Add mmap support (not yet used)

Also: const-correct ArgsBase,
add assert to mat.h checking element_bytes_
BUILD deps update (:shared provides shared.h, not :sfp)
PiperOrigin-RevId: 746073312

BUILD.bazel

@@ -138,7 +138,7 @@ cc_library(
     deps = [
         ":basics",
         "//compression:fields",
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:hwy",  # base.h
     ],
 )
@@ -159,10 +159,11 @@ cc_test(
     deps = [
         ":basics",
         ":common",
+        ":mat",
         ":weights",
         "@googletest//:gtest_main",  # buildcleaner: keep
         "//compression:compress",
-        "@highway//:hwy",
+        "@highway//:hwy",  # aligned_allocator.h
     ],
 )
 
@@ -176,7 +177,7 @@ cc_library(
         ":common",
         ":threading_context",
         "//compression:fields",
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:hwy",
         "@highway//:profiler",
     ],
@@ -348,7 +349,7 @@ cc_library(
         ":mat",
         "//compression:blob_store",
         "//compression:compress",
-        "//compression:io",
+        "//compression:io",  # Path
         "@highway//:hwy",
         "@highway//:profiler",
         "@highway//:stats",
@@ -362,8 +363,8 @@ cc_library(
     hdrs = ["gemma/tokenizer.h"],
     deps = [
         ":common",
-        "//compression:io",
-        "//compression:sfp",
+        "//compression:io",  # Path
+        "//compression:shared",
         "@highway//:hwy",
         "@highway//:profiler",
         "@com_google_sentencepiece//:sentencepiece_processor",
@@ -405,16 +406,17 @@ cc_library(
         ":allocator",
         ":basics",
         ":common",
-        ":ops",
-        ":mat",
-        ":tokenizer",
         ":kv_cache",
-        ":weights",
+        ":mat",
+        ":ops",
+        ":tokenizer",
         ":threading",
         ":threading_context",
+        ":weights",
         # Placeholder for internal dep, do not remove.,
+        "//compression:blob_store",
         "//compression:io",
-        "//compression:sfp",
+        "//compression:shared",
         "//paligemma:image",
         "@highway//:hwy",
         "@highway//:nanobenchmark",  # timer
@@ -445,7 +447,7 @@ cc_library(
         ":gemma_lib",
         ":ops",
         "//compression:io",
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:hwy",
     ],
 )
@@ -517,7 +519,7 @@ cc_binary(
         ":gemma_lib",
         ":ops",
         ":threading_context",
-        "//compression:sfp",
+        "//compression:shared",
         "//paligemma:image",
         "@highway//:hwy",
         "@highway//:profiler",
@@ -706,6 +708,7 @@ cc_library(
         ":mat",
         ":weights",
         "//compression:compress",
+        "//compression:shared",
         "@highway//:hwy",
         "@highway//:thread_pool",
     ],
@@ -731,7 +734,7 @@ cc_test(
         ":threading",
         ":weights",
         "@googletest//:gtest_main",  # buildcleaner: keep
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:thread_pool",
     ],
 )

compression/BUILD.bazel

@@ -40,6 +40,7 @@ cc_library(
         "//conditions:default": [],
     }),
     deps = [
+        "//:allocator",
         "@highway//:hwy",
     ] + FILE_DEPS,
 )
@@ -69,6 +70,7 @@ cc_library(
     hdrs = ["blob_store.h"],
     deps = [
         ":io",
+        "//:threading_context",
         "@highway//:hwy",
         "@highway//:thread_pool",
     ],
@@ -81,7 +83,7 @@ cc_test(
         ":blob_store",
         ":io",
         "@googletest//:gtest_main",  # buildcleaner: keep
-        "@highway//:hwy",
+        "//:threading_context",
         "@highway//:hwy_test_util",
         "@highway//:thread_pool",
     ],
@@ -115,21 +117,30 @@ cc_test(
 )
 
 cc_library(
-    name = "sfp",
+    name = "shared",
     hdrs = ["shared.h"],
-    textual_hdrs = ["sfp-inl.h"],
     deps = [
         "//:basics",
         "@highway//:hwy",
     ],
 )
 
