@img/sharp-libvips-dev-wasm32 1.3.0-rc.6 → 1.3.0-rc.7

package/include/hwy/ops/scalar-inl.h

@@ -253,13 +253,6 @@ HWY_API Vec1<T> operator^(const Vec1<T> a, const Vec1<T> b) {
   return Xor(a, b);
 }
 
-// ------------------------------ Xor3
-
-template <typename T>
-HWY_API Vec1<T> Xor3(Vec1<T> x1, Vec1<T> x2, Vec1<T> x3) {
-  return Xor(x1, Xor(x2, x3));
-}
-
 // ------------------------------ Or3
 
 template <typename T>
@@ -310,6 +303,17 @@ HWY_API Mask1<T> FirstN(D /*tag*/, size_t n) {
   return Mask1<T>::FromBool(n != 0);
 }
 
+#ifdef HWY_NATIVE_SET_MASK
+#undef HWY_NATIVE_SET_MASK
+#else
+#define HWY_NATIVE_SET_MASK
+#endif
+
+template <class D>
+HWY_API MFromD<D> SetMask(D /*d*/, bool val) {
+  return MFromD<D>::FromBool(val);
+}
+
 // ------------------------------ IfVecThenElse
 template <typename T>
 HWY_API Vec1<T> IfVecThenElse(Vec1<T> mask, Vec1<T> yes, Vec1<T> no) {
@@ -1673,6 +1677,17 @@ HWY_API V InterleaveOddBlocks(D, V a, V /*b*/) {
   return a;
 }
 
+// ------------------------------ InterleaveLowerBlocks
+template <class D, class V = VFromD<D>>
+HWY_API V InterleaveLowerBlocks(D, V a, V /*b*/) {
+  return a;
+}
+// ------------------------------ InterleaveUpperBlocks
+template <class D, class V = VFromD<D>>
+HWY_API V InterleaveUpperBlocks(D, V a, V /*b*/) {
+  return a;
+}
+
 // ------------------------------ TableLookupLanes
 
 // Returned by SetTableIndices for use by TableLookupLanes.

package/include/hwy/ops/set_macros-inl.h

@@ -1,5 +1,6 @@
 // Copyright 2020 Google LLC
-// Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// Copyright 2024-2025 Arm Limited and/or its affiliates
+// <open-source-office@arm.com>
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-License-Identifier: BSD-3-Clause
 //
@@ -34,18 +35,39 @@
 #undef HWY_NAMESPACE
 #undef HWY_ALIGN
 #undef HWY_MAX_BYTES
-#undef HWY_LANES
+#undef HWY_MIN_BYTES
 
 #undef HWY_HAVE_SCALABLE
 #undef HWY_HAVE_TUPLE
+#undef HWY_REGISTERS
 #undef HWY_HAVE_INTEGER64
 #undef HWY_HAVE_FLOAT16
 #undef HWY_HAVE_FLOAT64
 #undef HWY_MEM_OPS_MIGHT_FAULT
 #undef HWY_NATIVE_FMA
 #undef HWY_NATIVE_DOT_BF16
-#undef HWY_CAP_GE256
-#undef HWY_CAP_GE512
+#undef HWY_NATIVE_MASK
+#undef HWY_NATIVE_INTERLEAVE_WHOLE
+
+#ifndef HWY_CAP_GE256
+#define HWY_CAP_GE256 (HWY_MIN_BYTES >= 32)
+#endif
+#ifndef HWY_CAP_GE512
+#define HWY_CAP_GE512 (HWY_MIN_BYTES >= 64)
+#endif
+
+// Almost all targets (except RVV and SCALAR) use this definition.
+#undef HWY_LANES
+#define HWY_LANES(T) (HWY_MAX_BYTES / sizeof(T))
+
+// If 1, both __bf16 and a limited set of *_bf16 SVE intrinsics are available:
+// create/get/set/dup, ld/st, sel, rev, trn, uzp, zip.
+// Consulted below, hence define here rather than in arm_sve-inl.h.
+#if HWY_ARM_HAVE_SCALAR_BF16_TYPE && defined(__ARM_FEATURE_SVE_BF16)
+#define HWY_SVE_HAVE_BF16_FEATURE 1
+#else
+#define HWY_SVE_HAVE_BF16_FEATURE 0
+#endif
 
