89 sm_120) do (
   if errorlevel 1 goto failed
   call build_pool_miner.bat %%A
   if errorlevel 1 goto failed
+  call build_solver.bat %%A reference
+  if errorlevel 1 goto failed
 )
 
 echo.

build_all.sh

@@ -6,6 +6,7 @@ ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 for arch in sm_86 sm_89 sm_120; do
     "$ROOT/build_solver.sh" "$arch"
     "$ROOT/build_pool_miner.sh" "$arch"
+    "$ROOT/build_solver.sh" "$arch" reference
 done
 
 echo "All Linux GPU backends built successfully."

build_pool_miner.bat

@@ -16,7 +16,7 @@ set "CUCKAROO=%ROOT%third_party\cuckoo\src\cuckaroo"
 set "CRYPTO=%ROOT%third_party\cuckoo\src\crypto"
 
 echo Building tari_c29_pool_miner for %ARCH% ...
-"%NVCC%" -O3 -std=c++17 -arch=%ARCH% --default-stream per-thread -DXBITS=7 -DIDXSHIFT=9 -DGRAPH_UNION_SKIP=1 -DSOLVER_PRELAUNCH_NEXT=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 %EXTRA_FLAGS% -maxrregcount=96 -Xptxas -flcm=cg ^
+"%NVCC%" -O3 -std=c++17 -arch=%ARCH% --default-stream per-thread -DXBITS=7 -DIDXSHIFT=9 -DGRAPH_UNION_SKIP=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 %EXTRA_FLAGS% -maxrregcount=96 -Xptxas -flcm=cg ^
     -I"%ROOT%compat" ^
     -I"%CUCKAROO%" ^
     -I"%CRYPTO%" ^

build_pool_miner.sh

@@ -30,7 +30,7 @@ if [[ "$ARCH" == "sm_120" ]]; then
 fi
 
 echo "Building tari_c29_pool_miner_linux for $ARCH ..."
-"$NVCC" "${HOST_FLAGS[@]}" -O3 -std="$STD" --default-stream per-thread -DXBITS="$XBITS" -DIDXSHIFT="$IDXSHIFT" -DGRAPH_UNION_SKIP=1 -DSOLVER_PRELAUNCH_NEXT=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 \
+"$NVCC" "${HOST_FLAGS[@]}" -O3 -std="$STD" --default-stream per-thread -DXBITS="$XBITS" -DIDXSHIFT="$IDXSHIFT" -DGRAPH_UNION_SKIP=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 \
     "${EXTRA_FLAGS[@]}" \
     -maxrregcount="$MAXRREGCOUNT" -Xptxas -flcm=cg \
     -gencode "arch=$COMPUTE,code=$ARCH" \

build_solver.bat

@@ -5,18 +5,38 @@ setlocal
 set "ROOT=%~dp0"
 set ARCH=%1
 if "%ARCH%"=="" set ARCH=sm_120
+set BUILD_ARCH=%ARCH:sm_=%
+set PROFILE=%2
+if "%PROFILE%"=="" set PROFILE=release
 if not exist "%ROOT%bin" mkdir "%ROOT%bin"
+if /I "%PROFILE%"=="release" goto release_profile
+if /I "%PROFILE%"=="reference" goto reference_profile
+echo Unknown solver profile "%PROFILE%". Use release or reference.
+endlocal
+exit /b 2
+
+:release_profile
 set "OUTPUT=%ROOT%bin\tari_c29_solver_%ARCH%.exe"
+set "PROFILE_FLAGS=-DGRAPH_UNION_SKIP=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1"
 set EXTRA_FLAGS=
 if /I "%ARCH%"=="sm_120" set EXTRA_FLAGS=-DWARP_DST_ATOMICS_LATE=1 -DTARI_C29_DEFAULT_NTRIMS=48 -DSEEDB_REVERSE_LOOP=1 -DROUND0_DST_HASH_DYNAMIC_BITS=12 -DROUND0_DST_HASH_DYNAMIC_PROBES=4 -DROUND0_DST_HASH_FALLBACK_PLAIN=1 -DROUND0_DST_HASH_REPLAY_NORMAL_LOAD=1 -DROUND0_DST_HASH_REVERSE_INSERT=1 -DFUSE_FINAL_TAIL_CURRENT=1 -DFUSE_FINAL_TAIL_COUNT_NORMAL_LOAD=1 -DROUND23_TPB=960 -DROUND1_COUNT_NORMAL_LOAD=1 -DROUND23_COUNT_NORMAL_LOAD=1
+goto profile_ready
+
+:reference_profile
+if not exist "%ROOT%bin\validation" mkdir "%ROOT%bin\validation"
+set "OUTPUT=%ROOT%bin\validation\tari_c29_solver_%ARCH%_reference.exe"
+set "PROFILE_FLAGS=-DTARI_C29_REFERENCE_BUILD=1 -DGRAPH_UNION_SKIP=0 -DRECOVERY_SMALL_OUTPUT=0 -DSEEDA_REHASH=0 -DSKIP_LATE_NULL_CHECKS=0 -DWARP_DST_ATOMICS_LATE=0 -DTARI_C29_DEFAULT_NTRIMS=50 -DSEEDB_REVERSE_LOOP=0 -DROUND0_DST_HASH_DYNAMIC_BITS=0 -DROUND0_DST_HASH_DYNAMIC_PROBES=8 -DROUND0_DST_HASH_FALLBACK_PLAIN=0 -DROUND0_DST_HASH_REPLAY_NORMAL_LOAD=0 -DROUND0_DST_HASH_REVERSE_INSERT=0 -DFUSE_FINAL_TAIL_CURRENT=0 -DFUSE_FINAL_TAIL_COUNT_NORMAL_LOAD=0 -DROUND1_COUNT_NORMAL_LOAD=0 -DROUND23_COUNT_NORMAL_LOAD=0 -DROUND23_TPB=1024"
+set EXTRA_FLAGS=
+
+:profile_ready
 
