Add+use CanLookup8 helper function

PiperOrigin-RevId: 910004461

Author: jan-wassenberg

g3doc/quick_reference.md

@@ -2568,7 +2568,8 @@ The following `ReverseN` must not be called if `Lanes(D()) < N`:
     `T` denote `TFromD<D>`. Only available if `HWY_TARGET != HWY_SCALAR` and
     `HWY_MIN_BYTES / sizeof(T) >= 4`. The latter is guaranteed if `T` is four
     bytes and `D` is one of `FixedTag<T, 16/sizeof(T)>` or `ScalableTag<T>` or
-    `CappedTag<T, N/sizeof(T)>` (where `N >= 16`).
+    `CappedTag<T, N/sizeof(T)>` (where `N >= 16`). The constexpr function
+    `CanLookup8(d)` is the preferred way to check these conditions.
 
 *   <code>unspecified **IndicesFromVec**(D d, V idx)</code> prepares for
     `TableLookupLanes` or `TwoTablesLookupLanes` with integer indices in `idx`,

hwy/contrib/math/fast_math-inl.h

@@ -229,11 +229,9 @@ HWY_INLINE V FastTan(D d, V x) {
   V x_red, sign;
   impl::ReduceAngleTan(d, x, x_red, sign);
 
-  constexpr size_t kLanes = HWY_MAX_LANES_D(D);
   V b, c, d_val;
 
-  if constexpr ((kLanes >= 4 && !HWY_HAVE_SCALABLE) ||
-                (HWY_HAVE_SCALABLE && sizeof(T) == 4 && detail::IsFull(d))) {
+  if constexpr (CanLookup8(d)) {
     // --- Table Lookup ---
     const auto scale = Set(d, static_cast<T>(3.8197186342));
     auto idx_float = Floor(Mul(x_red, scale));
@@ -525,7 +523,6 @@ HWY_INLINE V FastTanh(D d, V val) {
   // Abs(val) and preserve sign for later
   auto y = Abs(val);
 
-  constexpr size_t kLanes = HWY_MAX_LANES_D(D);
   V a, b, c, d_val, e, f;
 
   const auto t0 = Set(d, static_cast<T>(0.168236118310606));
@@ -536,8 +533,7 @@ HWY_INLINE V FastTanh(D d, V val) {
   const auto t5 = Set(d, static_cast<T>(2.969315202883957));
   const auto t6 = Set(d, static_cast<T>(4.734657601441978));
 
-  if constexpr ((kLanes >= 4 && !HWY_HAVE_SCALABLE) ||
-                (HWY_HAVE_SCALABLE && sizeof(T) == 4 && detail::IsFull(d))) {
+  if constexpr (CanLookup8(d)) {
     using DI = RebindToSigned<D>;
     auto idx_i = Zero(DI());
     const auto one_i = Set(DI(), 1);
@@ -919,7 +915,6 @@ HWY_INLINE V FastLog(D d, V x) {
   V y, exp;
   impl::FastLogRangeReduction<kHandleSubnormals>(d, x, y, exp);
 
-  constexpr size_t kLanes = HWY_MAX_LANES_D(D);
   V approx;
 
   V a, b, c, d_val;
@@ -928,8 +923,7 @@ HWY_INLINE V FastLog(D d, V x) {
   // log(y) directly.
   const V z = Sub(y, Set(d, static_cast<T>(1.0)));
 
-  if constexpr ((kLanes >= 4 && !HWY_HAVE_SCALABLE) ||
-                (HWY_HAVE_SCALABLE && sizeof(T) == 4 && detail::IsFull(d))) {
+  if constexpr (CanLookup8(d)) {
     // --- Table Lookup ---
     const auto scale = Set(d, static_cast<T>(11.3137085));
     // Input is always non-negative, so Floor() + ConvertTo()
@@ -1390,7 +1384,6 @@ HWY_INLINE V FastLog2(D d, V x) {
   V y, exp;
   impl::FastLogRangeReduction<kHandleSubnormals>(d, x, y, exp);
 
-  constexpr size_t kLanes = HWY_MAX_LANES_D(D);
   V approx;
 
   V a, b, c, d_val;
@@ -1399,8 +1392,7 @@ HWY_INLINE V FastLog2(D d, V x) {
   // log(y) directly.
   const V z = Sub(y, Set(d, static_cast<T>(1.0)));
 
-  if constexpr ((kLanes >= 4 && !HWY_HAVE_SCALABLE) ||
-                (HWY_HAVE_SCALABLE && sizeof(T) == 4 && detail::IsFull(d))) {
+  if constexpr (CanLookup8(d)) {
     // --- Table Lookup ---
     const auto scale = Set(d, static_cast<T>(11.3137085));
     auto idx_i = ConvertInRangeTo(
@@ -1628,7 +1620,6 @@ HWY_INLINE V FastLog10(D d, V x) {
   V y, exp;
   impl::FastLogRangeReduction<kHandleSubnormals>(d, x, y, exp);
 
-  constexpr size_t kLanes = HWY_MAX_LANES_D(D);
   V approx;
 
   V a, b, c, d_val;
@@ -1637,8 +1628,7 @@ HWY_INLINE V FastLog10(D d, V x) {
   // log(y) directly.
   const V z = Sub(y, Set(d, static_cast<T>(1.0)));
 
-  if constexpr ((kLanes >= 4 && !HWY_HAVE_SCALABLE) ||
-                (HWY_HAVE_SCALABLE && sizeof(T) == 4 && detail::IsFull(d))) {
+  if constexpr (CanLookup8(d)) {
     // --- Table Lookup ---
     const auto scale = Set(d, static_cast<T>(11.3137085));
     auto idx_i = ConvertInRangeTo(

hwy/ops/shared-inl.h

@@ -560,6 +560,20 @@ HWY_API bool IsAligned(D d, T* ptr) {
   return reinterpret_cast<uintptr_t>(ptr) % (N * sizeof(T)) == 0;
 }
 
+// Returns whether `Lookup8` can be used for vectors created from tag `d`.
+template <class D, typename T = TFromD<D>>
+HWY_API constexpr bool CanLookup8(D d) {
+  // `Lookup8` can use two-register tables, so it is sufficient to ensure
+  // vectors have at least four lanes (8/2). For fixed-length vectors: check
+  // `MaxLanes` directly. For scalable vectors, first require full
+  // (non-partial) vectors, which implies they are at least 128 bits. Then also
+  // require 16 or 32-bit elements, which implies at least 128/{16,32} =
+  // {8,4} lanes per vector. For 8-bit T, `TableLookupBytes` is more efficient.
+  return (!HWY_HAVE_SCALABLE && MaxLanes(d) >= 4) ||
+         (HWY_HAVE_SCALABLE && detail::IsFull(d) &&
+          (sizeof(T) == 2 || sizeof(T) == 4));
+}
+
 // ------------------------------ Choosing overloads (SFINAE)
 
 // Same as base.h macros but with a Simd<T, N, kPow2> argument instead of T.