 #undef HWY_TARGET_IS_SVE
 #if HWY_TARGET & HWY_ALL_SVE
@@ -84,6 +106,13 @@
 #define HWY_HAVE_TUPLE 1
 #endif
 
+// Target-specific number of architectural vector registers available.
+#if !HWY_ARCH_X86 || (HWY_TARGET <= HWY_AVX3)
+#define HWY_REGISTERS 32
+#else
+#define HWY_REGISTERS 16
+#endif
+
 // For internal use (clamping/validating N for Simd<>)
 #undef HWY_MAX_N
 #if HWY_TARGET == HWY_SCALAR
@@ -141,57 +170,65 @@
 #define HWY_TARGET_STR_AVX2 \
   HWY_TARGET_STR_SSE4 ",avx,avx2" HWY_TARGET_STR_BMI2_FMA HWY_TARGET_STR_F16C
 
-#if (HWY_COMPILER_GCC_ACTUAL >= 1400 && HWY_COMPILER_GCC_ACTUAL < 1600) || \
-    HWY_COMPILER_CLANG >= 1800
+#ifndef HWY_HAVE_EVEX512  // allow override
+// evex512 has been removed from clang 22, see
+// https://github.com/llvm/llvm-project/pull/157034
+#if (1400 <= HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1600) || \
+    (1800 <= HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2200)
+#define HWY_HAVE_EVEX512 1
+#else
+#define HWY_HAVE_EVEX512 0
+#endif
+#endif
+
+#if (HWY_HAVE_EVEX512 == 1)
 #define HWY_TARGET_STR_AVX3_VL512 ",evex512"
 #else
 #define HWY_TARGET_STR_AVX3_VL512
 #endif
 
-#define HWY_TARGET_STR_AVX3_256 \
-  HWY_TARGET_STR_AVX2           \
+#define HWY_TARGET_STR_AVX3 \
+  HWY_TARGET_STR_AVX2       \
   ",avx512f,avx512cd,avx512vl,avx512dq,avx512bw" HWY_TARGET_STR_AVX3_VL512
 
-#define HWY_TARGET_STR_AVX3 HWY_TARGET_STR_AVX3_256 HWY_TARGET_STR_AVX3_VL512
-
-#define HWY_TARGET_STR_AVX3_DL_256                                   \
-  HWY_TARGET_STR_AVX3_256                                            \
+#define HWY_TARGET_STR_AVX3_DL                                       \
+  HWY_TARGET_STR_AVX3                                                \
   ",vpclmulqdq,avx512vbmi,avx512vbmi2,vaes,avx512vnni,avx512bitalg," \
   "avx512vpopcntdq,gfni"
 
-#define HWY_TARGET_STR_AVX3_DL \
-  HWY_TARGET_STR_AVX3_DL_256 HWY_TARGET_STR_AVX3_VL512
-
-// Force-disable for compilers that do not properly support avx512bf16.
-#if !defined(HWY_AVX3_DISABLE_AVX512BF16) &&                        \
+// Opt-out for compilers that do not properly support avx512bf16.
+#ifndef HWY_AVX3_ENABLE_AVX512BF16  // allow override
+// Default is to disable if the DISABLE macro is defined, or if old compiler.
+// clang-cl 21.1.4 reportedly works; feel free to define this to 1 there.
+#if defined(HWY_AVX3_DISABLE_AVX512BF16) ||                         \
     (HWY_COMPILER_CLANGCL ||                                        \
      (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000) || \
      (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 900))
-#define HWY_AVX3_DISABLE_AVX512BF16
+#define HWY_AVX3_ENABLE_AVX512BF16 0
+#else
+#define HWY_AVX3_ENABLE_AVX512BF16 1
 #endif
+#endif  // HWY_AVX3_ENABLE_AVX512BF16
 
-#if !defined(HWY_AVX3_DISABLE_AVX512BF16)
-#define HWY_TARGET_STR_AVX3_ZEN4_256 HWY_TARGET_STR_AVX3_DL ",avx512bf16"
+#if HWY_AVX3_ENABLE_AVX512BF16
+#define HWY_TARGET_STR_AVX3_ZEN4 HWY_TARGET_STR_AVX3_DL ",avx512bf16"
 #else
-#define HWY_TARGET_STR_AVX3_ZEN4_256 HWY_TARGET_STR_AVX3_DL
+#define HWY_TARGET_STR_AVX3_ZEN4 HWY_TARGET_STR_AVX3_DL
 #endif
 
