Fixed _mm512_kunpackb, reduce-max and reduce-min

`_mm512_kunpackb` was implemented wrong, and `simd_reduce_max` uses `maxnum` for comparison, which adheres to IEEE754, but Intel specifically says that they do NOT adhere to IEEE754 for NaNs, which can give wrong results

Author: sayantn

crates/core_arch/src/x86/avx512f.rs

@@ -27663,9 +27663,7 @@ pub unsafe fn _mm512_mask2int(k1: __mmask16) -> i32 {
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 #[cfg_attr(test, assert_instr(mov))] // generate normal and code instead of kunpckbw
 pub unsafe fn _mm512_kunpackb(a: __mmask16, b: __mmask16) -> __mmask16 {
-    let a = a & 0b00000000_11111111;
-    let b = b & 0b11111111_00000000;
-    a | b
+    ((a & 0xff) << 8) | (b & 0xff)
 }
 
 /// Performs bitwise OR between k1 and k2, storing the result in dst. CF flag is set if dst consists of all 1's.
@@ -31554,7 +31552,13 @@ pub unsafe fn _mm512_mask_reduce_max_epu64(k: __mmask8, a: __m512i) -> u64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_reduce_max_ps(a: __m512) -> f32 {
-    simd_reduce_max(a.as_f32x16())
+    let a = _mm256_max_ps(
+        simd_shuffle!(a, a, [0, 1, 2, 3, 4, 5, 6, 7]),
+        simd_shuffle!(a, a, [8, 9, 10, 11, 12, 13, 14, 15]),
+    );
+    let a = _mm_max_ps(_mm256_extractf128_ps::<0>(a), _mm256_extractf128_ps::<1>(a));
+    let a = _mm_max_ps(a, simd_shuffle!(a, a, [2, 3, 0, 1]));
+    _mm_cvtss_f32(_mm_max_ss(a, _mm_movehdup_ps(a)))
 }
 
 /// Reduce the packed single-precision (32-bit) floating-point elements in a by maximum using mask k. Returns the maximum of all active elements in a.
@@ -31564,11 +31568,7 @@ pub unsafe fn _mm512_reduce_max_ps(a: __m512) -> f32 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_max_ps(k: __mmask16, a: __m512) -> f32 {
-    simd_reduce_max(simd_select_bitmask(
-        k,
-        a.as_f32x16(),
-        _mm512_undefined_ps().as_f32x16(),
-    ))
+    _mm512_reduce_max_ps(_mm512_mask_mov_ps(_mm512_set1_ps(f32::MIN), k, a))
 }
 
 /// Reduce the packed double-precision (64-bit) floating-point elements in a by maximum. Returns the maximum of all elements in a.
@@ -31578,7 +31578,12 @@ pub unsafe fn _mm512_mask_reduce_max_ps(k: __mmask16, a: __m512) -> f32 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_reduce_max_pd(a: __m512d) -> f64 {
-    simd_reduce_max(a.as_f64x8())
+    let a = _mm256_max_pd(
+        _mm512_extractf64x4_pd::<0>(a),
+        _mm512_extractf64x4_pd::<1>(a),
+    );
+    let a = _mm_max_pd(_mm256_extractf128_pd::<0>(a), _mm256_extractf128_pd::<1>(a));
+    _mm_cvtsd_f64(_mm_max_sd(a, simd_shuffle!(a, a, [1, 0])))
 }
 
 /// Reduce the packed double-precision (64-bit) floating-point elements in a by maximum using mask k. Returns the maximum of all active elements in a.
@@ -31588,11 +31593,7 @@ pub unsafe fn _mm512_reduce_max_pd(a: __m512d) -> f64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_max_pd(k: __mmask8, a: __m512d) -> f64 {
-    simd_reduce_max(simd_select_bitmask(
-        k,
-        a.as_f64x8(),
-        _mm512_undefined_pd().as_f64x8(),
-    ))
+    _mm512_reduce_max_pd(_mm512_mask_mov_pd(_mm512_set1_pd(f64::MIN), k, a))
 }
 
 /// Reduce the packed signed 32-bit integers in a by minimum. Returns the minimum of all elements in a.
@@ -31698,7 +31699,13 @@ pub unsafe fn _mm512_mask_reduce_min_epu64(k: __mmask8, a: __m512i) -> u64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_reduce_min_ps(a: __m512) -> f32 {
-    simd_reduce_min(a.as_f32x16())
+    let a = _mm256_min_ps(
+        simd_shuffle!(a, a, [0, 1, 2, 3, 4, 5, 6, 7]),
+        simd_shuffle!(a, a, [8, 9, 10, 11, 12, 13, 14, 15]),
+    );
+    let a = _mm_min_ps(_mm256_extractf128_ps::<0>(a), _mm256_extractf128_ps::<1>(a));
+    let a = _mm_min_ps(a, simd_shuffle!(a, a, [2, 3, 0, 1]));
+    _mm_cvtss_f32(_mm_min_ss(a, _mm_movehdup_ps(a)))
 }
 
 /// Reduce the packed single-precision (32-bit) floating-point elements in a by maximum using mask k. Returns the minimum of all active elements in a.
@@ -31708,11 +31715,7 @@ pub unsafe fn _mm512_reduce_min_ps(a: __m512) -> f32 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_min_ps(k: __mmask16, a: __m512) -> f32 {
-    simd_reduce_min(simd_select_bitmask(
-        k,
-        a.as_f32x16(),
-        _mm512_undefined_ps().as_f32x16(),
-    ))
+    _mm512_reduce_min_ps(_mm512_mask_mov_ps(_mm512_set1_ps(f32::MAX), k, a))
 }
 
 /// Reduce the packed double-precision (64-bit) floating-point elements in a by minimum. Returns the minimum of all elements in a.
@@ -31722,7 +31725,12 @@ pub unsafe fn _mm512_mask_reduce_min_ps(k: __mmask16, a: __m512) -> f32 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_reduce_min_pd(a: __m512d) -> f64 {
-    simd_reduce_min(a.as_f64x8())
+    let a = _mm256_min_pd(
+        _mm512_extractf64x4_pd::<0>(a),
+        _mm512_extractf64x4_pd::<1>(a),
+    );
+    let a = _mm_min_pd(_mm256_extractf128_pd::<0>(a), _mm256_extractf128_pd::<1>(a));
+    _mm_cvtsd_f64(_mm_min_sd(a, simd_shuffle!(a, a, [1, 0])))
 }
 
 /// Reduce the packed double-precision (64-bit) floating-point elements in a by maximum using mask k. Returns the minimum of all active elements in a.
@@ -31732,11 +31740,7 @@ pub unsafe fn _mm512_reduce_min_pd(a: __m512d) -> f64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_min_pd(k: __mmask8, a: __m512d) -> f64 {
-    simd_reduce_min(simd_select_bitmask(
-        k,
-        a.as_f64x8(),
-        _mm512_undefined_pd().as_f64x8(),
-    ))
+    _mm512_reduce_min_pd(_mm512_mask_mov_pd(_mm512_set1_pd(f64::MAX), k, a))
 }
 
 /// Reduce the packed 32-bit integers in a by bitwise AND. Returns the bitwise AND of all elements in a.
@@ -54323,7 +54327,7 @@ mod tests {
         let a: u16 = 0b11001100_00110011;
         let b: u16 = 0b00101110_00001011;
         let r = _mm512_kunpackb(a, b);
-        let e: u16 = 0b00101110_00110011;
+        let e: u16 = 0b00110011_00001011;
         assert_eq!(r, e);
     }