Cleanup: Cycles: Remove unnecessary SSE4.2 CPU kernel

This is the minimum requirement, so just the regular kernel already
includes these instructions if supported by the CPU architecture.

intern/cycles/device/cpu/device.cpp

@@ -49,7 +49,6 @@ void device_cpu_info(vector<DeviceInfo> &devices)
 string device_cpu_capabilities()
 {
   string capabilities;
-  capabilities += system_cpu_support_sse42() ? "SSE42 " : "";
   capabilities += system_cpu_support_avx2() ? "AVX2" : "";
   if (capabilities[capabilities.size() - 1] == ' ') {
     capabilities.resize(capabilities.size() - 1);

intern/cycles/device/cpu/kernel.cpp

@@ -8,8 +8,7 @@
 
 CCL_NAMESPACE_BEGIN
 
-#define KERNEL_FUNCTIONS(name) \
+#define KERNEL_FUNCTIONS(name) KERNEL_NAME_EVAL(cpu, name), KERNEL_NAME_EVAL(cpu_avx2, name)
-  KERNEL_NAME_EVAL(cpu, name), KERNEL_NAME_EVAL(cpu_sse42, name), KERNEL_NAME_EVAL(cpu_avx2, name)
 
 #define REGISTER_KERNEL(name) name(KERNEL_FUNCTIONS(name))
 #define REGISTER_KERNEL_FILM_CONVERT(name) \

intern/cycles/device/cpu/kernel_function.h

@@ -13,14 +13,12 @@ CCL_NAMESPACE_BEGIN
  *
  * Provides a function-call-like API which gets routed to the most suitable implementation.
  *
- * For example, on a computer which only has SSE4.2 the kernel_sse42 will be used. */
+ * For example, on a computer which only has AVX2 the kernel_avx2 will be used. */
 template<typename FunctionType> class CPUKernelFunction {
  public:
-  CPUKernelFunction(FunctionType kernel_default,
+  CPUKernelFunction(FunctionType kernel_default, FunctionType kernel_avx2)
-                    FunctionType kernel_sse42,
-                    FunctionType kernel_avx2)
   {
-    kernel_info_ = get_best_kernel_info(kernel_default, kernel_sse42, kernel_avx2);
+    kernel_info_ = get_best_kernel_info(kernel_default, kernel_avx2);
   }
 
   template<typename... Args> auto operator()(Args... args) const
@@ -53,12 +51,9 @@ template<typename FunctionType> class CPUKernelFunction {
     FunctionType kernel;
   };
 
-  KernelInfo get_best_kernel_info(FunctionType kernel_default,
+  KernelInfo get_best_kernel_info(FunctionType kernel_default, FunctionType kernel_avx2)
-                                  FunctionType kernel_sse42,
-                                  FunctionType kernel_avx2)
   {
     /* Silence warnings about unused variables when compiling without some architectures. */
-    (void)kernel_sse42;
     (void)kernel_avx2;
 
 #ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX2
@@ -67,12 +62,6 @@ template<typename FunctionType> class CPUKernelFunction {
     }
 #endif
 
-#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE42
-    if (DebugFlags().cpu.has_sse42() && system_cpu_support_sse42()) {
-      return KernelInfo("SSE4.2", kernel_sse42);
-    }
-#endif
-
     return KernelInfo("default", kernel_default);
   }
 

intern/cycles/kernel/CMakeLists.txt

@@ -15,7 +15,6 @@ set(INC_SYS
 set(SRC_KERNEL_DEVICE_CPU
   device/cpu/globals.cpp
   device/cpu/kernel.cpp
-  device/cpu/kernel_sse42.cpp
   device/cpu/kernel_avx2.cpp
 )
 
@@ -1348,10 +1347,6 @@ if(DEFINED CYCLES_KERNEL_FLAGS)
   set_source_files_properties(device/cpu/kernel.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_KERNEL_FLAGS}")
 endif()
 
-if(CXX_HAS_SSE42)
-  set_source_files_properties(device/cpu/kernel_sse42.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE42_FLAGS}")
-endif()
-
 if(CXX_HAS_AVX2)
   set_source_files_properties(device/cpu/kernel_avx2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX2_FLAGS}")
 endif()

intern/cycles/kernel/device/cpu/kernel.h

@@ -35,9 +35,6 @@ void kernel_global_memory_copy(KernelGlobalsCPU *kg,
 #define KERNEL_ARCH cpu
 #include "kernel/device/cpu/kernel_arch.h"
 
