numkong 7.2.2__tar.gz → 7.2.4__tar.gz

{numkong-7.2.2/numkong.egg-info → numkong-7.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: numkong
-Version: 7.2.2
+Version: 7.2.4
 Summary: Portable mixed-precision BLAS-like vector math library for x86 and ARM
 Home-page: https://github.com/ashvardanian/NumKong
 Author: Ash Vardanian

numkong-7.2.4/VERSION

@@ -0,0 +1 @@
+7.2.4

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/capabilities.h

@@ -96,7 +96,7 @@
 
 #define NK_VERSION_MAJOR 7
 #define NK_VERSION_MINOR 2
-#define NK_VERSION_PATCH 2
+#define NK_VERSION_PATCH 4
 
 /**
  *  @brief  Removes compile-time dispatching, and replaces it with runtime dispatching.

{numkong-7.2.2 → numkong-7.2.4}/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_

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/mesh/serial.h

@@ -436,7 +436,7 @@ nk_define_det3x3_(f64)
             rotation[3] = 0, rotation[4] = 1, rotation[5] = 0;                                                     \
             rotation[6] = 0, rotation[7] = 0, rotation[8] = 1;                                                     \
         }                                                                                                          \
-        if (scale) *scale = 1.0;                                                                                   \
+        if (scale) *scale = (nk_##output_type##_t)1;                                                               \
         nk_##accumulator_type##_t sum_squared = 0, sum_squared_compensation = 0;                                   \
         for (nk_size_t i = 0; i < n; ++i) {                                                                        \
             load_and_convert(a + i * 3 + 0, &val_a_x), load_and_convert(b + i * 3 + 0, &val_b_x);                  \
@@ -534,7 +534,7 @@ nk_define_det3x3_(f64)
         if (rotation) {                                                                                              \
             for (int j = 0; j < 9; ++j) rotation[j] = (nk_##output_type##_t)rotation_matrix[j];                      \
         }                                                                                                            \
-        if (scale) *scale = 1.0;                                                                                     \
+        if (scale) *scale = (nk_##output_type##_t)1;                                                                 \
         /* Step 5: Compute RMSD after rotation */                                                                    \
         nk_##accumulator_type##_t sum_squared = 0, sum_squared_compensation = 0;                                     \
         for (nk_size_t i = 0; i < n; ++i) {                                                                          \
@@ -652,7 +652,7 @@ nk_define_det3x3_(f64)
         nk_##svd_type##_t trace_scaled_s = svd_s[0] + svd_s[4] + sign_det * svd_s[8];                                 \
         nk_##accumulator_type##_t scale_factor = (nk_##accumulator_type##_t)trace_scaled_s /                          \
                                                  ((nk_##accumulator_type##_t)n * variance_a);                         \
-        if (scale) *scale = scale_factor;                                                                             \
+        if (scale) *scale = (nk_##output_type##_t)scale_factor;                                                       \
         if (rotation_det < 0) {                                                                                       \
             svd_v[2] = -svd_v[2], svd_v[5] = -svd_v[5], svd_v[8] = -svd_v[8];                                         \
             nk_rotation_from_svd_##svd_type##_serial_(svd_u, svd_v, rotation_matrix);                                 \

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/scalar/haswell.h

@@ -89,8 +89,8 @@ NK_PUBLIC nk_u64_t nk_u64_saturating_mul_haswell(nk_u64_t a, nk_u64_t b) {
 }
 NK_PUBLIC nk_i64_t nk_i64_saturating_mul_haswell(nk_i64_t a, nk_i64_t b) {
     int sign = (a < 0) ^ (b < 0);
-    nk_u64_t abs_a = a < 0 ? -(nk_u64_t)a : (nk_u64_t)a;
-    nk_u64_t abs_b = b < 0 ? -(nk_u64_t)b : (nk_u64_t)b;
+    nk_u64_t abs_a = a < 0 ? (0u - (nk_u64_t)a) : (nk_u64_t)a;
+    nk_u64_t abs_b = b < 0 ? (0u - (nk_u64_t)b) : (nk_u64_t)b;
     unsigned long long high;
     unsigned long long low = _mulx_u64(abs_a, abs_b, &high);
     if (high || (sign && low > 9223372036854775808ull) || (!sign && low > 9223372036854775807ull))

