numkong 7.0.0 → 7.4.2

package/include/numkong/cast/README.md

@@ -6,17 +6,17 @@ All conversions use round-to-nearest-even (RNE) for narrowing and exact widening
 
 BFloat16 relates to Float32 by truncation with rounding:
 
-```math
+$$
 \text{bf16} \approx \text{f32} \gg 16
-```
+$$
 
 With RNE tie-breaking to preserve the least significant bit of the truncated result.
 
 Float16 range and precision:
 
-```math
+$$
 \text{f16} \in [-65504, 65504], \quad \text{min positive normal} = 2^{-14}
-```
+$$
 
 Reformulating as Python pseudocode:
 
@@ -194,69 +194,69 @@ Measured with Wasmtime v42 (Cranelift backend).
 | __f32 ↔ e4m3__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░ |
 | `nk_cast_serial` |       ? gb/s |       ? gb/s |      0.239 gb/s |       ? gb/s |       ? gb/s |      0.746 gb/s |
 
-### Apple M4
+### Apple M5
 
 #### Native
 
 | Kernel           |        ↓ 256 |         ↓ 1K |         ↓ 4K |        ↑ 256 |         ↑ 1K |         ↑ 4K |
 | :--------------- | -----------: | -----------: | -----------: | -----------: | -----------: | -----------: |
 | __f32 ↔ bf16__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    10.2 gb/s |    10.6 gb/s |    10.7 gb/s |    8.06 gb/s |    8.34 gb/s |    8.32 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    1.37 gb/s |    1.35 gb/s |    1.41 gb/s |    1.37 gb/s |    1.34 gb/s |    1.38 gb/s |
+| `nk_cast_neon`   |    19.3 gb/s |    23.7 gb/s |    23.2 gb/s |    59.4 gb/s |    58.9 gb/s |    57.3 gb/s |
 | __f32 ↔ f16__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    10.9 gb/s |    11.3 gb/s |    11.4 gb/s |    8.40 gb/s |    8.62 gb/s |    8.70 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    1.37 gb/s |    1.31 gb/s |    1.32 gb/s |    1.37 gb/s |    1.31 gb/s |    1.40 gb/s |
+| `nk_cast_neon`   |    20.1 gb/s |    21.9 gb/s |    25.0 gb/s |    52.1 gb/s |    60.2 gb/s |    70.2 gb/s |
 | __f32 ↔ e5m2__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    1.65 gb/s |    1.52 gb/s |    1.36 gb/s |    5.96 gb/s |    6.08 gb/s |    6.11 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.681 gb/s |   0.621 gb/s |   0.600 gb/s |    1.17 gb/s |    1.17 gb/s |    1.23 gb/s |
+| `nk_cast_neon`   |    8.50 gb/s |    8.45 gb/s |    8.35 gb/s |    40.6 gb/s |    46.5 gb/s |    46.5 gb/s |
 | __f32 ↔ e4m3__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    1.49 gb/s |    1.36 gb/s |    1.24 gb/s |    4.96 gb/s |    5.05 gb/s |    4.81 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.683 gb/s |   0.618 gb/s |   0.586 gb/s |    1.02 gb/s |    1.01 gb/s |    1.02 gb/s |
+| `nk_cast_neon`   |    7.85 gb/s |    7.91 gb/s |    7.66 gb/s |    18.9 gb/s |    19.2 gb/s |    18.3 gb/s |
 | __f32 ↔ e3m2__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    2.17 gb/s |    2.13 gb/s |    1.97 gb/s |    5.90 gb/s |    6.02 gb/s |    6.07 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.702 gb/s |   0.632 gb/s |   0.596 gb/s |    1.17 gb/s |    1.13 gb/s |    1.15 gb/s |
+| `nk_cast_neon`   |    8.94 gb/s |    9.02 gb/s |    8.91 gb/s |    24.9 gb/s |    25.0 gb/s |    24.4 gb/s |
 | __f32 ↔ e2m3__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    2.54 gb/s |    2.45 gb/s |    2.23 gb/s |    5.88 gb/s |    6.11 gb/s |    6.10 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.921 gb/s |   0.843 gb/s |   0.715 gb/s |    1.21 gb/s |    1.21 gb/s |    1.26 gb/s |
+| `nk_cast_neon`   |    8.89 gb/s |    9.03 gb/s |    8.82 gb/s |    24.9 gb/s |    25.1 gb/s |    24.6 gb/s |
 | __f32 ↔ i16__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    6.13 gb/s |    5.99 gb/s |    6.10 gb/s |    8.29 gb/s |    8.53 gb/s |    8.58 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.785 gb/s |   0.679 gb/s |   0.678 gb/s |    1.44 gb/s |    1.39 gb/s |    1.49 gb/s |
+| `nk_cast_neon`   |    19.4 gb/s |    22.6 gb/s |    23.9 gb/s |    19.9 gb/s |    23.2 gb/s |    25.9 gb/s |
 | __f32 ↔ u16__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    5.36 gb/s |    5.01 gb/s |    4.49 gb/s |    8.43 gb/s |    8.64 gb/s |    8.76 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.916 gb/s |   0.822 gb/s |   0.726 gb/s |    1.37 gb/s |    1.36 gb/s |    1.48 gb/s |
+| `nk_cast_neon`   |    20.3 gb/s |    20.6 gb/s |    22.1 gb/s |    15.6 gb/s |    18.5 gb/s |    17.4 gb/s |
 | __f32 ↔ i8__     | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    4.83 gb/s |    4.89 gb/s |    5.09 gb/s |    6.67 gb/s |    6.92 gb/s |    7.08 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.725 gb/s |   0.616 gb/s |   0.578 gb/s |    1.21 gb/s |    1.21 gb/s |    1.28 gb/s |
+| `nk_cast_neon`   |    18.2 gb/s |    24.5 gb/s |    21.7 gb/s |    16.3 gb/s |    18.9 gb/s |    19.8 gb/s |
 | __f32 ↔ u8__     | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    4.31 gb/s |    4.10 gb/s |    3.62 gb/s |    7.03 gb/s |    7.21 gb/s |    7.28 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |   0.967 gb/s |   0.795 gb/s |   0.723 gb/s |    1.29 gb/s |    1.25 gb/s |    1.40 gb/s |
+| `nk_cast_neon`   |    17.5 gb/s |    19.8 gb/s |    19.4 gb/s |    13.8 gb/s |    17.8 gb/s |    15.1 gb/s |
 | __f64 ↔ f32__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    17.3 gb/s |    17.8 gb/s |    18.1 gb/s |    17.9 gb/s |    18.5 gb/s |    18.5 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    2.65 gb/s |    2.60 gb/s |    2.70 gb/s |    2.59 gb/s |    2.55 gb/s |    2.65 gb/s |
+| `nk_cast_neon`   |    2.87 gb/s |    2.60 gb/s |    2.73 gb/s |    2.64 gb/s |    2.63 gb/s |    2.57 gb/s |
 | __f64 ↔ i64__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    16.8 gb/s |    17.2 gb/s |    17.0 gb/s |    23.9 gb/s |    24.7 gb/s |    24.8 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    2.42 gb/s |    2.00 gb/s |    1.86 gb/s |    3.79 gb/s |    3.61 gb/s |    4.03 gb/s |
+| `nk_cast_neon`   |    2.51 gb/s |    1.94 gb/s |    1.78 gb/s |    3.83 gb/s |    3.68 gb/s |    3.79 gb/s |
 | __f64 ↔ u64__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    13.5 gb/s |    12.8 gb/s |    11.3 gb/s |    24.4 gb/s |    25.0 gb/s |    25.1 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    2.56 gb/s |    2.19 gb/s |    2.06 gb/s |    3.71 gb/s |    3.50 gb/s |    3.87 gb/s |
+| `nk_cast_neon`   |    2.68 gb/s |    2.10 gb/s |    1.97 gb/s |    3.68 gb/s |    3.61 gb/s |    3.58 gb/s |
 | __f64 ↔ i32__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    12.1 gb/s |    12.4 gb/s |    12.6 gb/s |    18.2 gb/s |    18.9 gb/s |    19.2 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    1.58 gb/s |    1.32 gb/s |    1.29 gb/s |    2.65 gb/s |    2.58 gb/s |    2.84 gb/s |
+| `nk_cast_neon`   |    1.61 gb/s |    1.33 gb/s |    1.24 gb/s |    2.73 gb/s |    2.63 gb/s |    2.66 gb/s |
 | __f64 ↔ u32__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    10.9 gb/s |    10.6 gb/s |    9.58 gb/s |    17.6 gb/s |    18.0 gb/s |    18.1 gb/s |
-| `nk_cast_neon`   |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |       ? gb/s |
+| `nk_cast_serial` |    1.83 gb/s |    1.53 gb/s |    1.47 gb/s |    2.55 gb/s |    2.48 gb/s |    2.69 gb/s |
+| `nk_cast_neon`   |    1.89 gb/s |    1.53 gb/s |    1.38 gb/s |    2.56 gb/s |    2.54 gb/s |    2.59 gb/s |
 
 #### WASM
 
-Measured with Wasmtime v42 (Cranelift backend).
+Measured with Wasmtime v43 (Cranelift backend).
 