-#define HWY_TARGET_STR_AVX3_ZEN4 \
-  HWY_TARGET_STR_AVX3_ZEN4_256 HWY_TARGET_STR_AVX3_VL512
-
 #if HWY_COMPILER_GCC_ACTUAL >= 1200 || HWY_COMPILER_CLANG >= 1400
-#define HWY_TARGET_STR_AVX3_SPR_256 HWY_TARGET_STR_AVX3_ZEN4_256 ",avx512fp16"
+#define HWY_TARGET_STR_AVX3_SPR HWY_TARGET_STR_AVX3_ZEN4 ",avx512fp16"
 #else
-#define HWY_TARGET_STR_AVX3_SPR_256 HWY_TARGET_STR_AVX3_ZEN4_256
+#define HWY_TARGET_STR_AVX3_SPR HWY_TARGET_STR_AVX3_ZEN4
 #endif
 
-#define HWY_TARGET_STR_AVX3_SPR \
-  HWY_TARGET_STR_AVX3_SPR_256 HWY_TARGET_STR_AVX3_VL512
-
-#if HWY_COMPILER_GCC_ACTUAL >= 1500
-#define HWY_TARGET_STR_AVX10_2 HWY_TARGET_STR_AVX3_SPR ",avx10.2"
-#elif HWY_COMPILER_CLANG >= 2000
+// Support for avx10.2-512 was removed between clang 22 and 23 without a
+// feature test macro.
+#if HWY_COMPILER_CLANG >= 2200 && HWY_HAVE_EVEX512
 #define HWY_TARGET_STR_AVX10_2 HWY_TARGET_STR_AVX3_SPR ",avx10.2-512"
+// Recent compilers drop the -512 suffix because 512 bits are always available.
+#elif HWY_COMPILER_GCC_ACTUAL >= 1500 || HWY_COMPILER_CLANG >= 2200
+#define HWY_TARGET_STR_AVX10_2 HWY_TARGET_STR_AVX3_SPR ",avx10.2"
 #else
 #define HWY_TARGET_STR_AVX10_2 HWY_TARGET_STR_AVX3_SPR
 #endif
@@ -234,7 +271,7 @@
 #define HWY_NAMESPACE N_SSE2
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -243,8 +280,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0  // a few actually are
 
 #define HWY_TARGET_STR HWY_TARGET_STR_SSE2
 //-----------------------------------------------------------------------------
@@ -254,7 +290,7 @@
 #define HWY_NAMESPACE N_SSSE3
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -263,8 +299,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0  // a few actually are
 
 #define HWY_TARGET_STR HWY_TARGET_STR_SSSE3
 
@@ -275,7 +310,7 @@
 #define HWY_NAMESPACE N_SSE4
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -284,8 +319,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0  // a few actually are
 
 #define HWY_TARGET_STR HWY_TARGET_STR_SSE4
 
@@ -296,7 +330,7 @@
 #define HWY_NAMESPACE N_AVX2
 #define HWY_ALIGN alignas(32)
 #define HWY_MAX_BYTES 32
-#define HWY_LANES(T) (32 / sizeof(T))
+#define HWY_MIN_BYTES 32
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -310,26 +344,22 @@
 #define HWY_NATIVE_FMA 1
 #endif
 #define HWY_NATIVE_DOT_BF16 0
-
-#define HWY_CAP_GE256 1
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0  // a few actually are
 
 #define HWY_TARGET_STR HWY_TARGET_STR_AVX2
 
 //-----------------------------------------------------------------------------
-// AVX3[_DL]/AVX10
-#elif HWY_TARGET == HWY_AVX3 || HWY_TARGET == HWY_AVX3_DL ||     \
-    HWY_TARGET == HWY_AVX3_ZEN4 || HWY_TARGET == HWY_AVX3_SPR || \
-    HWY_TARGET == HWY_AVX10_2
+// AVX3[_DL/ZEN4/SPR]/AVX10
+#elif HWY_TARGET <= HWY_AVX3
 