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/scalar/neon.h

@@ -98,8 +98,8 @@ NK_PUBLIC nk_u64_t nk_u64_saturating_mul_neon(nk_u64_t a, nk_u64_t b) {
 }
 NK_PUBLIC nk_i64_t nk_i64_saturating_mul_neon(nk_i64_t a, nk_i64_t b) {
     int sign = (a < 0) ^ (b < 0);
-    nk_u64_t abs_a = a < 0 ? -(nk_u64_t)a : (nk_u64_t)a;
-    nk_u64_t abs_b = b < 0 ? -(nk_u64_t)b : (nk_u64_t)b;
+    nk_u64_t abs_a = a < 0 ? (0u - (nk_u64_t)a) : (nk_u64_t)a;
+    nk_u64_t abs_b = b < 0 ? (0u - (nk_u64_t)b) : (nk_u64_t)b;
     nk_u64_t high = nk_u64_mulhigh_neon_(abs_a, abs_b);
     nk_u64_t low = abs_a * abs_b;
     if (high || (sign && low > 9223372036854775808ull) || (!sign && low > 9223372036854775807ull))

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/scalar/rvv.h

@@ -185,8 +185,8 @@ NK_PUBLIC nk_u64_t nk_u64_saturating_mul_rvv(nk_u64_t a, nk_u64_t b) {
 }
 NK_PUBLIC nk_i64_t nk_i64_saturating_mul_rvv(nk_i64_t a, nk_i64_t b) {
     int sign = (a < 0) ^ (b < 0);
-    nk_u64_t abs_a = a < 0 ? -(nk_u64_t)a : (nk_u64_t)a;
-    nk_u64_t abs_b = b < 0 ? -(nk_u64_t)b : (nk_u64_t)b;
+    nk_u64_t abs_a = a < 0 ? (0u - (nk_u64_t)a) : (nk_u64_t)a;
+    nk_u64_t abs_b = b < 0 ? (0u - (nk_u64_t)b) : (nk_u64_t)b;
     vuint64m1_t a_u64m1 = __riscv_vmv_v_x_u64m1(abs_a, 1);
     vuint64m1_t b_u64m1 = __riscv_vmv_v_x_u64m1(abs_b, 1);
     nk_u64_t high = __riscv_vmv_x_s_u64m1_u64(__riscv_vmulhu_vv_u64m1(a_u64m1, b_u64m1, 1));

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/scalar/serial.h

@@ -238,8 +238,8 @@ NK_PUBLIC nk_i64_t nk_i64_saturating_mul_serial(nk_i64_t a, nk_i64_t b) {
     int sign = ((a < 0) ^ (b < 0)) ? -1 : 1; // Track the sign of the result
 
     // Take absolute values for easy multiplication and overflow detection
-    nk_u64_t abs_a = (a < 0) ? -(nk_u64_t)a : (nk_u64_t)a;
-    nk_u64_t abs_b = (b < 0) ? -(nk_u64_t)b : (nk_u64_t)b;
+    nk_u64_t abs_a = (a < 0) ? (0u - (nk_u64_t)a) : (nk_u64_t)a;
+    nk_u64_t abs_b = (b < 0) ? (0u - (nk_u64_t)b) : (nk_u64_t)b;
 
     // Split the absolute values into high and low 32-bit parts
     nk_u64_t a_high = abs_a >> 32;

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/spatial/serial.h

@@ -101,8 +101,8 @@ extern "C" {
         if (a_norm_sq == 0 && b_norm_sq == 0) { *result = 0; }                                                    \
         else if (dot_product == 0) { *result = 1; }                                                               \
         else {                                                                                                    \
-            nk_##output_type##_t unclipped_distance = 1 - dot_product * compute_rsqrt(a_norm_sq) *                \
-                                                              compute_rsqrt(b_norm_sq);                           \
+            nk_##output_type##_t unclipped_distance = (nk_##output_type##_t)(                                     \
+                1 - dot_product * compute_rsqrt(a_norm_sq) * compute_rsqrt(b_norm_sq));                           \
             *result = unclipped_distance > 0 ? unclipped_distance : 0;                                            \
         }                                                                                                         \
     }