 if not defined NVCC set "NVCC=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2\bin\nvcc.exe"
 if not exist "%NVCC%" set "NVCC=nvcc"
 set "CUCKAROO=%ROOT%third_party\cuckoo\src\cuckaroo"
 set "CRYPTO=%ROOT%third_party\cuckoo\src\crypto"
 
-echo Building tari_c29_solver for %ARCH% ...
-"%NVCC%" -O3 -std=c++17 -arch=%ARCH% --default-stream per-thread -DXBITS=7 -DIDXSHIFT=9 -DGRAPH_UNION_SKIP=1 -DSOLVER_PRELAUNCH_NEXT=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 %EXTRA_FLAGS% -maxrregcount=96 -Xptxas -flcm=cg ^
+echo Building tari_c29_solver for %ARCH% profile=%PROFILE% ...
+"%NVCC%" -O3 -std=c++17 -arch=%ARCH% --default-stream per-thread -DXBITS=7 -DIDXSHIFT=9 -DTARI_C29_BUILD_ARCH=%BUILD_ARCH% %PROFILE_FLAGS% %EXTRA_FLAGS% -maxrregcount=96 -Xptxas -flcm=cg ^
     -I"%ROOT%compat" ^
     -I"%CUCKAROO%" ^
     -I"%CRYPTO%" ^

build_solver.sh

@@ -6,9 +6,10 @@ set -euo pipefail
 # forward JIT. With a newer toolkit, pass NVCC=/path/to/nvcc and sm_89/sm_120.
 