 | Kernel           |        ↓ 256 |         ↓ 1K |         ↓ 4K |        ↑ 256 |         ↑ 1K |         ↑ 4K |
 | :--------------- | -----------: | -----------: | -----------: | -----------: | -----------: | -----------: |
 | __f32 ↔ bf16__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |    3.39 gb/s |    7.29 gb/s |    11.2 gb/s |    3.08 gb/s |    6.20 gb/s |    8.81 gb/s |
+| `nk_cast_serial` |   0.514 gb/s |   0.522 gb/s |   0.538 gb/s |   0.511 gb/s |   0.526 gb/s |   0.519 gb/s |
 | __f32 ↔ f16__    | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |   0.605 gb/s |   0.952 gb/s |    1.22 gb/s |    2.36 gb/s |    4.71 gb/s |    7.31 gb/s |
+| `nk_cast_serial` |   0.368 gb/s |   0.363 gb/s |   0.360 gb/s |   0.490 gb/s |   0.480 gb/s |   0.489 gb/s |
 | __f32 ↔ e5m2__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |   0.752 gb/s |    1.84 gb/s |    1.80 gb/s |    2.24 gb/s |    6.32 gb/s |    6.31 gb/s |
+| `nk_cast_serial` |   0.323 gb/s |   0.312 gb/s |   0.304 gb/s |   0.423 gb/s |   0.425 gb/s |   0.425 gb/s |
 | __f32 ↔ e4m3__   | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ | ░░░░░░░░░░░░ |
-| `nk_cast_serial` |   0.623 gb/s |    1.61 gb/s |    1.50 gb/s |    1.68 gb/s |    4.35 gb/s |    4.28 gb/s |
+| `nk_cast_serial` |   0.315 gb/s |   0.304 gb/s |   0.295 gb/s |   0.396 gb/s |   0.396 gb/s |   0.397 gb/s |

package/include/numkong/cast/diamond.h

@@ -0,0 +1,64 @@
+/**
+ *  @brief SIMD-accelerated Type Conversions for Diamond Rapids.
+ *  @file include/numkong/cast/diamond.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/cast/icelake.h
+ *
+ *  Uses VCVTHF82PH (E4M3→FP16) and VCVTBF82PH (E5M2→FP16) for native 1-instruction
+ *  FP8→FP16 conversion. Both conversions are exact (no rounding needed).
+ */
+#ifndef NK_CAST_DIAMOND_H
+#define NK_CAST_DIAMOND_H
+
+#if NK_TARGET_X86_
+#if NK_TARGET_DIAMOND
+
+#include "numkong/types.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute push(                                                                                    \
+    __attribute__((target("avx2,avx512f,avx512vl,avx512bw,avx512dq,avx512fp16,avx10.2-512,f16c,fma,bmi,bmi2"))), \
+    apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("avx2", "avx512f", "avx512vl", "avx512bw", "avx512dq", "avx512fp16", "avx10.2-512", "f16c", "fma", \
+                   "bmi", "bmi2")
+#endif
+
+NK_INTERNAL void nk_load_e4m3x32_to_f16x32_diamond_(nk_e4m3_t const *src, nk_b512_vec_t *dst) {
+    dst->zmm_ph = _mm512_cvthf8_ph(_mm256_loadu_epi8(src));
+}
+
+NK_INTERNAL void nk_partial_load_e4m3x32_to_f16x32_diamond_(nk_e4m3_t const *src, nk_b512_vec_t *dst, nk_size_t count) {
+    __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, count);
+    dst->zmm_ph = _mm512_cvthf8_ph(_mm256_maskz_loadu_epi8(mask, src));
+}
+
+NK_INTERNAL void nk_load_e5m2x32_to_f16x32_diamond_(nk_e5m2_t const *src, nk_b512_vec_t *dst) {
+    dst->zmm_ph = _mm512_cvtbf8_ph(_mm256_loadu_epi8(src));
+}
+
+NK_INTERNAL void nk_partial_load_e5m2x32_to_f16x32_diamond_(nk_e5m2_t const *src, nk_b512_vec_t *dst, nk_size_t count) {
+    __mmask32 mask = (__mmask32)_bzhi_u32(0xFFFFFFFF, count);
+    dst->zmm_ph = _mm512_cvtbf8_ph(_mm256_maskz_loadu_epi8(mask, src));
+}
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_DIAMOND
+#endif // NK_TARGET_X86_
+#endif // NK_CAST_DIAMOND_H

package/include/numkong/cast/haswell.h

@@ -6,12 +6,12 @@
  *
  *  @section haswell_cast_instructions Key F16C/AVX2 Conversion Instructions
  *
- *      Intrinsic                   Instruction                     Latency     Throughput  Ports
- *      _mm256_cvtph_ps             VCVTPH2PS (YMM, XMM)            5cy         1/cy        p01
- *      _mm256_cvtps_ph             VCVTPS2PH (XMM, YMM, I8)        4cy         1/cy        p01+p5
- *      _mm256_cvtepi16_epi32       VPMOVSXWD (YMM, XMM)            3cy         1/cy        p5
- *      _mm256_slli_epi32           VPSLLD (YMM, YMM, I8)           1cy         0.5/cy      p01
- *      _mm256_blendv_ps            VBLENDVPS (YMM, YMM, YMM, YMM)  2cy         1/cy        p015
+ *      Intrinsic              Instruction                     Haswell     Genoa
+ *      _mm256_cvtph_ps        VCVTPH2PS (YMM, XMM)            5cy @ p01   4cy @ p12+p23
+ *      _mm256_cvtps_ph        VCVTPS2PH (XMM, YMM, I8)        5cy @ p01   4cy @ p12+p23
+ *      _mm256_cvtepi16_epi32  VPMOVSXWD (YMM, XMM)            1cy @ p5    2cy @ p12
+ *      _mm256_slli_epi32      VPSLLD (YMM, YMM, I8)           1cy @ p0    1cy @ p23
+ *      _mm256_blendv_ps       VBLENDVPS (YMM, YMM, YMM, YMM)  2cy @ p015  1cy @ p01
  *
  *  F16C provides hardware F16<->F32 conversion. BF16 lacks hardware support and is emulated via
  *  bit manipulation (shift upper 16 bits). FP8 formats (E4M3/E5M2) use lookup tables for subnormal
@@ -38,14 +38,14 @@ extern "C" {
 #endif
 
