Split matmul into matvec; add large matrix benchmark

Rename var names to row/col for more clarity.
Better estimate error tolerance via max abs col sum.

PiperOrigin-RevId: 657601791

Author: jan-wassenberg

BUILD.bazel

@@ -25,6 +25,7 @@ cc_library(
     textual_hdrs = [
         "ops/ops-inl.h",
         "ops/matmul-inl.h",
+        "ops/matvec-inl.h",
     ],
     deps = [
         "//compression:compress",
@@ -56,6 +57,25 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "matvec_test",
+    size = "small",
+    timeout = "long",
+    srcs = ["ops/matvec_test.cc"],
+    local_defines = ["HWY_IS_TEST"],
+    # for test_suite.
+    tags = ["hwy_ops_test"],
+    deps = [
+        ":ops",
+        "@googletest//:gtest_main",  # buildcleaner: keep
+        "//compression:compress",
+        "@hwy//:hwy",
+        "@hwy//:hwy_test_util",
+        "@hwy//:nanobenchmark",
+        "@hwy//:thread_pool",
+    ],
+)
+
 cc_test(
     name = "matmul_test",
     size = "small",

CMakeLists.txt

@@ -98,6 +98,7 @@ set(SOURCES
   gemma/weights.cc
   gemma/weights.h
   ops/matmul-inl.h
+  ops/matvec-inl.h
   ops/ops-inl.h
   util/app.h
   util/args.h
@@ -148,7 +149,9 @@ set(GEMMA_TEST_FILES
   backprop/backward_test.cc
   backprop/backward_scalar_test.cc
   backprop/optimize_test.cc
-  gemma/ops_test.cc
+  ops/ops_test.cc
+  ops/matmul_test.cc
+  ops/matvec_test.cc
   evals/gemma_test.cc
 )
 

backprop/forward-inl.h

@@ -39,7 +39,7 @@
 #define THIRD_PARTY_GEMMA_CPP_FORWARD_TOGGLE
 #endif
 
-#include "ops/matmul-inl.h"
+#include "ops/matvec-inl.h"
 #include "ops/ops-inl.h"
 #include "hwy/highway.h"
 

gemma/gemma-inl.h

@@ -40,11 +40,11 @@
 #include "gemma/weights.h"
 // Placeholder for internal test4, do not remove
 #include "ops/matmul-inl.h"
+#include "ops/matvec-inl.h"
 #include "ops/ops-inl.h"
 #include "hwy/aligned_allocator.h"
 #include "hwy/base.h"
 #include "hwy/bit_set.h"
-#include "hwy/contrib/matvec/matvec-inl.h"
 #include "hwy/contrib/thread_pool/thread_pool.h"
 #include "hwy/contrib/thread_pool/topology.h"
 #include "hwy/highway.h"