numkong 7.0.0 → 7.4.2

package/include/numkong/each/sapphire.h

@@ -8,18 +8,18 @@
  *
  *  @section sapphire_elementwise_instructions Relevant Instructions
  *
- *      Intrinsic                   Instruction                     Sapphire    Genoa
- *      _mm512_add_ph               VADDPH (ZMM, ZMM, ZMM)          4cy @ p05   3cy @ p01
- *      _mm512_mul_ph               VMULPH (ZMM, ZMM, ZMM)          4cy @ p05   3cy @ p01
- *      _mm512_fmadd_ph             VFMADD (ZMM, ZMM, ZMM)          4cy @ p05   4cy @ p01
- *      _mm512_cvtepi16_ph          VCVTW2PH (ZMM, ZMM)             4cy @ p05   4cy @ p01
- *      _mm512_cvtph_epi16          VCVTPH2W (ZMM, ZMM)             4cy @ p05   4cy @ p01
- *      _mm512_cvtepi8_epi16        VPMOVSXBW (ZMM, YMM)            3cy @ p5    3cy @ p12
- *      _mm512_cvtsepi16_epi8       VPMOVSWB (YMM, ZMM)             4cy @ p5    4cy @ p12
- *      _mm512_packus_epi16         VPACKUSWB (ZMM, ZMM, ZMM)       1cy @ p5    1cy @ p12
- *      _mm256_add_ph               VADDPH (YMM, YMM, YMM)          4cy @ p05   3cy @ p01
- *      _mm512_maskz_loadu_epi16    VMOVDQU16 (ZMM {K}, M512)       7cy @ p23   7cy @ p23
- *      _mm512_mask_storeu_epi16    VMOVDQU16 (M512 {K}, ZMM)       4cy @ p4    4cy @ p4
+ *      Intrinsic                 Instruction                Sapphire   Genoa
+ *      _mm512_add_ph             VADDPH (ZMM, ZMM, ZMM)     4cy @ p05  3cy @ p01
+ *      _mm512_mul_ph             VMULPH (ZMM, ZMM, ZMM)     4cy @ p05  3cy @ p01
+ *      _mm512_fmadd_ph           VFMADD (ZMM, ZMM, ZMM)     4cy @ p05  4cy @ p01
+ *      _mm512_cvtepi16_ph        VCVTW2PH (ZMM, ZMM)        4cy @ p05  4cy @ p01
+ *      _mm512_cvtph_epi16        VCVTPH2W (ZMM, ZMM)        4cy @ p05  4cy @ p01
+ *      _mm512_cvtepi8_epi16      VPMOVSXBW (ZMM, YMM)       3cy @ p5   3cy @ p12
+ *      _mm512_cvtsepi16_epi8     VPMOVSWB (YMM, ZMM)        4cy @ p5   4cy @ p12
+ *      _mm512_packus_epi16       VPACKUSWB (ZMM, ZMM, ZMM)  1cy @ p5   1cy @ p12
+ *      _mm256_add_ph             VADDPH (YMM, YMM, YMM)     4cy @ p05  3cy @ p01
+ *      _mm512_maskz_loadu_epi16  VMOVDQU16 (ZMM {K}, M512)  7cy @ p23  7cy @ p23
+ *      _mm512_mask_storeu_epi16  VMOVDQU16 (M512 {K}, ZMM)  4cy @ p4   4cy @ p4
  */
 #ifndef NK_EACH_SAPPHIRE_H
 #define NK_EACH_SAPPHIRE_H
@@ -54,8 +54,8 @@ nk_each_sum_f16_sapphire_cycle:
         n = 0;
     }
     else {
-        a_f16_vec = _mm512_loadu_ph(a);
-        b_f16_vec = _mm512_loadu_ph(b);
+        a_f16_vec = _mm512_castsi512_ph(_mm512_loadu_epi16(a));
+        b_f16_vec = _mm512_castsi512_ph(_mm512_loadu_epi16(b));
         a += 32, b += 32, n -= 32;
     }
     sum_f16_vec = _mm512_add_ph(a_f16_vec, b_f16_vec);
@@ -287,146 +287,11 @@ nk_each_blend_i8_sapphire_cycle:
     if (n) goto nk_each_blend_i8_sapphire_cycle;
 }
 