 NK_PUBLIC void nk_f32_to_f16_haswell(nk_f32_t const *from, nk_f16_t *to) {
-    *to = _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(*from), _MM_FROUND_TO_NEAREST_INT));
+    *(nk_u16_t *)to = (nk_u16_t)_mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(*from), _MM_FROUND_TO_NEAREST_INT));
 }
 
 NK_PUBLIC void nk_f16_to_f32_haswell(nk_f16_t const *from, nk_f32_t *to) {
-    *to = _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(*from)));
+    *to = _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(*(nk_u16_t const *)from)));
 }
 
-#pragma region - Type Punned Loads and Stores
+#pragma region Type Punned Loads and Stores
 
 /** @brief Type-agnostic 256-bit full load (Haswell AVX2). */
 NK_INTERNAL void nk_load_b256_haswell_(void const *src, nk_b256_vec_t *dst) {
@@ -99,9 +99,9 @@ NK_INTERNAL void nk_partial_store_b64x4_haswell_(nk_b256_vec_t const *src, void
     _mm256_maskstore_pd((double *)dst, mask_i64x4, _mm256_castsi256_pd(src->ymm));
 }
 
-#pragma endregion - Type Punned Loads and Stores
+#pragma endregion Type Punned Loads and Stores
 
-#pragma region - Vectorized Conversions
+#pragma region Vectorized Conversions
 
 /** @brief Convert 8x bf16 → 8x f32 by shifting left 16 bits (AVX2). */
 NK_INTERNAL __m256 nk_bf16x8_to_f32x8_haswell_(__m128i bf16_i16x8) {
@@ -116,9 +116,9 @@ NK_INTERNAL __m128i nk_f32x8_to_bf16x8_haswell_(__m256 f32x8) {
     __m256i rounded_i32x8 = _mm256_add_epi32(bits_i32x8, _mm256_add_epi32(_mm256_set1_epi32(0x7FFF), lsb_i32x8));
     __m256i bf16_i32x8 = _mm256_srli_epi32(rounded_i32x8, 16);
     // Pack 8x i32 to 8x i16
-    __m128i lo_i32x4 = _mm256_castsi256_si128(bf16_i32x8);
-    __m128i hi_i32x4 = _mm256_extracti128_si256(bf16_i32x8, 1);
-    return _mm_packus_epi32(lo_i32x4, hi_i32x4);
+    __m128i low_i32x4 = _mm256_castsi256_si128(bf16_i32x8);
+    __m128i high_i32x4 = _mm256_extracti128_si256(bf16_i32x8, 1);
+    return _mm_packus_epi32(low_i32x4, high_i32x4);
 }
 
 /** @brief Integer upcasts to f32x8 (AVX2). */
@@ -132,10 +132,10 @@ NK_INTERNAL __m256 nk_u16x8_to_f32x8_haswell_(__m128i u16x8) {
 }
 NK_INTERNAL __m256 nk_i32x8_to_f32x8_haswell_(__m256i i32x8) { return _mm256_cvtepi32_ps(i32x8); }
 NK_INTERNAL __m256 nk_u32x8_to_f32x8_haswell_(__m256i u32x8) {
-    __m256i lo_i32x8 = _mm256_and_si256(u32x8, _mm256_set1_epi32(0xFFFF));
-    __m256i hi_i32x8 = _mm256_srli_epi32(u32x8, 16);
-    return _mm256_add_ps(_mm256_cvtepi32_ps(lo_i32x8),
-                         _mm256_mul_ps(_mm256_cvtepi32_ps(hi_i32x8), _mm256_set1_ps(65536.0f)));
+    __m256i low_i32x8 = _mm256_and_si256(u32x8, _mm256_set1_epi32(0xFFFF));
+    __m256i high_i32x8 = _mm256_srli_epi32(u32x8, 16);
+    return _mm256_add_ps(_mm256_cvtepi32_ps(low_i32x8),
+                         _mm256_mul_ps(_mm256_cvtepi32_ps(high_i32x8), _mm256_set1_ps(65536.0f)));
 }
 
 /** @brief Saturating f32x8 downcasts to integers (AVX2). */
@@ -172,167 +172,10 @@ NK_INTERNAL __m128i nk_f32x8_to_u8x8_haswell_(__m256 f32x8) {
     return _mm_packus_epi16(u16x8, _mm_setzero_si128());
 }
 
