From f99d4e252fed9d5ff7bc8460efe85e822b7e56ef Mon Sep 17 00:00:00 2001
From: Guillaume Piolat <contact@auburnsounds.com>
Date: Tue, 3 Oct 2023 16:57:36 +0200
Subject: [PATCH] Add support for _mm256_max_epu32

---
 source/inteli/avx2intrin.d | 47 ++++++++++++++++++++++++++++++++++++--
 source/inteli/smmintrin.d  | 17 +++++++++++---
 2 files changed, 59 insertions(+), 5 deletions(-)

diff --git a/source/inteli/avx2intrin.d b/source/inteli/avx2intrin.d
index 6464ca1..e748407 100644
--- a/source/inteli/avx2intrin.d
+++ b/source/inteli/avx2intrin.d
@@ -1475,10 +1475,53 @@ unittest
 }
 
 // TODO __m256i _mm256_max_epi8 (__m256i a, __m256i b) pure @safe
+// TODO __m256i _mm256_max_epu16 (__m256i a, __m256i b) pure @safe
 
+/// Compare packed unsigned 32-bit integers in `a` and `b`, and return packed maximum values.
+__m256i _mm256_max_epu32 (__m256i a, __m256i b) pure @safe
+{
+    // PERF D_SIMD
+    version(GNU)
+        enum bool split = true;
+    else static if (SIMD_COMPARISON_MASKS_32B)
+        enum bool split = false;
+    else
+        enum bool split = true;
+
+    static if (GDC_with_AVX2)
+    {
+        return cast(__m256i) __builtin_ia32_pmaxud256(cast(int8)a, cast(int8)b);
+    }
+    else static if (split)
+    {
+        // split
+        __m128i a_lo = _mm256_extractf128_si256!0(a);
+        __m128i a_hi = _mm256_extractf128_si256!1(a);
+        __m128i b_lo = _mm256_extractf128_si256!0(b);
+        __m128i b_hi = _mm256_extractf128_si256!1(b);
+        __m128i r_lo = _mm_max_epu32(a_lo, b_lo);
+        __m128i r_hi = _mm_max_epu32(a_hi, b_hi);
+        return _mm256_set_m128i(r_hi, r_lo);
+    }
+    else static if (SIMD_COMPARISON_MASKS_32B) 
+    {
+        // catastrophic with GDC x86 for some reason, like for 16-bit numbers.
+        uint8 sa = cast(uint8)a;
+        uint8 sb = cast(uint8)b;
+        uint8 greater = sa > sb;
+        return cast(__m256i)( (greater & sa) | (~greater & sb) );
+    }
+    else
+        static assert(0);
+}
+unittest
+{
+    int8 R = cast(int8) _mm256_max_epu32(_mm256_setr_epi32(0x7fffffff, 1,  4, -7, 0x7fffffff, 1, 11, -7),
+                                         _mm256_setr_epi32(        -4,-8,  9, -8,         -4,-8,  9, -8));
+    int[8] correct =                                      [        -4,-8,  9, -7,         -4,-8, 11, -7];
+    assert(R.array == correct);
+}
 
-// TODO __m256i _mm256_max_epu16 (__m256i a, __m256i b) pure @safe
-// TODO __m256i _mm256_max_epu32 (__m256i a, __m256i b) pure @safe
 // TODO __m256i _mm256_max_epu8 (__m256i a, __m256i b) pure @safe
 
 // Compare packed signed 16-bit integers in `a` and `b`, and return packed minimum values.
diff --git a/source/inteli/smmintrin.d b/source/inteli/smmintrin.d
index bef80e9..651de32 100644
--- a/source/inteli/smmintrin.d
+++ b/source/inteli/smmintrin.d
@@ -1302,7 +1302,7 @@ unittest
 }
 
 /// Compare packed unsigned 32-bit integers in `a` and `b`, returns packed maximum values.
-__m128i _mm_max_epu32 (__m128i a, __m128i b) @trusted
+__m128i _mm_max_epu32 (__m128i a, __m128i b) pure @trusted
 {
     // PERF DMD
     static if (GDC_with_SSE41)
@@ -1323,6 +1323,17 @@ __m128i _mm_max_epu32 (__m128i a, __m128i b) @trusted
     }
     else
     {
+        // PERF: LLVM suggests to replace the _mm_add_epi32 by _mm_xor_si128, and the last xor by an "_mm_or_si128"
+        /+
+        movdqa  xmm2, xmmword ptr [-0x80000000, -0x80000000, -0x80000000, -0x80000000]
+        movdqa  xmm3, xmm1
+        pxor    xmm3, xmm2
+        pxor    xmm2, xmm0
+        pcmpgtd xmm2, xmm3
+        pand    xmm0, xmm2
+        pandn   xmm2, xmm1
+        por     xmm0, xmm2
+        +/
         __m128i valueShift = _mm_set1_epi32(-0x80000000);
         __m128i higher = _mm_cmpgt_epi32(_mm_add_epi32(a, valueShift), _mm_add_epi32(b, valueShift));
         __m128i aTob = _mm_xor_si128(a, b); // a ^ (a ^ b) == b
@@ -1448,7 +1459,7 @@ unittest
 }
 
 /// Compare packed unsigned 32-bit integers in a and b, and store packed minimum values in dst.
-__m128i _mm_min_epu32 (__m128i a, __m128i b) @trusted
+__m128i _mm_min_epu32 (__m128i a, __m128i b) pure @trusted
 {
     // PERF DMD
     static if (GDC_with_SSE41)
@@ -1463,7 +1474,7 @@ __m128i _mm_min_epu32 (__m128i a, __m128i b) @trusted
         uint4 sb = cast(uint4)b;
         static if (SIMD_COMPARISON_MASKS_16B)
             uint4 greater = sa > sb;
-        else        
+        else
             uint4 greater = cast(uint4) greaterMask!uint4(sa, sb);
         return cast(__m128i)( (~greater & sa) | (greater & sb) );
     }