-NK_PUBLIC void nk_each_fma_i8_sapphire(                                //
-    nk_i8_t const *a, nk_i8_t const *b, nk_i8_t const *c, nk_size_t n, //
-    nk_f32_t const *alpha, nk_f32_t const *beta, nk_i8_t *result) {
-
-    short alpha_short, beta_short;
-    nk_f32_to_f16_sapphire(alpha, (nk_f16_t *)&alpha_short);
-    nk_f32_to_f16_sapphire(beta, (nk_f16_t *)&beta_short);
-    __mmask64 mask = 0xFFFFFFFFFFFFFFFF;
-    __m512h alpha_f16x32 = _mm512_castsi512_ph(_mm512_set1_epi16(alpha_short));
-    __m512h beta_f16x32 = _mm512_castsi512_ph(_mm512_set1_epi16(beta_short));
-    __m256i a_low_i8x32, a_high_i8x32, b_low_i8x32, b_high_i8x32, c_low_i8x32, c_high_i8x32;
-    __m512i result_i8x64;
-    __m512h a_low_f16x32, a_high_f16x32, b_low_f16x32, b_high_f16x32;
-    __m512h c_low_f16x32, c_high_f16x32, ab_low_f16x32, ab_high_f16x32;
-    __m512h ab_scaled_low_f16x32, ab_scaled_high_f16x32, result_low_f16x32, result_high_f16x32;
-    __m512i result_low_i16x32, result_high_i16x32;
-    __m512h min_f16x32 = _mm512_cvtepi16_ph(_mm512_set1_epi16(-128));
-    __m512h max_f16x32 = _mm512_cvtepi16_ph(_mm512_set1_epi16(127));
-
-nk_each_fma_i8_sapphire_cycle:
-    if (n < 64) {
-        // Tail: use masked 512-bit loads and extract (runs once)
-        mask = (__mmask64)_bzhi_u64(0xFFFFFFFFFFFFFFFFull, n);
-        __m512i a_i8x64 = _mm512_maskz_loadu_epi8(mask, a);
-        __m512i b_i8x64 = _mm512_maskz_loadu_epi8(mask, b);
-        __m512i c_i8x64 = _mm512_maskz_loadu_epi8(mask, c);
-        a_low_i8x32 = _mm512_castsi512_si256(a_i8x64);
-        a_high_i8x32 = _mm512_extracti64x4_epi64(a_i8x64, 1);
-        b_low_i8x32 = _mm512_castsi512_si256(b_i8x64);
-        b_high_i8x32 = _mm512_extracti64x4_epi64(b_i8x64, 1);
-        c_low_i8x32 = _mm512_castsi512_si256(c_i8x64);
-        c_high_i8x32 = _mm512_extracti64x4_epi64(c_i8x64, 1);
-        n = 0;
-    }
-    else {
-        // Hot path: 2×256-bit loads per vector to avoid VEXTRACTI64X4 (Port 5)
-        a_low_i8x32 = _mm256_loadu_epi8(a);
-        a_high_i8x32 = _mm256_loadu_epi8(a + 32);
-        b_low_i8x32 = _mm256_loadu_epi8(b);
-        b_high_i8x32 = _mm256_loadu_epi8(b + 32);
-        c_low_i8x32 = _mm256_loadu_epi8(c);
-        c_high_i8x32 = _mm256_loadu_epi8(c + 32);
-        a += 64, b += 64, c += 64, n -= 64;
-    }
-    // Upcast from 256-bit halves:
-    a_low_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(a_low_i8x32));
-    a_high_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(a_high_i8x32));
-    b_low_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(b_low_i8x32));
-    b_high_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(b_high_i8x32));