+cc_library(
+    name = "sfp",
+    textual_hdrs = ["sfp-inl.h"],
+    deps = [
+        ":shared",
+        "//:basics",
+        "@highway//:hwy",
+    ],
+)
+
 cc_library(
     name = "nuq",
-    hdrs = ["shared.h"],
     textual_hdrs = ["nuq-inl.h"],
     deps = [
         ":sfp",
+        ":shared",
         "//:basics",
         "@highway//:hwy",
         "@highway//hwy/contrib/sort:vqsort",
@@ -144,6 +155,7 @@ cc_library(
     deps = [
         ":compress",
         ":distortion",
+        "//:mat",
         "@highway//:hwy",
         "@highway//:hwy_test_util",
         "@highway//:thread_pool",
@@ -254,6 +266,16 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "io_win",
+    srcs = ["io_win.cc"],
+    deps = [
+        ":io",
+        "//:allocator",
+        "@highway//:hwy",
+    ],
+)
+
 cc_binary(
     name = "blob_compare",
     srcs = ["blob_compare.cc"],

compression/fields.cc

@@ -87,7 +87,7 @@ class PrintVisitor : public VisitorBase {
   }
 
   void operator()(uint64_t& value) override {
-    fprintf(stderr, "%sU64 %zu\n", indent_.c_str(), value);
+    fprintf(stderr, "%sU64 %zu\n", indent_.c_str(), static_cast<size_t>(value));
   }
 
   void operator()(float& value) override {

compression/io.cc

@@ -36,12 +36,16 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <stdio.h>     // SEEK_END - unistd isn't enough for IDE.
+#include <sys/types.h>
+// Old OSX may require sys/types.h before sys/mman.h.
+#include <sys/mman.h>  // mmap
 #include <sys/stat.h>  // O_RDONLY
 #include <unistd.h>    // read, write, close
 
 #include <memory>
 
 #include "compression/io.h"
+#include "util/allocator.h"
 #include "hwy/base.h"  // HWY_ASSERT
 
 namespace gcpp {
@@ -93,6 +97,28 @@ class FilePosix : public File {
     }
     return pos == size;  // success if managed to write desired size
   }
+
+  MapPtr Map() override {
+    const size_t mapping_size = FileSize();
+    // No `MAP_POPULATE` because we do not want to wait for I/O, and
+    // `MAP_NONBLOCK` is not guaranteed. `MAP_HUGETLB` fails. `MAP_SHARED` is
+    // more efficient than `MAP_PRIVATE`; the main difference is that the former
+    // will eventually see subsequent changes to the file.
+    const int flags = MAP_SHARED;
+    void* mapping =
+        mmap(nullptr, mapping_size, PROT_READ, flags, fd_, /*offset=*/0);
+    if (mapping == MAP_FAILED) return MapPtr();
+
+#ifdef MADV_WILLNEED  // Missing on some OSX.
+    // (Maybe) initiate readahead.
+    madvise(mapping, mapping_size, MADV_WILLNEED);
+#endif
+
+    return MapPtr(static_cast<const uint8_t*>(mapping),
+                  DeleterFunc2([mapping_size](void* ptr) {
+                    HWY_ASSERT(munmap(ptr, mapping_size) == 0);
+                  }));
+  }
 };  // FilePosix
 
 HWY_MAYBE_UNUSED extern std::unique_ptr<File> OpenFileGoogle(

compression/io.h

@@ -23,6 +23,7 @@
 #include <string>
 #include <utility>  // std::move
 
+#include "util/allocator.h"
 #include "hwy/base.h"
 
 namespace gcpp {
@@ -32,6 +33,8 @@ namespace gcpp {
 // prefer to define Exists inline because there are multiple io*.cc files.
 struct Path;
 
+using MapPtr = AlignedPtr2<const uint8_t[]>;
+
 // Abstract base class enables multiple I/O backends in the same binary.
 class File {
  public:
@@ -50,6 +53,12 @@ class File {
 
   // Returns true if all the requested bytes were written.
   virtual bool Write(const void* from, uint64_t size, uint64_t offset) = 0;
+
+  // Maps the entire file into read-only memory or returns nullptr on failure.
+  // We do not support offsets because Windows requires them to be a multiple of
+  // the allocation granularity, which is 64 KiB. Some implementations may fail
+  // if the file is zero-sized and return a nullptr.
+  virtual MapPtr Map() = 0;
 };
 
