Internal for now (because we might still need iterate a bit).

BUILD

@@ -702,3 +702,43 @@ test_suite(
     name = "hwy_ops_tests",
     tags = ["hwy_ops_test"],
 )
+
+# copybara:strip_begin(internal)
+cc_library(
+    name = "dual_tier_prefetch",
+    hdrs = [
+        "hwy/contrib/pipeline/dual_tier_prefetch.h",
+    ],
+    compatible_with = [],
+    copts = COPTS,
+    deps = [
+        ":hwy",
+    ],
+)
+
+cc_library(
+    name = "dual_tier_prefetch_2d",
+    hdrs = [
+        "hwy/contrib/pipeline/dual_tier_prefetch_2d.h",
+    ],
+    compatible_with = [],
+    copts = COPTS,
+    deps = [
+        ":dual_tier_prefetch",
+        ":hwy",
+    ],
+)
+
+cc_test(
+    name = "dual_tier_prefetch_benchmark",
+    size = "medium",
+    srcs = ["hwy/contrib/pipeline/dual_tier_prefetch_benchmark.cc"],
+    copts = COPTS + HWY_TEST_COPTS,
+    deps = [
+        ":dual_tier_prefetch",
+        ":hwy",
+        "//testing/base/public:gunit_main",
+        "//third_party/benchmark",
+    ],
+)
+# copybara:strip_end

hwy/cache_control.h

@@ -92,8 +92,10 @@ HWY_INLINE HWY_ATTR_CACHE void FlushStream() {
 #endif
 }
 
-// Optionally begins loading the cache line containing "p" to reduce latency of
-// subsequent actual loads.
+// Optionally begins loading the cache line containing "p" into all cache
+// levels, including L1, to reduce latency of subsequent actual loads. This
+// corresponds to the T0 temporal locality hint on x86, which is ideal when data
+// is about to be directly consumed.
 template <typename T>
 HWY_INLINE HWY_ATTR_CACHE void Prefetch(const T* p) {
   (void)p;
@@ -109,6 +111,38 @@ HWY_INLINE HWY_ATTR_CACHE void Prefetch(const T* p) {
 #endif  //  HWY_DISABLE_CACHE_CONTROL
 }
 
+// Begins loading the cache line containing "p" into the L1 cache only, passing
+// a Non-Temporal Access (NTA) hint. This minimizes pollution of outer memory
+// caches (L2/L3) and is ideal for data accessed exactly once.
+template <typename T>
+HWY_INLINE HWY_ATTR_CACHE void PrefetchNTA(const T* p) {
+  (void)p;
+#ifndef HWY_DISABLE_CACHE_CONTROL
+#if HWY_ARCH_X86 && !(HWY_COMPILER_CLANGCL && !defined(__MMX__))
+  _mm_prefetch(reinterpret_cast<const char*>(p), _MM_HINT_NTA);
+#elif HWY_COMPILER_GCC || HWY_COMPILER_CLANGCL  // includes clang
+  // Hint=0 specifically sets Non-Temporal local locality
+  __builtin_prefetch(p, /*write=*/0, /*hint=*/0);
+#endif
+#endif  //  HWY_DISABLE_CACHE_CONTROL
+}
+
+// Attempts to stage the cache line containing "p" into the L3/L2 outer caches
+// without aggressively staging it immediately into the L1. This restricts L1
+// and LFB thrashing on architectures like Intel when hiding massive DRAM delay.
+template <typename T>
+HWY_INLINE HWY_ATTR_CACHE void DeepPrefetch(const T* p) {
+  (void)p;
+#ifndef HWY_DISABLE_CACHE_CONTROL
+#if HWY_ARCH_X86 && !(HWY_COMPILER_CLANGCL && !defined(__MMX__))
+  _mm_prefetch(reinterpret_cast<const char*>(p), _MM_HINT_T2);
+#elif HWY_COMPILER_GCC || HWY_COMPILER_CLANGCL  // includes clang
+  // Hint=1 requests Moderate degrees of temporal locality (L2/L3 bounds)
+  __builtin_prefetch(p, /*write=*/0, /*hint=*/1);
+#endif
+#endif  //  HWY_DISABLE_CACHE_CONTROL
+}
+
 // Invalidates and flushes the cache line containing "p", if possible.
 HWY_INLINE HWY_ATTR_CACHE void FlushCacheline(const void* p) {
 #if HWY_ARCH_X86 && !defined(HWY_DISABLE_CACHE_CONTROL)