-    c_low_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(c_low_i8x32));
-    c_high_f16x32 = _mm512_cvtepi16_ph(_mm512_cvtepi8_epi16(c_high_i8x32));
-    // Multiply:
-    ab_low_f16x32 = _mm512_mul_ph(a_low_f16x32, b_low_f16x32);
-    ab_high_f16x32 = _mm512_mul_ph(a_high_f16x32, b_high_f16x32);
-    // Scale:
-    ab_scaled_low_f16x32 = _mm512_mul_ph(ab_low_f16x32, alpha_f16x32);
-    ab_scaled_high_f16x32 = _mm512_mul_ph(ab_high_f16x32, alpha_f16x32);
-    // Add:
-    result_low_f16x32 = _mm512_fmadd_ph(c_low_f16x32, beta_f16x32, ab_scaled_low_f16x32);
-    result_high_f16x32 = _mm512_fmadd_ph(c_high_f16x32, beta_f16x32, ab_scaled_high_f16x32);
-    // Clip the 16-bit result to 8-bit:
-    result_low_f16x32 = _mm512_max_ph(_mm512_min_ph(result_low_f16x32, max_f16x32), min_f16x32);
-    result_high_f16x32 = _mm512_max_ph(_mm512_min_ph(result_high_f16x32, max_f16x32), min_f16x32);
-    // Downcast:
-    result_low_i16x32 = _mm512_cvtph_epi16(result_low_f16x32);
-    result_high_i16x32 = _mm512_cvtph_epi16(result_high_f16x32);
-    // Merge back:
-    result_i8x64 = _mm512_inserti64x4(_mm512_castsi256_si512(_mm512_cvtsepi16_epi8(result_low_i16x32)),
-                                      _mm512_cvtsepi16_epi8(result_high_i16x32), 1);
-    _mm512_mask_storeu_epi8(result, mask, result_i8x64);
-    result += 64;
-    if (n) goto nk_each_fma_i8_sapphire_cycle;
-}
-
-NK_PUBLIC void nk_each_fma_u8_sapphire(                                //
-    nk_u8_t const *a, nk_u8_t const *b, nk_u8_t const *c, nk_size_t n, //
-    nk_f32_t const *alpha, nk_f32_t const *beta, nk_u8_t *result) {
-
-    short alpha_short, beta_short;
-    nk_f32_to_f16_sapphire(alpha, (nk_f16_t *)&alpha_short);
-    nk_f32_to_f16_sapphire(beta, (nk_f16_t *)&beta_short);
-    __mmask64 mask = 0xFFFFFFFFFFFFFFFF;
-    __m512h alpha_f16x32 = _mm512_castsi512_ph(_mm512_set1_epi16(alpha_short));
-    __m512h beta_f16x32 = _mm512_castsi512_ph(_mm512_set1_epi16(beta_short));
-    __m512i a_u8x64, b_u8x64, c_u8x64, result_u8x64;
-    __m512h a_low_f16x32, a_high_f16x32, b_low_f16x32, b_high_f16x32;
-    __m512h c_low_f16x32, c_high_f16x32, ab_low_f16x32, ab_high_f16x32;
-    __m512h ab_scaled_low_f16x32, ab_scaled_high_f16x32, result_low_f16x32, result_high_f16x32;
-    __m512i result_low_i16x32, result_high_i16x32;
-    __m512h min_f16x32 = _mm512_cvtepi16_ph(_mm512_set1_epi16(0));
-    __m512h max_f16x32 = _mm512_cvtepi16_ph(_mm512_set1_epi16(255));
-
-nk_each_fma_u8_sapphire_cycle:
-    if (n < 64) {
-        mask = (__mmask64)_bzhi_u64(0xFFFFFFFFFFFFFFFFull, n);
-        a_u8x64 = _mm512_maskz_loadu_epi8(mask, a);
