From 16b8d223b718925d6aadd30b802967f010076f5d Mon Sep 17 00:00:00 2001
From: Ray Molenkamp <github@lazydodo.com>
Date: Mon, 17 Sep 2018 16:27:13 -0600
Subject: [PATCH] Cycles: Fix usage of AVX2 intrinsics in AVX kernel

While building the AVX kernel, util_avxf.h/avxb.h were using some AVX2 intrinsics,
these were never called, so it wasn't a run-time issue, but the intrinsics headers
on centos excluded the AVX2 prototypes when building the AVX kernel causing build errors.

This commit cleans up the improper usage of the AVX2 intrinsics and provides AVX
fallback implementations for future use.

Differential Revision: https://developer.blender.org/D3696
---
 intern/cycles/util/util_avxb.h | 18 ++++++++++++++++--
 intern/cycles/util/util_avxf.h |  6 ++++--
 2 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/intern/cycles/util/util_avxb.h b/intern/cycles/util/util_avxb.h
index b6d77857c6f..4add69bcb91 100644
--- a/intern/cycles/util/util_avxb.h
+++ b/intern/cycles/util/util_avxb.h
@@ -49,7 +49,7 @@ struct avxb
 	////////////////////////////////////////////////////////////////////////////////
 
 	__forceinline avxb( FalseTy ) : m256(_mm256_setzero_ps()) {}
-	__forceinline avxb( TrueTy  ) : m256(_mm256_castsi256_ps(_mm256_cmpeq_epi32(_mm256_setzero_si256(), _mm256_setzero_si256()))) {}
+	__forceinline avxb( TrueTy  ) : m256(_mm256_castsi256_ps(_mm256_set1_epi32(-1))) {}
 
 	////////////////////////////////////////////////////////////////////////////////
 	/// Array Access
@@ -86,7 +86,21 @@ __forceinline const avxb operator ^=( avxb& a, const avxb& b ) { return a = a ^
 ////////////////////////////////////////////////////////////////////////////////
 
 __forceinline const avxb operator !=( const avxb& a, const avxb& b ) { return _mm256_xor_ps(a, b); }
-__forceinline const avxb operator ==( const avxb& a, const avxb& b ) { return _mm256_castsi256_ps(_mm256_cmpeq_epi32(a, b)); }
+__forceinline const avxb operator ==( const avxb& a, const avxb& b )
+{
+#ifdef __KERNEL_AVX2__
+	return _mm256_castsi256_ps(_mm256_cmpeq_epi32(a, b));
+#else
+	__m128i a_lo = _mm_castps_si128(_mm256_extractf128_ps(a, 0));
+	__m128i a_hi = _mm_castps_si128(_mm256_extractf128_ps(a, 1));
+	__m128i b_lo = _mm_castps_si128(_mm256_extractf128_ps(b, 0));
+	__m128i b_hi = _mm_castps_si128(_mm256_extractf128_ps(b, 1));
+	__m128i c_lo = _mm_cmpeq_epi32(a_lo, b_lo);
+	__m128i c_hi = _mm_cmpeq_epi32(a_hi, b_hi);
+	__m256i result = _mm256_insertf128_si256(_mm256_castsi128_si256(c_lo), c_hi, 1);
+	return _mm256_castsi256_ps(result);
+#endif 
+}
 
 __forceinline const avxb select( const avxb& m, const avxb& t, const avxb& f ) {
 #if defined(__KERNEL_SSE41__)
diff --git a/intern/cycles/util/util_avxf.h b/intern/cycles/util/util_avxf.h
index 5596702ca20..0c15ba5cbbd 100644
--- a/intern/cycles/util/util_avxf.h
+++ b/intern/cycles/util/util_avxf.h
@@ -214,17 +214,19 @@ __forceinline const avxf nmadd(const avxf& a, const avxf& b, const avxf& c) {
 #endif
 }
 __forceinline const avxf msub(const avxf& a, const avxf& b, const avxf& c) {
+#ifdef __KERNEL_AVX2__
 	return _mm256_fmsub_ps(a, b, c);
+#else
+	return (a*b) - c;
+#endif
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 /// Comparison Operators
 ////////////////////////////////////////////////////////////////////////////////
-#ifdef __KERNEL_AVX2__
 __forceinline const avxb operator <=(const avxf& a, const avxf& b) {
 	return _mm256_cmp_ps(a.m256, b.m256, _CMP_LE_OS);
 }
-#endif
 
 #endif