embree-4.3.0-e2k.patch

From 46cab529d9b854214c058561bc58d62a698ca72a Mon Sep 17 00:00:00 2001
From: Ilya Kurdyukov <jpegqs@gmail.com>
Date: Fri, 27 Oct 2023 10:53:19 +0700
Subject: [PATCH] embree-4.3.0 e2k support

Use cmake option -DEMBREE_MAX_ISA=DEFAULT
---
 common/math/vec2fa.h      | 2 +-
 common/math/vec3fa.h      | 4 ++--
 common/sys/intrinsics.h   | 6 ++++++
 common/sys/platform.h     | 2 +-
 common/sys/sysinfo.cpp    | 7 +++++++
 common/sys/sysinfo.h      | 1 +
 kernels/common/device.cpp | 5 +++++
 7 files changed, 23 insertions(+), 4 deletions(-)

diff --git a/common/math/vec2fa.h b/common/math/vec2fa.h
index d57e549..922b252 100644
--- a/common/math/vec2fa.h
+++ b/common/math/vec2fa.h
@@ -252,7 +252,7 @@ namespace embree
   /// Euclidean Space Operators
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) && !defined(__e2k__)
   __forceinline float dot ( const Vec2fa& a, const Vec2fa& b ) {
     return _mm_cvtss_f32(_mm_dp_ps(a,b,0x3F));
   }
diff --git a/common/math/vec3fa.h b/common/math/vec3fa.h
index 967e75d..fbb0432 100644
--- a/common/math/vec3fa.h
+++ b/common/math/vec3fa.h
@@ -321,7 +321,7 @@ namespace embree
   /// Euclidean Space Operators
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) && !defined(__e2k__)
   __forceinline float dot ( const Vec3fa& a, const Vec3fa& b ) {
     return _mm_cvtss_f32(_mm_dp_ps(a.m128,b.m128,0x7F));
   }
@@ -691,7 +691,7 @@ namespace embree
   /// Euclidean Space Operators
   ////////////////////////////////////////////////////////////////////////////////
 
-#if defined(__SSE4_1__)
+#if defined(__SSE4_1__) && !defined(__e2k__)
   __forceinline float dot ( const Vec3fx& a, const Vec3fx& b ) {
     return _mm_cvtss_f32(_mm_dp_ps(a.m128,b.m128,0x7F));
   }
diff --git a/common/sys/intrinsics.h b/common/sys/intrinsics.h
index f5074bb..69b67f3 100644
--- a/common/sys/intrinsics.h
+++ b/common/sys/intrinsics.h
@@ -18,6 +18,10 @@
 #endif
 #endif
 
+#ifdef __e2k__
+#include <x86intrin.h>
+#endif
+
 #if defined(__BMI__) && defined(__GNUC__) && !defined(__INTEL_COMPILER)
   #if !defined(_tzcnt_u32)
     #define _tzcnt_u32 __tzcnt_u32
@@ -201,6 +205,8 @@ namespace embree
     uint32_t high,low;
     asm volatile ("rdtsc" : "=d"(high), "=a"(low));
     return (((uint64_t)high) << 32) + (uint64_t)low;
+#elif defined(__e2k__)
+    return __rdtsc();
 #else
     /* Not supported yet, meaning measuring traversal cost per pixel does not work. */
     return 0;
diff --git a/common/sys/platform.h b/common/sys/platform.h
index 6dc0cf3..002e5a2 100644
--- a/common/sys/platform.h
+++ b/common/sys/platform.h
@@ -58,7 +58,7 @@
 #endif
 
 /* detect 64 bit platform */
-#if defined(__X86_64__) || defined(__aarch64__)
+#if defined(__X86_64__) || defined(__aarch64__) || defined(__e2k__)
 #define __64BIT__
 #endif
 
diff --git a/common/sys/sysinfo.cpp b/common/sys/sysinfo.cpp
index 1e45061..d6bc3dc 100644
--- a/common/sys/sysinfo.cpp
+++ b/common/sys/sysinfo.cpp
@@ -100,6 +100,8 @@ namespace embree
     return (char*)name;
 #elif defined(__ARM_NEON)
     return "ARM";
+#elif defined(__e2k__)
+    return "MCST";
 #else
     return "Unknown";
 #endif
@@ -176,6 +178,8 @@ namespace embree
     
 #elif defined(__ARM_NEON)
     return CPU::ARM;
+#elif defined(__e2k__)
+    return CPU::E2K;
 #endif
     
     return CPU::UNKNOWN;
@@ -205,6 +209,7 @@ namespace embree
     case CPU::CORE2                   : return "Core2";
     case CPU::CORE1                   : return "Core";
     case CPU::ARM                     : return "ARM";
+    case CPU::E2K                     : return "Elbrus 2000";
     case CPU::UNKNOWN                 : return "Unknown CPU";
     }
     return "Unknown CPU (error)";
@@ -363,6 +368,8 @@ namespace embree
     cpu_features |= CPU_FEATURE_NEON_2X;
     return cpu_features;
 
+#elif defined(__e2k__)
+    return ISA;
 #else
     /* Unknown CPU. */
     return 0;
diff --git a/common/sys/sysinfo.h b/common/sys/sysinfo.h
index cefd39a..ffe82ca 100644
--- a/common/sys/sysinfo.h
+++ b/common/sys/sysinfo.h
@@ -89,6 +89,7 @@ namespace embree
     CORE2,
     CORE1,
     ARM,
+    E2K,
     UNKNOWN,
   };
   
diff --git a/kernels/common/device.cpp b/kernels/common/device.cpp
index c2e7c7c..c27a2dd 100644
--- a/kernels/common/device.cpp
+++ b/kernels/common/device.cpp
@@ -78,6 +78,11 @@ namespace embree
     case CPU::XEON_PHI_KNIGHTS_MILL   : frequency_level = FREQUENCY_SIMD512; break;
     case CPU::XEON_PHI_KNIGHTS_LANDING: frequency_level = FREQUENCY_SIMD512; break;
     case CPU::ARM:             frequency_level = FREQUENCY_SIMD256; break;
+#ifdef __AVX__
+    case CPU::E2K:             frequency_level = FREQUENCY_SIMD256; break;
+#else
+    case CPU::E2K:             frequency_level = FREQUENCY_SIMD128; break;
+#endif
     }
 
     /* initialize global state */
-- 
2.34.1