-        b_u8x64 = _mm512_maskz_loadu_epi8(mask, b);
-        c_u8x64 = _mm512_maskz_loadu_epi8(mask, c);
-        n = 0;
-    }
-    else {
-        a_u8x64 = _mm512_loadu_epi8(a);
-        b_u8x64 = _mm512_loadu_epi8(b);
-        c_u8x64 = _mm512_loadu_epi8(c);
-        a += 64, b += 64, c += 64, n -= 64;
-    }
-    // Upcast:
-    a_low_f16x32 = _mm512_cvtepi16_ph(_mm512_unpacklo_epi8(a_u8x64, _mm512_setzero_si512()));
-    a_high_f16x32 = _mm512_cvtepi16_ph(_mm512_unpackhi_epi8(a_u8x64, _mm512_setzero_si512()));
-    b_low_f16x32 = _mm512_cvtepi16_ph(_mm512_unpacklo_epi8(b_u8x64, _mm512_setzero_si512()));
-    b_high_f16x32 = _mm512_cvtepi16_ph(_mm512_unpackhi_epi8(b_u8x64, _mm512_setzero_si512()));
-    c_low_f16x32 = _mm512_cvtepi16_ph(_mm512_unpacklo_epi8(c_u8x64, _mm512_setzero_si512()));
-    c_high_f16x32 = _mm512_cvtepi16_ph(_mm512_unpackhi_epi8(c_u8x64, _mm512_setzero_si512()));
-    // Multiply:
-    ab_low_f16x32 = _mm512_mul_ph(a_low_f16x32, b_low_f16x32);
-    ab_high_f16x32 = _mm512_mul_ph(a_high_f16x32, b_high_f16x32);
-    // Scale:
-    ab_scaled_low_f16x32 = _mm512_mul_ph(ab_low_f16x32, alpha_f16x32);
-    ab_scaled_high_f16x32 = _mm512_mul_ph(ab_high_f16x32, alpha_f16x32);
-    // Add:
-    result_low_f16x32 = _mm512_fmadd_ph(c_low_f16x32, beta_f16x32, ab_scaled_low_f16x32);
-    result_high_f16x32 = _mm512_fmadd_ph(c_high_f16x32, beta_f16x32, ab_scaled_high_f16x32);
-    // Clip the 16-bit result to 8-bit:
-    result_low_f16x32 = _mm512_max_ph(_mm512_min_ph(result_low_f16x32, max_f16x32), min_f16x32);
-    result_high_f16x32 = _mm512_max_ph(_mm512_min_ph(result_high_f16x32, max_f16x32), min_f16x32);
-    // Downcast:
-    result_low_i16x32 = _mm512_cvtph_epi16(result_low_f16x32);
-    result_high_i16x32 = _mm512_cvtph_epi16(result_high_f16x32);
-    // Merge back:
-    result_u8x64 = _mm512_packus_epi16(result_low_i16x32, result_high_i16x32);
-    _mm512_mask_storeu_epi8(result, mask, result_u8x64);
-    result += 64;
-    if (n) goto nk_each_fma_u8_sapphire_cycle;
-}
-
 NK_PUBLIC void nk_each_sum_e4m3_sapphire(nk_e4m3_t const *a, nk_e4m3_t const *b, nk_size_t n, nk_e4m3_t *result) {
     __m256i a_e4m3x32, b_e4m3x32;
-    __m256h a_lo_f16x16, a_hi_f16x16, b_lo_f16x16, b_hi_f16x16;
-    __m256h sum_lo_f16x16, sum_hi_f16x16;
-    __m128i result_lo_e4m3x16, result_hi_e4m3x16;
+    __m256h a_low_f16x16, a_high_f16x16, b_low_f16x16, b_high_f16x16;
+    __m256h sum_low_f16x16, sum_high_f16x16;
+    __m128i result_low_e4m3x16, result_high_e4m3x16;
     __mmask32 mask = 0xFFFFFFFF;
 nk_each_sum_e4m3_sapphire_cycle:
     if (n < 32) {
@@ -442,21 +307,22 @@ nk_each_sum_e4m3_sapphire_cycle:
     }
 
