From 1415f5b0b13e5706aa7afb0bb7ecf90781eae418 Mon Sep 17 00:00:00 2001
From: Dmitry Babokin <babokin@gmail.com>
Date: Fri, 19 Feb 2016 15:19:23 +0300
Subject: [PATCH] Adding avx512knl-32x16 target to examples Makefiles

---
 examples/aobench/Makefile          | 2 +-
 examples/common.mk                 | 5 ++++-
 examples/deferred/Makefile         | 2 +-
 examples/gmres/Makefile            | 2 +-
 examples/mandelbrot/Makefile       | 2 +-
 examples/mandelbrot_tasks/Makefile | 2 +-
 examples/noise/Makefile            | 2 +-
 examples/options/Makefile          | 2 +-
 examples/perfbench/Makefile        | 2 +-
 examples/rt/Makefile               | 2 +-
 examples/sort/Makefile             | 2 +-
 examples/stencil/Makefile          | 2 +-
 examples/volume_rendering/Makefile | 2 +-
 perf.py                            | 2 +-
 14 files changed, 17 insertions(+), 14 deletions(-)

diff --git a/examples/aobench/Makefile b/examples/aobench/Makefile
index c8122c07..16417eaa 100644
--- a/examples/aobench/Makefile
+++ b/examples/aobench/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=ao
 CPP_SRC=ao.cpp ao_serial.cpp
 ISPC_SRC=ao.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x4,avx1-i32x8,avx2-i32x8
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x4,avx1-i32x8,avx2-i32x8,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/common.mk b/examples/common.mk
index 04a566bb..b337dcae 100644
--- a/examples/common.mk
+++ b/examples/common.mk
@@ -35,6 +35,9 @@ ifeq ($(ARCH),x86)
     ifneq (,$(findstring avx2,$(ISPC_IA_TARGETS)))
       ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_avx2.o)
     endif
+    ifneq (,$(findstring avx512knl,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_avx512knl.o)
+    endif
   endif
   ISPC_TARGETS=$(ISPC_IA_TARGETS)
   ARCH_BIT:=$(shell getconf LONG_BIT)
@@ -86,7 +89,7 @@ objs/%.o: ../%.cpp dirs
 
 objs/$(EXAMPLE).o: objs/$(EXAMPLE)_ispc.h dirs
 
-objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o objs/%_ispc_avx11.o objs/%_ispc_avx2.o: %.ispc dirs
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o objs/%_ispc_avx11.o objs/%_ispc_avx2.o objs/%_ispc_avx512knl.o : %.ispc dirs
 	$(ISPC) $(ISPC_FLAGS) --target=$(ISPC_TARGETS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h
 
 objs/$(ISPC_SRC:.ispc=)_sse4.cpp: $(ISPC_SRC)
diff --git a/examples/deferred/Makefile b/examples/deferred/Makefile
index be8ce7c4..31813527 100644
--- a/examples/deferred/Makefile
+++ b/examples/deferred/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=deferred_shading
 CPP_SRC=common.cpp main.cpp dynamic_c.cpp dynamic_cilk.cpp
 ISPC_SRC=kernels.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 ISPC_FLAGS=--opt=fast-math
 