 // Returns nullptr on failure. `mode` is either "r" or "w+". This is not just
@@ -87,6 +96,7 @@ struct Path {
   std::string path;
 };
 
+// Aborts on error.
 static inline HWY_MAYBE_UNUSED std::string ReadFileToString(const Path& path) {
   std::unique_ptr<File> file = OpenFileOrNull(path, "r");
   if (!file) {

compression/io_win.cc

@@ -22,6 +22,7 @@
 #include <stdint.h>
 
 #include "compression/io.h"
+#include "util/allocator.h"
 #include "hwy/base.h"  // HWY_ASSERT
 #ifndef WIN32_LEAN_AND_MEAN
 #define WIN32_LEAN_AND_MEAN
@@ -96,6 +97,22 @@ class FileWin : public File {
     }
     return true;  // wrote everything => success
   }
+
+  MapPtr Map() override {
+    if (hFile_ == INVALID_HANDLE_VALUE) return MapPtr();
+
+    // Size=0 means the entire file.
+    HANDLE hMapping =
+        CreateFileMappingA(hFile_, nullptr, PAGE_READONLY, 0, 0, nullptr);
+    // Offset zero and size=0 means the entire file.
+    void* ptr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+    if (!ptr) return MapPtr();
+    return MapPtr(static_cast<const uint8_t*>(ptr),
+                  DeleterFunc2([hMapping](void* ptr) {
+                    HWY_ASSERT(UnmapViewOfFile(ptr));
+                    HWY_ASSERT(CloseHandle(hMapping));
+                  }));
+  }
 };  // FileWin
 
 std::unique_ptr<File> OpenFileOrNull(const Path& filename, const char* mode) {

compression/python/BUILD.bazel

@@ -32,7 +32,7 @@ pybind_extension(
     deps = [
         ":compression_clif_aux",
         "@abseil-cpp//absl/types:span",
-        "//compression:sfp",
+        "//compression:shared",
     ],
 )
 

paligemma/BUILD.bazel

@@ -43,7 +43,7 @@ cc_test(
         "//:benchmark_helper",
         "//:common",
         "//:gemma_lib",
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:hwy",
         "@highway//:hwy_test_util",
     ],

python/BUILD.bazel

@@ -13,7 +13,7 @@ pybind_extension(
     srcs = ["configs.cc"],
     deps = [
         "//:common",
-        "//compression:sfp",
+        "//compression:shared",
     ],
 )
 
@@ -25,7 +25,7 @@ pybind_extension(
         "//:benchmark_helper",
         "//:gemma_args",
         "//:gemma_lib",
-        "//compression:sfp",
+        "//compression:shared",
         "@highway//:hwy",
     ],
 )

util/args.h

@@ -181,6 +181,10 @@ class ArgsBase {
   void ForEach(Visitor& visitor) {
     static_cast<Args*>(this)->ForEach(visitor);
   }
+  template <class Visitor>
+  void ForEach(Visitor& visitor) const {
+    const_cast<ArgsBase*>(this)->ForEach(visitor);
+  }
 
  public:
   // WARNING: cannot call from ctor because the derived ctor has not yet run.
@@ -189,12 +193,12 @@ class ArgsBase {
     ForEach(visitor);
   }
 
-  void Help() {
+  void Help() const {
     HelpVisitor visitor;
     ForEach(visitor);
   }
 
-  void Print(int verbosity = 0) {
+  void Print(int verbosity = 0) const {
     PrintVisitor visitor(verbosity);
     ForEach(visitor);
   }

util/mat.h

@@ -112,6 +112,7 @@ class MatPtr : public IFields {
     type_ = type;
     element_bytes_ = static_cast<uint32_t>(hwy::DivCeil(TypeBits(type), 8));
     num_elements_ = static_cast<uint32_t>(ComputeNumElements(type, Extents()));
+    HWY_DASSERT(0 != element_bytes_ && element_bytes_ <= 16);
   }
 
   bool IsEmpty() const { return rows_ == 0 || cols_ == 0; }