 #define HWY_ALIGN alignas(64)
 #define HWY_MAX_BYTES 64
-#define HWY_LANES(T) (64 / sizeof(T))
+#define HWY_MIN_BYTES 64
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
 #if HWY_TARGET <= HWY_AVX3_SPR &&                              \
-    (HWY_COMPILER_GCC_ACTUAL || HWY_COMPILER_CLANG >= 1901) && \
+    (HWY_COMPILER_GCC_ACTUAL || HWY_COMPILER_CLANG >= 2200) && \
     HWY_HAVE_SCALAR_F16_TYPE
 #define HWY_HAVE_FLOAT16 1
 #else
@@ -338,18 +368,12 @@
 #define HWY_HAVE_FLOAT64 1
 #define HWY_MEM_OPS_MIGHT_FAULT 0
 #define HWY_NATIVE_FMA 1
-#if (HWY_TARGET <= HWY_AVX3_ZEN4) && !defined(HWY_AVX3_DISABLE_AVX512BF16)
+#if (HWY_TARGET <= HWY_AVX3_ZEN4) && HWY_AVX3_ENABLE_AVX512BF16
 #define HWY_NATIVE_DOT_BF16 1
 #else
 #define HWY_NATIVE_DOT_BF16 0
 #endif
-#define HWY_CAP_GE256 1
-
-#if HWY_MAX_BYTES >= 64
-#define HWY_CAP_GE512 1
-#else
-#define HWY_CAP_GE512 0
-#endif
+#define HWY_NATIVE_MASK 1
 
 #if HWY_TARGET == HWY_AVX3
 
@@ -386,7 +410,7 @@
 
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -395,8 +419,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 1
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #if HWY_TARGET == HWY_PPC8
 
@@ -423,7 +446,7 @@
 
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -432,8 +455,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 1
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #if HWY_TARGET == HWY_Z14
 
@@ -478,7 +500,7 @@
 
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -508,8 +530,7 @@
 #define HWY_NATIVE_DOT_BF16 0
 #endif
 
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #if HWY_TARGET == HWY_NEON_WITHOUT_AES
 #define HWY_NAMESPACE N_NEON_WITHOUT_AES
@@ -553,12 +574,27 @@
 #define HWY_TARGET_STR_FP16 "+fp16"
 #endif
 
+#if HWY_OS_APPLE
+// Enable i8mm for the NEON_BF16 target if compiling for macOS, iOS, or iPadOS
+// as all Apple Silicon CPU's that support BF16 have support for I8MM.
+#define HWY_TARGET_STR_NEON_BF16_EXTRA "+i8mm"
+#else
+#define HWY_TARGET_STR_NEON_BF16_EXTRA ""
+#endif
+
 #if HWY_TARGET == HWY_NEON_WITHOUT_AES
+#if HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1400
+// Prevents inadvertent use of SVE by GCC 13.4 and earlier, see #2689.
+#define HWY_TARGET_STR "+nosve"
+#else
 // Do not define HWY_TARGET_STR (no pragma).
+#endif  // HWY_COMPILER_GCC_ACTUAL
 #elif HWY_TARGET == HWY_NEON
 #define HWY_TARGET_STR HWY_TARGET_STR_NEON
 #elif HWY_TARGET == HWY_NEON_BF16
-#define HWY_TARGET_STR HWY_TARGET_STR_FP16 "+bf16+dotprod" HWY_TARGET_STR_NEON
+#define HWY_TARGET_STR \
+  HWY_TARGET_STR_FP16  \
+      "+bf16+dotprod" HWY_TARGET_STR_NEON_BF16_EXTRA HWY_TARGET_STR_NEON
 #else
 #error "Logic error, missing case"
 #endif  // HWY_TARGET
@@ -575,51 +611,65 @@
 // SVE only requires lane alignment, not natural alignment of the entire vector.
 #define HWY_ALIGN alignas(8)
 