-/** @brief Convert 16x e4m3 → 16x bf16 via arithmetic + small LUT for subnormals (AVX2).
- *  E4M3 format: S EEEE MMM (bias=7). BF16: S EEEEEEEE MMMMMMM (bias=127).
- *  Normal values: BF16 = sign | ((lower7 << 4) + 0x3C00).
- *  Subnormals (8 values): looked up via vpshufb from an 8-entry LUT.
- *  Handles all corner cases: zero, subnormals, normals, and NaN. */
-NK_INTERNAL __m256i nk_e4m3x16_to_bf16x16_haswell_(__m128i e4m3x16) {
-    __m256i e4m3_i16x16 = _mm256_cvtepu8_epi16(e4m3x16);
-    __m256i sign_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16((short)0x80));
-    __m256i lower7_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16(0x7F));
-
-    // Normal path: BF16 = ((lower7 << 4) + 0x3C00) | (sign << 8)
-    __m256i normal_abs_i16x16 = _mm256_add_epi16(_mm256_slli_epi16(lower7_i16x16, 4), _mm256_set1_epi16(0x3C00));
-    sign_i16x16 = _mm256_slli_epi16(sign_i16x16, 8);
-    __m256i normal_i16x16 = _mm256_or_si256(sign_i16x16, normal_abs_i16x16);
-
-    // Subnormal LUT via shuffle_epi8 (8 entries: mantissa 0-7 → BF16)
-    // E4M3 subnormal BF16 values: 0x0000, 0x3B00, 0x3B80, 0x3BC0, 0x3C00, 0x3C20, 0x3C40, 0x3C60
-    // Split into low bytes and high bytes for reconstruction
-    __m256i const lo_lut_i8x32 = _mm256_broadcastsi128_si256(_mm_set_epi8( //
-        0x60, 0x40, 0x20, 0x00, (char)0xC0, (char)0x80, 0x00, 0x00,        //
-        0x60, 0x40, 0x20, 0x00, (char)0xC0, (char)0x80, 0x00, 0x00));      //
-    __m256i const hi_lut_i8x32 = _mm256_broadcastsi128_si256(_mm_set_epi8( //
-        0x3C, 0x3C, 0x3C, 0x3C, 0x3B, 0x3B, 0x3B, 0x00,                    //
-        0x3C, 0x3C, 0x3C, 0x3C, 0x3B, 0x3B, 0x3B, 0x00));                  //
-
-    // Extract mantissa (bits 0-2) as byte indices for shuffle
-    __m256i byte_idx_i8x32 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi8(0x07));
-    __m256i lo_bytes_i8x32 = _mm256_shuffle_epi8(lo_lut_i8x32, byte_idx_i8x32);
-    __m256i hi_bytes_i8x32 = _mm256_shuffle_epi8(hi_lut_i8x32, byte_idx_i8x32);
-
-    // Combine low and high bytes into 16-bit values
-    __m256i subnorm_abs_i16x16 = _mm256_or_si256(                    //
-        _mm256_and_si256(lo_bytes_i8x32, _mm256_set1_epi16(0x00FF)), //
-        _mm256_slli_epi16(hi_bytes_i8x32, 8));                       //
-    __m256i subnorm_i16x16 = _mm256_or_si256(subnorm_abs_i16x16, sign_i16x16);
-
-    // Blend: if exponent == 0, use subnormal result; else use normal result
-    __m256i exp_bits_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16(0x78));
-    __m256i is_subnormal_i16x16 = _mm256_cmpeq_epi16(exp_bits_i16x16, _mm256_setzero_si256());
-    __m256i result_i16x16 = _mm256_blendv_epi8(normal_i16x16, subnorm_i16x16, is_subnormal_i16x16);
-
-    // Handle NaN: E4M3 index 127 (0x7F) → BF16 NaN (0x7FC0)
-    __m256i is_nan_i16x16 = _mm256_cmpeq_epi16(lower7_i16x16, _mm256_set1_epi16(0x7F));
-    __m256i nan_i16x16 = _mm256_or_si256(sign_i16x16, _mm256_set1_epi16(0x7FC0));
-    return _mm256_blendv_epi8(result_i16x16, nan_i16x16, is_nan_i16x16);
-}
-
-/** @brief Convert 16x e5m2 → 16x bf16 via arithmetic + small LUT for subnormals (AVX2).
- *  E5M2 format: S EEEEE MM (bias=15). BF16: S EEEEEEEE MMMMMMM (bias=127).
- *  Normal values: BF16 = sign | ((lower7 << 5) + 0x3800).
- *  Subnormals (4 values): looked up via vpshufb from a 4-entry LUT.
- *  Handles all corner cases: zero, subnormals, normals, infinity, and NaN. */
-NK_INTERNAL __m256i nk_e5m2x16_to_bf16x16_haswell_(__m128i e5m2x16) {
-    __m256i e5m2_i16x16 = _mm256_cvtepu8_epi16(e5m2x16);