     // Convert e4m3x16 → f16x16 (two halves)
-    a_lo_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_castsi256_si128(a_e4m3x32));
-    a_hi_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_extracti128_si256(a_e4m3x32, 1));
-    b_lo_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_castsi256_si128(b_e4m3x32));
-    b_hi_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_extracti128_si256(b_e4m3x32, 1));
+    a_low_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_castsi256_si128(a_e4m3x32));
+    a_high_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_extracti128_si256(a_e4m3x32, 1));
+    b_low_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_castsi256_si128(b_e4m3x32));
+    b_high_f16x16 = nk_e4m3x16_to_f16x16_sapphire_(_mm256_extracti128_si256(b_e4m3x32, 1));
 
     // Add in F16 - e4m3 sum is safe (max 896 < 65504)
-    sum_lo_f16x16 = _mm256_add_ph(a_lo_f16x16, b_lo_f16x16);
-    sum_hi_f16x16 = _mm256_add_ph(a_hi_f16x16, b_hi_f16x16);
+    sum_low_f16x16 = _mm256_add_ph(a_low_f16x16, b_low_f16x16);
+    sum_high_f16x16 = _mm256_add_ph(a_high_f16x16, b_high_f16x16);
 
     // Convert f16x16 → e4m3x16
-    result_lo_e4m3x16 = nk_f16x16_to_e4m3x16_sapphire_(sum_lo_f16x16);
-    result_hi_e4m3x16 = nk_f16x16_to_e4m3x16_sapphire_(sum_hi_f16x16);
+    result_low_e4m3x16 = nk_f16x16_to_e4m3x16_sapphire_(sum_low_f16x16);
+    result_high_e4m3x16 = nk_f16x16_to_e4m3x16_sapphire_(sum_high_f16x16);
 
     // Pack and store
-    __m256i result_e4m3x32 = _mm256_inserti128_si256(_mm256_castsi128_si256(result_lo_e4m3x16), result_hi_e4m3x16, 1);
+    __m256i result_e4m3x32 = _mm256_inserti128_si256(_mm256_castsi128_si256(result_low_e4m3x16), result_high_e4m3x16,
+                                                     1);
     _mm256_mask_storeu_epi8(result, mask, result_e4m3x32);
     result += 32;
     if (n) goto nk_each_sum_e4m3_sapphire_cycle;

package/include/numkong/each/serial.h

@@ -107,8 +107,8 @@ nk_define_each_scale_(i16, f32, nk_assign_from_to_, nk_f32_to_i16_serial)      /
 nk_define_each_scale_(u16, f32, nk_assign_from_to_, nk_f32_to_u16_serial)      // nk_each_scale_u16_serial
 nk_define_each_scale_(i32, f64, nk_assign_from_to_, nk_f64_to_i32_serial)      // nk_each_scale_i32_serial
 nk_define_each_scale_(u32, f64, nk_assign_from_to_, nk_f64_to_u32_serial)      // nk_each_scale_u32_serial
-nk_define_each_scale_(i64, f64, nk_assign_from_to_, nk_f64_to_i64_serial)      // nk_each_scale_i64_serial
-nk_define_each_scale_(u64, f64, nk_assign_from_to_, nk_f64_to_u64_serial)      // nk_each_scale_u64_serial
+nk_define_each_scale_(i64, f64, nk_f64_from_i64_, nk_f64_to_i64_serial)        // nk_each_scale_i64_serial
+nk_define_each_scale_(u64, f64, nk_f64_from_u64_, nk_f64_to_u64_serial)        // nk_each_scale_u64_serial
 
 nk_define_each_blend_(f64, f64, nk_assign_from_to_, nk_assign_from_to_)        // nk_each_blend_f64_serial
 nk_define_each_blend_(f32, f32, nk_assign_from_to_, nk_assign_from_to_)        // nk_each_blend_f32_serial
@@ -124,8 +124,8 @@ nk_define_each_blend_(i16, f32, nk_assign_from_to_, nk_f32_to_i16_serial)      /
 nk_define_each_blend_(u16, f32, nk_assign_from_to_, nk_f32_to_u16_serial)      // nk_each_blend_u16_serial
 nk_define_each_blend_(i32, f64, nk_assign_from_to_, nk_f64_to_i32_serial)      // nk_each_blend_i32_serial
 nk_define_each_blend_(u32, f64, nk_assign_from_to_, nk_f64_to_u32_serial)      // nk_each_blend_u32_serial
-nk_define_each_blend_(i64, f64, nk_assign_from_to_, nk_f64_to_i64_serial)      // nk_each_blend_i64_serial
-nk_define_each_blend_(u64, f64, nk_assign_from_to_, nk_f64_to_u64_serial)      // nk_each_blend_u64_serial
+nk_define_each_blend_(i64, f64, nk_f64_from_i64_, nk_f64_to_i64_serial)        // nk_each_blend_i64_serial
+nk_define_each_blend_(u64, f64, nk_f64_from_u64_, nk_f64_to_u64_serial)        // nk_each_blend_u64_serial
 
