perf(encoding): add row-based match finder backend

.github/scripts/run-benchmarks.sh

@@ -105,7 +105,6 @@ timing_point_count = 0
 DELTA_LOW = 0.99
 DELTA_HIGH = 1.05
 REGRESSION_STAGES = {"compress", "decompress"}
-REGRESSION_LEVELS = {"default", "better"}
 REGRESSION_SCENARIOS = {
     "small-4k-log-lines",
     "decodecorpus-z000033",
@@ -162,10 +161,10 @@ def normalize_impl(impl):
         return "ffi"
     return impl
 
-def include_in_regression_set(parsed_name):
+def include_in_regression_set(parsed_name, regression_levels):
     return (
         parsed_name["stage"] in REGRESSION_STAGES
-        and parsed_name["level"] in REGRESSION_LEVELS
+        and parsed_name["level"] in regression_levels
         and parsed_name["scenario"] in REGRESSION_SCENARIOS
     )
 
@@ -199,12 +198,6 @@ with open(raw_path) as f:
             timings.append((name, ms))
             parsed = parse_benchmark_name(name)
             timing_point_count += 1
-            if include_in_regression_set(parsed):
-                benchmark_results.append({
-                    "name": name,
-                    "unit": "ms",
-                    "value": round(ms, 3),
-                })
             timing_rows.append({
                 "name": name,
                 "stage": parsed["stage"],
@@ -320,6 +313,17 @@ if timing_point_count == 0:
     print("ERROR: No benchmark timings parsed from compare_ffi output.", file=sys.stderr)
     sys.exit(1)
 
+regression_levels = {row["level"] for row in ratios}
+benchmark_results = [
+    {
+        "name": row["name"],
+        "unit": "ms",
+        "value": round(row["ms_per_iter"], 3),
+    }
+    for row in timing_rows
+    if include_in_regression_set(row, regression_levels)
+]
+
 if not benchmark_results:
     print(
         "WARN: No regression-set benchmark rows matched smoke filter; "

.github/workflows/ci.yml

@@ -114,19 +114,22 @@ jobs:
   benchmark:
     name: Performance regression check (${{ matrix.bench.id }})
     needs: lint
-    timeout-minutes: 40
+    timeout-minutes: ${{ matrix.bench.timeout_minutes }}
     strategy:
       fail-fast: false
       matrix:
         bench:
           - id: x86_64-gnu
             target_triple: x86_64-unknown-linux-gnu
+            timeout_minutes: 40
             setup_command: ""
           - id: i686-gnu
             target_triple: i686-unknown-linux-gnu
+            timeout_minutes: 55
             setup_command: "sudo apt-get update && sudo apt-get install -y gcc-multilib libc6-dev-i386"
           - id: x86_64-musl
             target_triple: x86_64-unknown-linux-musl
+            timeout_minutes: 40
             setup_command: "sudo apt-get update && sudo apt-get install -y musl-tools"
     runs-on: ubuntu-latest
     steps:

zstd/benches/support/mod.rs

@@ -72,7 +72,8 @@ pub(crate) fn benchmark_scenarios() -> Vec<Scenario> {
     scenarios
 }
 
-pub(crate) fn supported_levels() -> [LevelConfig; 4] {
+/// Benchmark levels mapped to comparable Rust and FFI compression settings.
+pub(crate) fn supported_levels() -> [LevelConfig; 5] {
     [
         LevelConfig {
             name: "fastest",
@@ -89,6 +90,11 @@ pub(crate) fn supported_levels() -> [LevelConfig; 4] {
             rust_level: CompressionLevel::Better,
             ffi_level: 7,
         },
+        LevelConfig {
+            name: "level4-row",
+            rust_level: CompressionLevel::Level(4),
+            ffi_level: 4,
+        },
         LevelConfig {
             name: "best",
             rust_level: CompressionLevel::Best,