fix: nothing

NoNaeAbC · Apr 3, 2021 · ce36428 · ce36428
1 parent e1ef4dc
commit ce36428
Show file tree

Hide file tree

Showing 8 changed files with 220 additions and 227 deletions.
diff --git a/.gitignore b/.gitignore
@@ -4,6 +4,7 @@ cmake-build-debug
 *.html
 *.js
 *.wasm
+*.out
 
 samples/mandelbrot
 

diff --git a/amathlib.h b/amathlib.h
@@ -231,7 +231,6 @@
 
 
 #endif
-
 #if defined(AML_USE_GMP) && !defined(AML_USE_AML_NUMBER)
 #define AML_USE_AML_NUMBER
 #endif
@@ -1090,159 +1089,6 @@ class AML_PREFIX(VectorU8_8D) {
 
 #include "amllahead.h"
 
-#if defined(USE_AVX512) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
-#define COMPLEX_DEFINITIONS
-
-#define MAX_COMPLEX_64_SIZE 8
-#define MAX_COMPLEX_64_TYPE Array8Complex64
-#define MAX_COMPLEX_64_MASK_TYPE VectorU8_8D
-#define MAX_COMPLEX_64_VECTOR_TYPE Vector8D_64
-#define IDEAL_COMPLEX_64_SIZE 8
-#define IDEAL_COMPLEX_64_TYPE Array8Complex64
-#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_8D
-#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector8D_64
-#define MIN_COMPLEX_64_SIZE 4
-#define MIN_COMPLEX_64_TYPE Array4Complex64
-#define MIN_COMPLEX_64_MASK_TYPE VectorU8_4D
-#define MIN_COMPLEX_64_VECTOR_TYPE Vector4D_64
-
-#define MAX_COMPLEX_32_SIZE 8
-#define MAX_COMPLEX_32_TYPE Array8Complex32
-#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
-#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
-#define IDEAL_COMPLEX_32_SIZE 8
-#define IDEAL_COMPLEX_32_TYPE Array8Complex32
-#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_8D
-#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector8D_32
-#define MIN_COMPLEX_32_SIZE 4
-#define MIN_COMPLEX_32_TYPE Array4Complex32
-#define MIN_COMPLEX_32_MASK_TYPE VectorU8_4D
-#define MIN_COMPLEX_32_VECTOR_TYPE Vector4D_64
-
-
-#elif defined(USE_AVX) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
-#define COMPLEX_DEFINITIONS
-
-#define MAX_COMPLEX_64_SIZE 8
-#define MAX_COMPLEX_64_TYPE Array8Complex64
-#define MAX_COMPLEX_64_MASK_TYPE VectorU8_8D
-#define MAX_COMPLEX_64_VECTOR_TYPE Vector8D_64
-#define IDEAL_COMPLEX_64_SIZE 4
-#define IDEAL_COMPLEX_64_TYPE Array4Complex64
-#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_4D
-#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector4D_64
-#define MIN_COMPLEX_64_SIZE 2
-#define MIN_COMPLEX_64_TYPE Array2Complex64
-#define MIN_COMPLEX_64_MASK_TYPE VectorU8_2D
-#define MIN_COMPLEX_64_VECTOR_TYPE Vector2D_64
-
-#define MAX_COMPLEX_32_SIZE 8
-#define MAX_COMPLEX_32_TYPE Array8Complex32
-#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
-#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
-#define IDEAL_COMPLEX_32_SIZE 8
-#define IDEAL_COMPLEX_32_TYPE Array8Complex32
-#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_8D
-#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector8D_32
-#define MIN_COMPLEX_32_SIZE 4
-#define MIN_COMPLEX_32_TYPE Array4Complex32
-#define MIN_COMPLEX_32_MASK_TYPE VectorU8_4D
-#define MIN_COMPLEX_32_VECTOR_TYPE Vector4D_32
-
-
-#elif defined(USE_SSE) || defined(USE_NEON) || defined(USE_WASM_SIMD) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
-#define COMPLEX_DEFINITIONS
-
-#define MAX_COMPLEX_64_SIZE 4