-    __m256i sign_i16x16 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi16((short)0x80));
-    __m256i lower7_i16x16 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi16(0x7F));
-
-    // Normal path: BF16 = ((lower7 << 5) + 0x3800) | (sign << 8)
-    __m256i normal_abs_i16x16 = _mm256_add_epi16(_mm256_slli_epi16(lower7_i16x16, 5), _mm256_set1_epi16(0x3800));
-    sign_i16x16 = _mm256_slli_epi16(sign_i16x16, 8);
-    __m256i normal_i16x16 = _mm256_or_si256(sign_i16x16, normal_abs_i16x16);
-
-    // Subnormal LUT via shuffle_epi8 (4 entries: mantissa 0-3 → BF16)
-    // E5M2 subnormal BF16 values: 0x0000, 0x3780, 0x3800, 0x3840
-    __m256i const lo_lut_i8x32 = _mm256_broadcastsi128_si256(_mm_set_epi8( //
-        0x00, 0x00, 0x00, 0x00, 0x40, 0x00, (char)0x80, 0x00,              //
-        0x00, 0x00, 0x00, 0x00, 0x40, 0x00, (char)0x80, 0x00));            //
-    __m256i const hi_lut_i8x32 = _mm256_broadcastsi128_si256(_mm_set_epi8( //
-        0x00, 0x00, 0x00, 0x00, 0x38, 0x38, 0x37, 0x00,                    //
-        0x00, 0x00, 0x00, 0x00, 0x38, 0x38, 0x37, 0x00));                  //
-
-    // Extract mantissa (bits 0-1) as byte indices for shuffle
-    __m256i byte_idx_i8x32 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi8(0x03));
-    __m256i lo_bytes_i8x32 = _mm256_shuffle_epi8(lo_lut_i8x32, byte_idx_i8x32);
-    __m256i hi_bytes_i8x32 = _mm256_shuffle_epi8(hi_lut_i8x32, byte_idx_i8x32);
-
-    // Combine low and high bytes into 16-bit values
-    __m256i subnorm_abs_i16x16 = _mm256_or_si256(                    //
-        _mm256_and_si256(lo_bytes_i8x32, _mm256_set1_epi16(0x00FF)), //
-        _mm256_slli_epi16(hi_bytes_i8x32, 8));                       //
-    __m256i subnorm_i16x16 = _mm256_or_si256(subnorm_abs_i16x16, sign_i16x16);
-
-    // Blend: if exponent == 0, use subnormal result; else use normal result
-    __m256i exp_bits_i16x16 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi16(0x7C));
-    __m256i is_subnormal_i16x16 = _mm256_cmpeq_epi16(exp_bits_i16x16, _mm256_setzero_si256());
-    __m256i result_i16x16 = _mm256_blendv_epi8(normal_i16x16, subnorm_i16x16, is_subnormal_i16x16);
-
-    // Handle Inf (0x7C) and NaN (0x7D-0x7F)
-    __m256i is_inf_i16x16 = _mm256_cmpeq_epi16(lower7_i16x16, _mm256_set1_epi16(0x7C));
-    __m256i is_nan_i16x16 = _mm256_cmpgt_epi16(lower7_i16x16, _mm256_set1_epi16(0x7C));
-    __m256i inf_i16x16 = _mm256_or_si256(sign_i16x16, _mm256_set1_epi16(0x7F80));
-    __m256i nan_i16x16 = _mm256_or_si256(sign_i16x16, _mm256_set1_epi16(0x7FC0));
-    result_i16x16 = _mm256_blendv_epi8(result_i16x16, inf_i16x16, is_inf_i16x16);
-    return _mm256_blendv_epi8(result_i16x16, nan_i16x16, is_nan_i16x16);
-}
-
-/** @brief Convert 16x e4m3 → 16x f16 via arithmetic + small LUT for subnormals (AVX2).
- *  E4M3 format: S EEEE MMM (bias=7). F16: S EEEEE MMMMMMMMMM (bias=15).
- *  Normal values: F16 = sign | ((lower7 << 7) + 0x2000).
- *  Subnormals (8 values): looked up via vpshufb from an 8-entry LUT.
- *  Handles all corner cases: zero, subnormals, normals, and NaN. */
-NK_INTERNAL __m256i nk_e4m3x16_to_f16x16_haswell_(__m128i e4m3x16) {
-    __m256i e4m3_i16x16 = _mm256_cvtepu8_epi16(e4m3x16);
-    __m256i sign_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16((short)0x80));
-    __m256i lower7_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16(0x7F));
-
-    // Normal path: F16 = ((lower7 << 7) + 0x2000) | (sign << 8)
-    __m256i normal_abs_i16x16 = _mm256_add_epi16(_mm256_slli_epi16(lower7_i16x16, 7), _mm256_set1_epi16(0x2000));