-// Value ensures MaxLanes() is the tightest possible upper bound to reduce
-// overallocation.
-#define HWY_LANES(T) ((HWY_MAX_BYTES) / sizeof(T))
-
 #define HWY_HAVE_INTEGER64 1
 #define HWY_HAVE_FLOAT16 1
 #define HWY_HAVE_FLOAT64 1
 #define HWY_MEM_OPS_MIGHT_FAULT 0
 #define HWY_NATIVE_FMA 1
-#if HWY_SVE_HAVE_BF16_FEATURE
+#if HWY_SVE_HAVE_BF16_FEATURE || HWY_TARGET == HWY_SVE2_128
 #define HWY_NATIVE_DOT_BF16 1
 #else
 #define HWY_NATIVE_DOT_BF16 0
 #endif
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 1
 
 #if HWY_TARGET == HWY_SVE2
 #define HWY_NAMESPACE N_SVE2
 #define HWY_MAX_BYTES 256
+#define HWY_MIN_BYTES 16
 #define HWY_HAVE_SCALABLE 1
 #elif HWY_TARGET == HWY_SVE_256
 #define HWY_NAMESPACE N_SVE_256
 #define HWY_MAX_BYTES 32
+#define HWY_MIN_BYTES 32
 #define HWY_HAVE_SCALABLE 0
 #elif HWY_TARGET == HWY_SVE2_128
 #define HWY_NAMESPACE N_SVE2_128
 #define HWY_MAX_BYTES 16
+#define HWY_MIN_BYTES 16
 #define HWY_HAVE_SCALABLE 0
 #else
 #define HWY_NAMESPACE N_SVE
 #define HWY_MAX_BYTES 256
+#define HWY_MIN_BYTES 16
 #define HWY_HAVE_SCALABLE 1
 #endif
 
-// Can use pragmas instead of -march compiler flag
-#if HWY_HAVE_RUNTIME_DISPATCH
-#if HWY_TARGET == HWY_SVE2 || HWY_TARGET == HWY_SVE2_128
+// Note: -march strings are delimited by + and GCC actually requires + before
+// each pragma target, which are also comma-separated.
+
+#undef HWY_TARGET_STR_SVE2_AES
 // Static dispatch with -march=armv8-a+sve2+aes, or no baseline, hence dynamic
 // dispatch, which checks for AES support at runtime.
 #if defined(__ARM_FEATURE_SVE2_AES) || (HWY_BASELINE_SVE2 == 0)
-#define HWY_TARGET_STR "+sve2+sve2-aes,+sve"
+#define HWY_TARGET_STR_SVE2_AES ",+sve2-aes"
 #else  // SVE2 without AES
-#define HWY_TARGET_STR "+sve2,+sve"
+#define HWY_TARGET_STR_SVE2_AES ""
 #endif
+
+#undef HWY_TARGET_STR_SVE2_128
+// SVE2_128 implies/requires I8MM and BF16, see #2973.
+#if HWY_TARGET == HWY_SVE2_128
+#define HWY_TARGET_STR_SVE2_128 ",+i8mm,+bf16"
+#else
+#define HWY_TARGET_STR_SVE2_128 ""
+#endif
+
+// Can use pragmas instead of -march compiler flag
+#if HWY_HAVE_RUNTIME_DISPATCH
+#if HWY_TARGET == HWY_SVE2 || HWY_TARGET == HWY_SVE2_128
+#define HWY_TARGET_STR \
+  "+sve,+sve2" HWY_TARGET_STR_SVE2_AES HWY_TARGET_STR_SVE2_128
 #else  // not SVE2 target
 #define HWY_TARGET_STR "+sve"
 #endif
@@ -633,7 +683,7 @@
 
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -642,8 +692,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #define HWY_NAMESPACE N_WASM
 
@@ -655,7 +704,7 @@
 
 #define HWY_ALIGN alignas(32)
 #define HWY_MAX_BYTES 32
-#define HWY_LANES(T) (32 / sizeof(T))
+#define HWY_MIN_BYTES 32
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -664,8 +713,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 1
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #define HWY_NAMESPACE N_WASM_EMU256
 
@@ -681,9 +729,11 @@
 
 // The spec requires VLEN <= 2^16 bits, so the limit is 2^16 bytes (LMUL=8).
 #define HWY_MAX_BYTES 65536
+#define HWY_MIN_BYTES 16
 
 // = HWY_MAX_BYTES divided by max LMUL=8 because MaxLanes includes the actual
 // LMUL. This is the tightest possible upper bound.
+#undef HWY_LANES
 #define HWY_LANES(T) (8192 / sizeof(T))
 
