From 463a3682be768244acecc51e4f6f9987b32c5ba8 Mon Sep 17 00:00:00 2001
From: Jan Wassenberg <janwas@google.com>
Date: Mon, 23 Feb 2026 08:55:04 -0800
Subject: [PATCH] Internal change

PiperOrigin-RevId: 874097322
---
 compression/test_util-inl.h | 57 ++++++++++++++++++++++---------------
 1 file changed, 34 insertions(+), 23 deletions(-)

diff --git a/compression/test_util-inl.h b/compression/test_util-inl.h
index 99b34b5b..bb2fadb0 100644
--- a/compression/test_util-inl.h
+++ b/compression/test_util-inl.h
@@ -18,6 +18,7 @@
 #define THIRD_PARTY_GEMMA_CPP_COMPRESSION_TEST_UTIL_INL_H_
 
 #include <stddef.h>
+#include <stdio.h>
 
 #include <vector>
 
@@ -219,6 +220,8 @@ void AssertClose(const MatPtrT<TA>& A, const MatPtrT<TB>& B,
   // magnitude, but also to f32 accumulation of rows in A and B.
   const double norm = MaxRowAbsSum(a_batch) * MaxRowAbsSum(b_trans_batch);
   const float max_abs = MaxAbs(a_batch) * MaxAbs(b_trans_batch);
+  HWY_ASSERT(hn::AllTrue(df, hn::IsFinite(hn::Set(df, norm))));
+  HWY_ASSERT(hn::AllTrue(df, hn::IsFinite(hn::Set(df, max_abs))));
   const double eps_bf16 = hwy::ConvertScalarTo<double>(hwy::Epsilon<BF16>());
   const double eps_f32 = hwy::ConvertScalarTo<double>(hwy::Epsilon<float>());
   // Dot() uses double-precision summation.
@@ -232,10 +235,7 @@ void AssertClose(const MatPtrT<TA>& A, const MatPtrT<TB>& B,
   if (tolerance > 500.0) {
     HWY_WARN("high tolerance %f norm %f maxabs %f\n", tolerance, norm, max_abs);
   }
-  const double rel_tolerance =
-      1.0 + hwy::ConvertScalarTo<double>(hwy::Epsilon<TC>());
-
-  double max_rel = 0.0;
+  double worst_l1 = 0.0;
   size_t worst_r = 0;
   size_t worst_c = 0;
   double worst_actual = 0.0;
@@ -247,34 +247,45 @@ void AssertClose(const MatPtrT<TA>& A, const MatPtrT<TB>& B,
     for (size_t c = 0; c < B.Rows(); c++) {
       const double expected_value = static_cast<double>(expected_row[c]);
       const double actual_value = static_cast<double>(actual_row[c]);
-      const bool in_range = expected_value - tolerance <= actual_value &&
-                            actual_value <= expected_value + tolerance;
+      const double l1 = hwy::ScalarAbs(expected_value - actual_value);
+      if (l1 > HWY_MAX(tolerance, tolerance * hwy::ScalarAbs(expected_value))) {
+        fprintf(stderr, "%zu,%zu\n", r, c);
+        ++num_outside;
+      }
 
-      if (!in_range) {
-        const double max = HWY_MAX(expected_value, actual_value);
-        const double min = HWY_MIN(expected_value, actual_value);
-        const double rel = max / HWY_MAX(min, 1E-6);
-        if (rel > max_rel) {
-          worst_expected = expected_value;
-          worst_actual = actual_value;
-          worst_r = r;
-          worst_c = c;
-          max_rel = rel;
-          ++num_outside;
-        }
+      if (l1 > worst_l1) {
+        worst_l1 = l1;
+        worst_expected = expected_value;
+        worst_actual = actual_value;
+        worst_r = r;
+        worst_c = c;
       }
     }
   }
 
-  if (max_rel > rel_tolerance) {
+  if (num_outside > 0) {
+    const size_t r_begin = worst_r >= 1 ? worst_r - 1 : 0;
+    const size_t r_end = HWY_MIN(r_begin + 3, A.Rows());
+    const size_t c_begin = worst_c >= 3 ? worst_c - 3 : 0;
+    const size_t c_end = HWY_MIN(c_begin + 7, B.Rows());
+    fprintf(stderr,
+            "%zu outside. Printing rows [%zu, %zu) and columns [%zu, %zu)\n",
+            num_outside, r_begin, r_end, c_begin, c_end);
+    for (size_t r = r_begin; r < r_end; r++) {
+      const float* expected_row = c_slow_batch.Row(r);
+      const float* actual_row = c_batch.Row(r);
+      for (size_t c = c_begin; c < c_end; c++) {
+        fprintf(stderr, "%6.3f=%6.3f ", expected_row[c], actual_row[c]);
+      }
+      fprintf(stderr, "\n");
+    }
+
     hwy::Abort(__FILE__, line,
                "(%zu,%zu): expected %f, actual %f, norm %f maxabs %f "
-               "tolerance %f rel %E max_rel %E num_outside %zu\n",
+               "tolerance %f worst_l1 %E\n",
                worst_r, worst_c, worst_expected, worst_actual, norm, max_abs,
-               tolerance, max_rel, rel_tolerance, num_outside);
+               tolerance, worst_l1);
   }
-  HWY_ASSERT(hn::AllFalse(
-      df, hn::IsEitherNaN(hn::Set(df, norm), hn::Set(df, max_abs))));
 }
 
 // NOLINTNEXTLINE(google-readability-namespace-comments)