-    sign_i16x16 = _mm256_slli_epi16(sign_i16x16, 8);
-    __m256i normal_i16x16 = _mm256_or_si256(sign_i16x16, normal_abs_i16x16);
-
-    // Subnormal LUT via shuffle_epi8 (8 entries: mantissa 0-7 → F16)
-    // E4M3 subnormal F16 values: 0x0000, 0x1800, 0x1C00, 0x1E00, 0x2000, 0x2100, 0x2200, 0x2300
-    // All low bytes are 0x00, high bytes: 0x00, 0x18, 0x1C, 0x1E, 0x20, 0x21, 0x22, 0x23
-    // _mm_set_epi8 order: b15..u1 (unused), b7=idx7, b6=idx6, ..., b0=idx0
-    __m256i const lo_lut_i8x32 = _mm256_setzero_si256();
-    __m256i const hi_lut_i8x32 = _mm256_broadcastsi128_si256(_mm_set_epi8( //
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,                    //
-        0x23, 0x22, 0x21, 0x20, 0x1E, 0x1C, 0x18, 0x00));                  //
-
-    // Extract mantissa (bits 0-2) as byte indices for shuffle
-    __m256i byte_idx_i8x32 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi8(0x07));
-    __m256i lo_bytes_i8x32 = _mm256_shuffle_epi8(lo_lut_i8x32, byte_idx_i8x32);
-    __m256i hi_bytes_i8x32 = _mm256_shuffle_epi8(hi_lut_i8x32, byte_idx_i8x32);
-
-    // Combine low and high bytes into 16-bit values
-    __m256i subnorm_abs_i16x16 = _mm256_or_si256(                    //
-        _mm256_and_si256(lo_bytes_i8x32, _mm256_set1_epi16(0x00FF)), //
-        _mm256_slli_epi16(hi_bytes_i8x32, 8));                       //
-    __m256i subnorm_i16x16 = _mm256_or_si256(subnorm_abs_i16x16, sign_i16x16);
-
-    // Blend: if exponent == 0, use subnormal result; else use normal result
-    __m256i exp_bits_i16x16 = _mm256_and_si256(e4m3_i16x16, _mm256_set1_epi16(0x78));
-    __m256i is_subnormal_i16x16 = _mm256_cmpeq_epi16(exp_bits_i16x16, _mm256_setzero_si256());
-    __m256i result_i16x16 = _mm256_blendv_epi8(normal_i16x16, subnorm_i16x16, is_subnormal_i16x16);
-
-    // Handle NaN: E4M3 index 127 (0x7F) → F16 NaN (0x7E00)
-    __m256i is_nan_i16x16 = _mm256_cmpeq_epi16(lower7_i16x16, _mm256_set1_epi16(0x7F));
-    __m256i nan_i16x16 = _mm256_or_si256(sign_i16x16, _mm256_set1_epi16(0x7E00));
-    return _mm256_blendv_epi8(result_i16x16, nan_i16x16, is_nan_i16x16);
-}
-
-/** @brief Convert 16x e5m2 → 16x f16 via simple bit shift (AVX2).
- *  E5M2 format: S EEEEE MM (bias=15). F16: S EEEEE MMMMMMMMMM (bias=15).
- *  Same exponent bias means F16 = (lower7 << 8) | (sign << 15).
- *  Handles all corner cases: zero, subnormals, normals, infinity, and NaN. */
-NK_INTERNAL __m256i nk_e5m2x16_to_f16x16_haswell_(__m128i e5m2x16) {
-    __m256i e5m2_i16x16 = _mm256_cvtepu8_epi16(e5m2x16);
-    __m256i sign_i16x16 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi16((short)0x80));
-    __m256i lower7_i16x16 = _mm256_and_si256(e5m2_i16x16, _mm256_set1_epi16(0x7F));
-
-    // F16 = (lower7 << 8) | (sign << 15)
-    // Works for all cases: subnormals, normals, infinity, and NaN
-    __m256i result_i16x16 = _mm256_slli_epi16(lower7_i16x16, 8);
-    sign_i16x16 = _mm256_slli_epi16(sign_i16x16, 8);
-    return _mm256_or_si256(result_i16x16, sign_i16x16);
-}
-
 /** @brief Convert 8x e4m3 → 8x f32 via bit manipulation (AVX2).
  *  E4M3 format: S EEEE MMM (bias=7). F32: sign<<31, (exp+120)<<23, mant<<20.
- *  Subnormals (exp=0): value = mantissa × 2⁽¹⁻⁷⁾ × 2⁻³ = mantissa ÷ 512. */
+ *  Subnormals (exp=0): looked up via vpermps from an 8-entry register LUT.
+ *  NaN detection uses a single comparison on the 7-bit magnitude (0x7F). */
 NK_INTERNAL __m256 nk_e4m3x8_to_f32x8_haswell_(__m128i e4m3_i8x8) {
     __m256i e4m3_i32x8 = _mm256_cvtepu8_epi32(e4m3_i8x8);
 