{numkong-7.2.2 → numkong-7.2.4}/include/numkong/types.h

@@ -1541,6 +1541,12 @@ NK_INTERNAL nk_bf16_t nk_bf16_from_u16_(nk_u16_t bits) {
     return c.bf;
 }
 
+NK_INTERNAL void nk_f64_from_i64_(nk_i64_t const *src, nk_f64_t *dest) { *dest = (nk_f64_t)*src; }
+NK_INTERNAL void nk_f64_from_u64_(nk_u64_t const *src, nk_f64_t *dest) { *dest = (nk_f64_t)*src; }
+NK_INTERNAL void nk_f32_from_i32_(nk_i32_t const *src, nk_f32_t *dest) { *dest = (nk_f32_t)*src; }
+NK_INTERNAL void nk_f32_from_u32_(nk_u32_t const *src, nk_f32_t *dest) { *dest = (nk_f32_t)*src; }
+NK_INTERNAL void nk_f32_from_f64_(nk_f64_t const *src, nk_f32_t *dest) { *dest = (nk_f32_t)*src; }
+
 /** @brief E4M3: NaN when (raw & 0x7F) == 0x7F  (two NaN values: 0x7F, 0xFF). */
 NK_INTERNAL int nk_e4m3_is_nan_(nk_e4m3_t x) { return (x & 0x7F) == 0x7F; }
 

{numkong-7.2.2 → numkong-7.2.4/numkong.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: numkong
-Version: 7.2.2
+Version: 7.2.4
 Summary: Portable mixed-precision BLAS-like vector math library for x86 and ARM
 Home-page: https://github.com/ashvardanian/NumKong
 Author: Ash Vardanian

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neon.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON (AArch64 baseline SIMD) */
-#if !defined(__ARM_NEON)
-#error "NEON not available"
-#endif
 #include <arm_neon.h>
 int main(void) {
     float32x4_t a = vdupq_n_f32(1.0f);

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neon_bfdot.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON BF16 (ARMv8.6-A bfloat16 dot product) */
-#if !defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) && !(defined(__ARM_ARCH) && __ARM_ARCH >= 806)
-#error "NEON BF16 not available"
-#endif
 #include <arm_neon.h>
 int main(void) {
     bfloat16x8_t a = vdupq_n_bf16(1.0f);

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neon_fhm.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON FHM (ARMv8.2-A FP16 fused multiply-add) */
-#if !defined(__ARM_FEATURE_FP16_FML) && !(defined(__ARM_ARCH) && __ARM_ARCH >= 804)
-#error "NEON FHM not available"
-#endif
 #include <arm_neon.h>
 int main(void) {
     float16x8_t a = vdupq_n_f16(1.0f);

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neon_half.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON F16 (ARMv8.2-A half-precision) */
-#if !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !(defined(__ARM_ARCH) && __ARM_ARCH >= 802)
-#error "NEON F16 not available"
-#endif
 #include <arm_neon.h>
 int main(void) {
     float16x8_t a = vdupq_n_f16(1.0f);

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neon_sdot.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON SDOT (ARMv8.2-A dot product) */
-#if !defined(__ARM_FEATURE_DOTPROD) && !(defined(__ARM_ARCH) && __ARM_ARCH >= 804)
-#error "NEON SDOT not available"
-#endif
 #include <arm_neon.h>
 int main(void) {
     int8x16_t a = vdupq_n_s8(1);

{numkong-7.2.2 → numkong-7.2.4}/probes/arm_neonfp8.c

@@ -1,7 +1,4 @@
 /* NumKong ISA probe: NEON FP8 (fp8dot4) */
-#if !defined(__ARM_NEON)
-#error "Feature not available"
-#endif
 #include <arm_neon.h>
 int test_neonfp8(void) {
     mfloat8x16_t a = vreinterpretq_mf8_u8(vdupq_n_u8(0));

numkong-7.2.2/VERSION

@@ -1 +0,0 @@
-7.2.2