 ARCH="${1:-sm_86}"
+BUILD_ARCH="${ARCH#sm_}"
+PROFILE="${2:-release}"
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 mkdir -p "$ROOT/bin"
-OUTPUT="$ROOT/bin/tari_c29_solver_$ARCH"
 CUCKOO_ROOT="${CUCKOO_ROOT:-"$ROOT/third_party/cuckoo"}"
 CUCKAROO="$CUCKOO_ROOT/src/cuckaroo"
 CRYPTO="$CUCKOO_ROOT/src/crypto"
@@ -24,13 +25,30 @@ HOST_FLAGS=()
 if [[ -n "$NVCC_HOST" ]]; then
     HOST_FLAGS=(-ccbin "$NVCC_HOST" --allow-unsupported-compiler)
 fi
+PROFILE_FLAGS=()
 EXTRA_FLAGS=()
-if [[ "$ARCH" == "sm_120" ]]; then
-    EXTRA_FLAGS=(-DWARP_DST_ATOMICS_LATE=1 -DTARI_C29_DEFAULT_NTRIMS=48 -DSEEDB_REVERSE_LOOP=1 -DROUND0_DST_HASH_DYNAMIC_BITS=12 -DROUND0_DST_HASH_DYNAMIC_PROBES=4 -DROUND0_DST_HASH_FALLBACK_PLAIN=1 -DROUND0_DST_HASH_REPLAY_NORMAL_LOAD=1 -DROUND0_DST_HASH_REVERSE_INSERT=1 -DFUSE_FINAL_TAIL_CURRENT=1 -DFUSE_FINAL_TAIL_COUNT_NORMAL_LOAD=1 -DROUND23_TPB=960 -DROUND1_COUNT_NORMAL_LOAD=1 -DROUND23_COUNT_NORMAL_LOAD=1)
-fi
+case "$PROFILE" in
+    release)
+        OUTPUT="$ROOT/bin/tari_c29_solver_$ARCH"
+        PROFILE_FLAGS=(-DGRAPH_UNION_SKIP=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1)
+        if [[ "$ARCH" == "sm_120" ]]; then
+            EXTRA_FLAGS=(-DWARP_DST_ATOMICS_LATE=1 -DTARI_C29_DEFAULT_NTRIMS=48 -DSEEDB_REVERSE_LOOP=1 -DROUND0_DST_HASH_DYNAMIC_BITS=12 -DROUND0_DST_HASH_DYNAMIC_PROBES=4 -DROUND0_DST_HASH_FALLBACK_PLAIN=1 -DROUND0_DST_HASH_REPLAY_NORMAL_LOAD=1 -DROUND0_DST_HASH_REVERSE_INSERT=1 -DFUSE_FINAL_TAIL_CURRENT=1 -DFUSE_FINAL_TAIL_COUNT_NORMAL_LOAD=1 -DROUND23_TPB=960 -DROUND1_COUNT_NORMAL_LOAD=1 -DROUND23_COUNT_NORMAL_LOAD=1)
+        fi
+        ;;
+    reference)
+        mkdir -p "$ROOT/bin/validation"
+        OUTPUT="$ROOT/bin/validation/tari_c29_solver_${ARCH}_reference"
+        PROFILE_FLAGS=(-DTARI_C29_REFERENCE_BUILD=1 -DGRAPH_UNION_SKIP=0 -DRECOVERY_SMALL_OUTPUT=0 -DSEEDA_REHASH=0 -DSKIP_LATE_NULL_CHECKS=0 -DWARP_DST_ATOMICS_LATE=0 -DTARI_C29_DEFAULT_NTRIMS=50 -DSEEDB_REVERSE_LOOP=0 -DROUND0_DST_HASH_DYNAMIC_BITS=0 -DROUND0_DST_HASH_DYNAMIC_PROBES=8 -DROUND0_DST_HASH_FALLBACK_PLAIN=0 -DROUND0_DST_HASH_REPLAY_NORMAL_LOAD=0 -DROUND0_DST_HASH_REVERSE_INSERT=0 -DFUSE_FINAL_TAIL_CURRENT=0 -DFUSE_FINAL_TAIL_COUNT_NORMAL_LOAD=0 -DROUND1_COUNT_NORMAL_LOAD=0 -DROUND23_COUNT_NORMAL_LOAD=0 -DROUND23_TPB=1024)
+        ;;
+    *)
+        echo "Unknown solver profile '$PROFILE'. Use release or reference." >&2
+        exit 2
+        ;;
+esac
 
-echo "Building tari_c29_solver_linux for $ARCH ..."
-"$NVCC" "${HOST_FLAGS[@]}" -O3 -std="$STD" --default-stream per-thread -DXBITS="$XBITS" -DIDXSHIFT="$IDXSHIFT" -DGRAPH_UNION_SKIP=1 -DSOLVER_PRELAUNCH_NEXT=1 -DRECOVERY_SMALL_OUTPUT=1 -DSEEDA_REHASH=1 \
+echo "Building tari_c29_solver_linux for $ARCH profile=$PROFILE ..."
+"$NVCC" "${HOST_FLAGS[@]}" -O3 -std="$STD" --default-stream per-thread -DXBITS="$XBITS" -DIDXSHIFT="$IDXSHIFT" -DTARI_C29_BUILD_ARCH="$BUILD_ARCH" \
+    "${PROFILE_FLAGS[@]}" \
     "${EXTRA_FLAGS[@]}" \
     -maxrregcount="$MAXRREGCOUNT" -Xptxas -flcm=cg \
     -gencode "arch=$COMPUTE,code=$ARCH" \

compress.hpp

@@ -1,5 +1,6 @@
 // Derived from tromp/cuckoo. SPDX-License-Identifier: GPL-2.0-or-later
 
+#include <stdio.h>
 #include <new>
 
 // compressor for cuckaroo nodes where edgetrimming
@@ -63,7 +64,7 @@ class compressor {
       word_t cu = nodes[ui];
       if (cu == NIL) {
         if (nnodes >= SIZE) {
-          print_log("NODE OVERFLOW at %x\n", u);
+          fprintf(stderr, "NODE OVERFLOW at %x\n", u);
           return 0;
         }
         nodes[ui] = u << SIZEBITS | nnodes;

mean_c29.cu

@@ -1148,7 +1148,7 @@ struct solver_ctx {
       return 0;
     }
     if (nedges > MAXEDGES) {
-      print_log("OOPS; losing %d edges beyond MAXEDGES=%d\n", nedges-MAXEDGES, MAXEDGES);
+      fprintf(stderr, "OOPS; losing %d edges beyond MAXEDGES=%d\n", nedges-MAXEDGES, MAXEDGES);
       nedges = MAXEDGES;
     }
 #if TRIM_STAGE_TIMING