-#define KERNEL_ARCH cpu_sse42
-#include "kernel/device/cpu/kernel_arch.h"
-
 #define KERNEL_ARCH cpu_avx2
 #include "kernel/device/cpu/kernel_arch.h"
 

intern/cycles/kernel/device/cpu/kernel_sse42.cpp

@@ -1,27 +0,0 @@
-/* SPDX-FileCopyrightText: 2011-2022 Blender Foundation
- *
- * SPDX-License-Identifier: Apache-2.0 */
-
-/* Optimized CPU kernel entry points. This file is compiled with SSE42
- * optimization flags and nearly all functions inlined, while kernel.cpp
- * is compiled without for other CPU's. */
-
-#include "util/optimization.h"
-
-#ifndef WITH_CYCLES_OPTIMIZED_KERNEL_SSE42
-#  define KERNEL_STUB
-#else
-/* SSE optimization disabled for now on 32 bit, see bug #36316. */
-#  if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
-#    define __KERNEL_SSE__
-#    define __KERNEL_SSE2__
-#    define __KERNEL_SSE3__
-#    define __KERNEL_SSSE3__
-#    define __KERNEL_SSE42__
-#  endif
-#endif /* WITH_CYCLES_OPTIMIZED_KERNEL_SSE42 */
-
-#include "kernel/device/cpu/globals.h"
-#include "kernel/device/cpu/kernel.h"
-#define KERNEL_ARCH cpu_sse42
-#include "kernel/device/cpu/kernel_arch_impl.h"

intern/cycles/util/CMakeLists.txt

@@ -27,7 +27,6 @@ set(SRC
   time.cpp
   transform.cpp
   transform_avx2.cpp
-  transform_sse42.cpp
   windows.cpp
 )
 
@@ -125,9 +124,6 @@ set(SRC_HEADERS
   xml.h
 )
 
-if(CXX_HAS_SSE42)
-  set_source_files_properties(transform_sse42.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE42_FLAGS}")
-endif()
 if(CXX_HAS_AVX2)
   set_source_files_properties(transform_avx2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX2_FLAGS}")
 endif()

intern/cycles/util/debug.cpp

@@ -27,7 +27,6 @@ void DebugFlags::CPU::reset()
   } while (0)
 
   CHECK_CPU_FLAGS(avx2, "CYCLES_CPU_NO_AVX2");
-  CHECK_CPU_FLAGS(sse42, "CYCLES_CPU_NO_SSE42");
 
 #undef STRINGIFY
 #undef CHECK_CPU_FLAGS

intern/cycles/util/optimization.h

@@ -8,16 +8,10 @@
 
 /* x86
  *
- * Compile a regular and SSE42 kernel. */
+ * Compile a regular kernel. */
 
 #  if defined(i386) || defined(_M_IX86)
 
-/* We require minimum SSE4.2 support on x86, so auto enable. */
-#    define __KERNEL_SSE42__
-#    ifdef WITH_KERNEL_SSE42
-#      define WITH_CYCLES_OPTIMIZED_KERNEL_SSE42
-#    endif
-
 /* x86-64
  *
  * Compile a regular (includes SSE4.2) and AVX2 kernel. */

intern/cycles/util/transform.h

@@ -408,7 +408,6 @@ ccl_device_inline float4 quat_interpolate(const float4 q1, const float4 q2, cons
 }
 
 #ifndef __KERNEL_GPU__
-void transform_inverse_cpu_sse42(const Transform &tfm, Transform &itfm);
 void transform_inverse_cpu_avx2(const Transform &tfm, Transform &itfm);
 #endif
 
@@ -497,11 +496,6 @@ ccl_device_inline Transform transform_inverse(const Transform tfm)
     transform_inverse_cpu_avx2(tfm, itfm);
     return itfm;
   }
-  if (system_cpu_support_sse42()) {
-    Transform itfm;
-    transform_inverse_cpu_sse42(tfm, itfm);
-    return itfm;
-  }
 #endif
 
   return transform_inverse_impl(tfm);

intern/cycles/util/transform_sse42.cpp

@@ -1,14 +0,0 @@
-/* SPDX-FileCopyrightText: 2011-2022 Blender Foundation
- *
- * SPDX-License-Identifier: Apache-2.0 */
-
-#include "util/transform.h"
-
-CCL_NAMESPACE_BEGIN
-
-void transform_inverse_cpu_sse42(const Transform &tfm, Transform &itfm)
-{
-  itfm = transform_inverse_impl(tfm);
-}
-
-CCL_NAMESPACE_END