@@ -348,21 +191,26 @@ NK_INTERNAL __m256 nk_e4m3x8_to_f32x8_haswell_(__m128i e4m3_i8x8) {
     __m256i f32_mant_i32x8 = _mm256_slli_epi32(mant_i32x8, 20);
     __m256i normal_bits_i32x8 = _mm256_or_si256(f32_sign_i32x8, _mm256_or_si256(f32_exp_i32x8, f32_mant_i32x8));
 
-    // Subnormal path: value = mantissa / 512.0f, then apply sign
-    __m256 subnorm_abs_f32x8 = _mm256_mul_ps(_mm256_cvtepi32_ps(mant_i32x8), _mm256_set1_ps(1.0f / 512.0f));
-    __m256 subnorm_f32x8 = _mm256_or_ps(subnorm_abs_f32x8, _mm256_castsi256_ps(f32_sign_i32x8));
+    // Subnormal path: vpermps from 8-entry register LUT (3 cy latency, no memory access)
+    __m256 subnorm_lut_f32x8 = _mm256_setr_ps(0, 1.0f / 512, 2.0f / 512, 3.0f / 512, //
+                                              4.0f / 512, 5.0f / 512, 6.0f / 512, 7.0f / 512);
+    __m256i subnorm_bits_i32x8 = _mm256_or_si256( //
+        _mm256_castps_si256(_mm256_permutevar8x32_ps(subnorm_lut_f32x8, mant_i32x8)), f32_sign_i32x8);
 