diff --git a/examples/gmres/Makefile b/examples/gmres/Makefile
index 07765069..c4bacf0b 100644
--- a/examples/gmres/Makefile
+++ b/examples/gmres/Makefile
@@ -3,7 +3,7 @@ EXAMPLE=gmres
 CPP_SRC=algorithm.cpp main.cpp matrix.cpp
 CC_SRC=mmio.c
 ISPC_SRC=matrix.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/mandelbrot/Makefile b/examples/mandelbrot/Makefile
index d225037d..59517034 100644
--- a/examples/mandelbrot/Makefile
+++ b/examples/mandelbrot/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=mandelbrot
 CPP_SRC=mandelbrot.cpp mandelbrot_serial.cpp
 ISPC_SRC=mandelbrot.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/mandelbrot_tasks/Makefile b/examples/mandelbrot_tasks/Makefile
index 51866b32..64f76c2a 100644
--- a/examples/mandelbrot_tasks/Makefile
+++ b/examples/mandelbrot_tasks/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=mandelbrot_tasks
 CPP_SRC=mandelbrot_tasks.cpp mandelbrot_tasks_serial.cpp
 ISPC_SRC=mandelbrot_tasks.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/noise/Makefile b/examples/noise/Makefile
index 6dd5ae42..0bc4e244 100644
--- a/examples/noise/Makefile
+++ b/examples/noise/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=noise
 CPP_SRC=noise.cpp noise_serial.cpp
 ISPC_SRC=noise.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x4,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x4,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/options/Makefile b/examples/options/Makefile
index 2da7e01a..53f0332f 100644
--- a/examples/options/Makefile
+++ b/examples/options/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=options
 CPP_SRC=options.cpp options_serial.cpp
 ISPC_SRC=options.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/perfbench/Makefile b/examples/perfbench/Makefile
index cc2e681f..39c1f866 100644
--- a/examples/perfbench/Makefile
+++ b/examples/perfbench/Makefile
@@ -1,5 +1,5 @@
 
-EXAMPLE=perbench
+EXAMPLE=perfbench
 CPP_SRC=perfbench.cpp perfbench_serial.cpp
 ISPC_SRC=perfbench.ispc
 ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x4,avx1-i32x8,avx2-i32x8
diff --git a/examples/rt/Makefile b/examples/rt/Makefile
index e52b02e4..0b7803ec 100644
--- a/examples/rt/Makefile
+++ b/examples/rt/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=rt
 CPP_SRC=rt.cpp rt_serial.cpp
 ISPC_SRC=rt.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/sort/Makefile b/examples/sort/Makefile
index 6ae43fae..d099aac9 100644
--- a/examples/sort/Makefile
+++ b/examples/sort/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=sort
 CPP_SRC=sort.cpp sort_serial.cpp
 ISPC_SRC=sort.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 #ISPC_FLAGS=-DDEBUG
 
diff --git a/examples/stencil/Makefile b/examples/stencil/Makefile
index 1b9c2717..9fe1ba08 100644
--- a/examples/stencil/Makefile
+++ b/examples/stencil/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=stencil
 CPP_SRC=stencil.cpp stencil_serial.cpp
 ISPC_SRC=stencil.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x16,avx2-i32x16,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/examples/volume_rendering/Makefile b/examples/volume_rendering/Makefile
index 1bc81e4e..1458edc1 100644
--- a/examples/volume_rendering/Makefile
+++ b/examples/volume_rendering/Makefile
@@ -2,7 +2,7 @@
 EXAMPLE=volume
 CPP_SRC=volume.cpp volume_serial.cpp
 ISPC_SRC=volume.ispc
-ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8
+ISPC_IA_TARGETS=sse2-i32x4,sse4-i32x8,avx1-i32x8,avx2-i32x8,avx512knl-i32x16
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk
diff --git a/perf.py b/perf.py
index 30189040..2ea2b510 100755
--- a/perf.py
+++ b/perf.py
@@ -339,7 +339,7 @@ def perf(options1, args):
     if options.perf_target != "":
         test_only_r = " sse2-i32x4 sse2-i32x8 sse4-i32x4 sse4-i32x8 sse4-i16x8 \
                         sse4-i8x16 avx1-i32x4 avx1-i32x8 avx1-i32x16 avx1-i64x4 avx1.1-i32x8 \
-                        avx1.1-i32x16 avx1.1-i64x4 avx2-i32x8 avx2-i32x16 avx2-i64x4 "
+                        avx1.1-i32x16 avx1.1-i64x4 avx2-i32x8 avx2-i32x16 avx2-i64x4 avx512knl-i32x16"
         test_only = options.perf_target.split(",")
         for iterator in test_only:
             if not (" " + iterator + " " in test_only_r):