 nk_define_each_fma_(f64, f64, nk_assign_from_to_, nk_assign_from_to_)        // nk_each_fma_f64_serial
 nk_define_each_fma_(f32, f32, nk_assign_from_to_, nk_assign_from_to_)        // nk_each_fma_f32_serial
@@ -141,8 +141,8 @@ nk_define_each_fma_(i16, f32, nk_assign_from_to_, nk_f32_to_i16_serial)      //
 nk_define_each_fma_(u16, f32, nk_assign_from_to_, nk_f32_to_u16_serial)      // nk_each_fma_u16_serial
 nk_define_each_fma_(i32, f64, nk_assign_from_to_, nk_f64_to_i32_serial)      // nk_each_fma_i32_serial
 nk_define_each_fma_(u32, f64, nk_assign_from_to_, nk_f64_to_u32_serial)      // nk_each_fma_u32_serial
-nk_define_each_fma_(i64, f64, nk_assign_from_to_, nk_f64_to_i64_serial)      // nk_each_fma_i64_serial
-nk_define_each_fma_(u64, f64, nk_assign_from_to_, nk_f64_to_u64_serial)      // nk_each_fma_u64_serial
+nk_define_each_fma_(i64, f64, nk_f64_from_i64_, nk_f64_to_i64_serial)        // nk_each_fma_i64_serial
+nk_define_each_fma_(u64, f64, nk_f64_from_u64_, nk_f64_to_u64_serial)        // nk_each_fma_u64_serial
 
 #undef nk_define_each_scale_
 #undef nk_define_each_sum_

package/include/numkong/each/skylake.h

@@ -8,13 +8,13 @@
  *
  *  @section skylake_elementwise_instructions Relevant Instructions
  *
- *      Intrinsic                   Instruction                     SKL         ICL         Genoa
- *      _mm512_add_ps               VADDPS (ZMM, ZMM, ZMM)          4cy @ p05   4cy @ p0    3cy @ p01
- *      _mm512_fmadd_ps             VFMADD132PS (ZMM, ZMM, ZMM)     4cy @ p05   4cy @ p0    4cy @ p01
- *      _mm512_mul_ps               VMULPS (ZMM, ZMM, ZMM)          4cy @ p05   4cy @ p0    3cy @ p01
- *      _mm512_cvtph_ps             VCVTPH2PS (ZMM, YMM)            5cy @ p05   7cy @ p0    5cy @ p01
- *      _mm512_maskz_loadu_ps       VMOVUPS (ZMM {K}, M512)         7cy @ p23   7cy @ p23   7cy @ p23
- *      _mm512_mask_storeu_ps       VMOVUPS (M512 {K}, ZMM)         4cy @ p4    4cy @ p4    4cy @ p4
+ *      Intrinsic              Instruction                  SKL        ICL        Genoa
+ *      _mm512_add_ps          VADDPS (ZMM, ZMM, ZMM)       4cy @ p05  4cy @ p0   3cy @ p01
+ *      _mm512_fmadd_ps        VFMADD132PS (ZMM, ZMM, ZMM)  4cy @ p05  4cy @ p0   4cy @ p01
+ *      _mm512_mul_ps          VMULPS (ZMM, ZMM, ZMM)       4cy @ p05  4cy @ p0   3cy @ p01
+ *      _mm512_cvtph_ps        VCVTPH2PS (ZMM, YMM)         5cy @ p05  7cy @ p0   5cy @ p01
+ *      _mm512_maskz_loadu_ps  VMOVUPS (ZMM {K}, M512)      7cy @ p23  7cy @ p23  7cy @ p23
+ *      _mm512_mask_storeu_ps  VMOVUPS (M512 {K}, ZMM)      4cy @ p4   4cy @ p4   4cy @ p4
  *
  *  Skylake-X server chips have dual 512-bit FMA units enabling 0.5cy throughput for arithmetic operations.
  *  AVX-512 masked loads and stores eliminate branch misprediction penalties for partial vector processing.