-    // Blend: if exp==0, use subnormal result; otherwise use normal bits
+    // Bitwise select: if exp==0, use subnormal; otherwise use normal
     __m256i exp_zero_mask = _mm256_cmpeq_epi32(exp_i32x8, _mm256_setzero_si256());
-    __m256 result = _mm256_blendv_ps(_mm256_castsi256_ps(normal_bits_i32x8), subnorm_f32x8,
-                                     _mm256_castsi256_ps(exp_zero_mask));
+    __m256i result_i32x8 = _mm256_or_si256(                  //
+        _mm256_and_si256(exp_zero_mask, subnorm_bits_i32x8), //
+        _mm256_andnot_si256(exp_zero_mask, normal_bits_i32x8));
 
-    // NaN path: E4M3FN has NaN only when exp=15 AND mant=7 (0x7F or 0xFF)
-    __m256i is_nan_mask = _mm256_and_si256(                                            //
-        _mm256_cmpeq_epi32(exp_i32x8, _mm256_set1_epi32(15)),                          //
-        _mm256_cmpeq_epi32(mant_i32x8, _mm256_set1_epi32(7)));                         //
-    __m256i nan_bits = _mm256_or_si256(f32_sign_i32x8, _mm256_set1_epi32(0x7FC00000)); // F32 quiet NaN
-    return _mm256_blendv_ps(result, _mm256_castsi256_ps(nan_bits), _mm256_castsi256_ps(is_nan_mask));
+    // NaN: E4M3FN has NaN only at magnitude 0x7F (exp=15, mant=7)
+    __m256i lower7_i32x8 = _mm256_and_si256(e4m3_i32x8, _mm256_set1_epi32(0x7F));
+    __m256i is_nan_mask = _mm256_cmpeq_epi32(lower7_i32x8, _mm256_set1_epi32(0x7F));
+    __m256i nan_i32x8 = _mm256_or_si256(f32_sign_i32x8, _mm256_set1_epi32(0x7FC00000));
+    result_i32x8 = _mm256_or_si256(               //
+        _mm256_and_si256(is_nan_mask, nan_i32x8), //
+        _mm256_andnot_si256(is_nan_mask, result_i32x8));
+    return _mm256_castsi256_ps(result_i32x8);
 }
 
 /** @brief Convert 8x e5m2 → 8x f32 via bit manipulation (AVX2).
@@ -676,9 +524,9 @@ NK_INTERNAL __m128i nk_f32x8_to_e3m2x8_haswell_(__m256 f32x8) {
     return packed_i8x8;
 }
 
-#pragma endregion - Vectorized Conversions
+#pragma endregion Vectorized Conversions
 
-#pragma region - Converting Loads and Stores
+#pragma region Converting Loads and Stores
 
 /** @brief Full load for f16 elements (8) with conversion to f32 via F16C. */
 NK_INTERNAL void nk_load_f16x8_to_f32x8_haswell_(void const *src, nk_b256_vec_t *dst) {
@@ -794,9 +642,9 @@ NK_INTERNAL void nk_partial_load_u32x8_to_f32x8_haswell_(nk_u32_t const *src, nk
     dst->ymm_ps = nk_u32x8_to_f32x8_haswell_(vec.ymm);
 }
 
-#pragma endregion - Converting Loads and Stores
+#pragma endregion Converting Loads and Stores
 
-#pragma region - Public API
+#pragma region Public API
 
 NK_PUBLIC void nk_cast_haswell(void const *from, nk_dtype_t from_type, nk_size_t n, void *to, nk_dtype_t to_type) {
     // Same-type fast path
@@ -958,7 +806,7 @@ NK_PUBLIC void nk_cast_haswell(void const *from, nk_dtype_t from_type, nk_size_t
     }
 }
 
-#pragma endregion - Public API
+#pragma endregion Public API
 
 #if defined(__clang__)
 #pragma clang attribute pop