 #define HWY_HAVE_SCALABLE 1
@@ -692,8 +742,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 0
 #define HWY_NATIVE_FMA 1
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 1
 
 #if HWY_RVV_HAVE_F16_VEC
 #define HWY_HAVE_FLOAT16 1
@@ -717,12 +766,18 @@
 #if HWY_TARGET == HWY_LSX
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
+#define HWY_MIN_BYTES 16
+#ifndef __loongarch_sx
+#define HWY_TARGET_STR "lsx"
+#endif
 #else
 #define HWY_ALIGN alignas(32)
 #define HWY_MAX_BYTES 32
+#define HWY_MIN_BYTES 32
+#ifndef __loongarch_asx
+#define HWY_TARGET_STR "lsx,lasx"
+#endif
 #endif
-
-#define HWY_LANES(T) (HWY_MAX_BYTES / sizeof(T))
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -731,14 +786,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 1
 #define HWY_NATIVE_DOT_BF16 0
-
-#if HWY_TARGET == HWY_LSX
-#define HWY_CAP_GE256 0
-#else
-#define HWY_CAP_GE256 1
-#endif
-
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #if HWY_TARGET == HWY_LSX
 #define HWY_NAMESPACE N_LSX
@@ -754,7 +802,7 @@
 
 #define HWY_ALIGN alignas(16)
 #define HWY_MAX_BYTES 16
-#define HWY_LANES(T) (16 / sizeof(T))
+#define HWY_MIN_BYTES 16
 
 #define HWY_HAVE_SCALABLE 0
 #define HWY_HAVE_INTEGER64 1
@@ -763,8 +811,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 1
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #define HWY_NAMESPACE N_EMU128
 
@@ -776,6 +823,8 @@
 
 #define HWY_ALIGN
 #define HWY_MAX_BYTES 8
+#define HWY_MIN_BYTES 8
+#undef HWY_LANES
 #define HWY_LANES(T) 1
 
 #define HWY_HAVE_SCALABLE 0
@@ -785,8 +834,7 @@
 #define HWY_MEM_OPS_MIGHT_FAULT 0
 #define HWY_NATIVE_FMA 0
 #define HWY_NATIVE_DOT_BF16 0
-#define HWY_CAP_GE256 0
-#define HWY_CAP_GE512 0
+#define HWY_NATIVE_MASK 0
 
 #define HWY_NAMESPACE N_SCALAR
 
@@ -812,7 +860,7 @@
 
 // Clang <9 requires this be invoked at file scope, before any namespace.
 #undef HWY_BEFORE_NAMESPACE
-#if defined(HWY_TARGET_STR)
+#if defined(HWY_TARGET_STR) && !defined(HWY_DISABLE_ATTR)
 #define HWY_BEFORE_NAMESPACE()        \
   HWY_PUSH_ATTRIBUTES(HWY_TARGET_STR) \
   static_assert(true, "For requiring trailing semicolon")
@@ -824,7 +872,7 @@
 
 // Clang <9 requires any namespaces be closed before this macro.
 #undef HWY_AFTER_NAMESPACE
-#if defined(HWY_TARGET_STR)
+#if defined(HWY_TARGET_STR) && !defined(HWY_DISABLE_ATTR)
 #define HWY_AFTER_NAMESPACE() \
   HWY_POP_ATTRIBUTES          \
   static_assert(true, "For requiring trailing semicolon")
@@ -835,8 +883,16 @@
 #endif
 
 #undef HWY_ATTR
-#if defined(HWY_TARGET_STR) && HWY_HAS_ATTRIBUTE(target)
+#if defined(HWY_TARGET_STR) && HWY_HAS_ATTRIBUTE(target) && \
+    !defined(HWY_DISABLE_ATTR)
 #define HWY_ATTR __attribute__((target(HWY_TARGET_STR)))
 #else
 #define HWY_ATTR
 #endif
+
+#if (HWY_MAX_BYTES <= 16) || HWY_TARGET_IS_SVE || (HWY_TARGET == HWY_RVV) || \
+    (HWY_TARGET == HWY_WASM_EMU256)
+#define HWY_NATIVE_INTERLEAVE_WHOLE 1
+#else
+#define HWY_NATIVE_INTERLEAVE_WHOLE 0
+#endif