-#define MAX_COMPLEX_64_TYPE Array4Complex64
-#define MAX_COMPLEX_64_MASK_TYPE VectorU8_4D
-#define MAX_COMPLEX_64_VECTOR_TYPE Vector4D_64
-#define IDEAL_COMPLEX_64_SIZE 2
-#define IDEAL_COMPLEX_64_TYPE Array2Complex64
-#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_2D
-#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector2D_64
-#define MIN_COMPLEX_64_SIZE 1
-#define MIN_COMPLEX_64_TYPE Complex64
-#define MIN_COMPLEX_64_MASK_TYPE VectorU8_1D
-#define MIN_COMPLEX_64_VECTOR_TYPE Vecto1D_64
-
-
-#define MAX_COMPLEX_32_SIZE 8
-#define MAX_COMPLEX_32_TYPE Array8Complex32
-#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
-#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
-#define IDEAL_COMPLEX_32_SIZE 4
-#define IDEAL_COMPLEX_32_TYPE Array4Complex32
-#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_4D
-#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector4D_32
-#define MIN_COMPLEX_32_SIZE 2
-#define MIN_COMPLEX_32_TYPE Array2Complex32
-#define MIN_COMPLEX_32_MASK_TYPE VectorU8_2D
-#define MIN_COMPLEX_32_VECTOR_TYPE Vector2D_32
-
-#elif defined(AML_USE_ARRAY_STRICT)
-
-#define MAX_COMPLEX_64_SIZE 2
-#define MAX_COMPLEX_64_TYPE Array2Complex64
-#define MAX_COMPLEX_64_MASK_TYPE VectorU8_2D
-#define MAX_COMPLEX_64_VECTOR_TYPE Vector2D_64
-#define IDEAL_COMPLEX_64_SIZE 2
-#define IDEAL_COMPLEX_64_TYPE Array2Complex64
-#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_2D
-#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector2D_64
-#define MIN_COMPLEX_64_SIZE 2
-#define MIN_COMPLEX_64_TYPE Array2Complex64
-#define MIN_COMPLEX_64_MASK_TYPE VectorU8_2D
-#define MIN_COMPLEX_64_VECTOR_TYPE Vector2D_64
-
-
-#define MAX_COMPLEX_32_SIZE 2
-#define MAX_COMPLEX_32_TYPE Array2Complex32
-#define MAX_COMPLEX_32_MASK_TYPE VectorU8_2D
-#define MAX_COMPLEX_32_VECTOR_TYPE Vector2D_32
-#define IDEAL_COMPLEX_32_SIZE 2
-#define IDEAL_COMPLEX_32_TYPE Array2Complex32
-#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_2D
-#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector2D_32
-#define MIN_COMPLEX_32_SIZE 2
-#define MIN_COMPLEX_32_TYPE Array2Complex32
-#define MIN_COMPLEX_32_MASK_TYPE VectorU8_2D
-#define MIN_COMPLEX_32_VECTOR_TYPE Vector2D_32
-
-
-#elif !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
-#define COMPLEX_DEFINITIONS
-
-
-#define MAX_COMPLEX_64_SIZE 1
-#define MAX_COMPLEX_64_TYPE Complex64
-#define MAX_COMPLEX_64_MASK_TYPE VectorU8_1D
-#define MAX_COMPLEX_64_VECTOR_TYPE Vecto1D_64
-#define IDEAL_COMPLEX_64_SIZE 1
-#define IDEAL_COMPLEX_64_TYPE Complex64
-#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_1D
-#define IDEAL_COMPLEX_64_VECTOR_TYPE Vecto1D_64
-#define MIN_COMPLEX_64_SIZE 1
-#define MIN_COMPLEX_64_TYPE Complex64
-#define MIN_COMPLEX_64_MASK_TYPE VectorU8_1D
-#define MIN_COMPLEX_64_VECTOR_TYPE Vecto1D_64
-
-#define MAX_COMPLEX_32_SIZE 1
-#define MAX_COMPLEX_32_TYPE Complex32
-#define MAX_COMPLEX_32_MASK_TYPE VectorU8_1D
-#define MAX_COMPLEX_32_VECTOR_TYPE Vecto1D_32
-#define IDEAL_COMPLEX_32_SIZE 1
-#define IDEAL_COMPLEX_32_TYPE Complex32
-#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_1D
-#define IDEAL_COMPLEX_32_VECTOR_TYPE Vecto1D_32
-#define MIN_COMPLEX_32_SIZE 1
-#define MIN_COMPLEX_32_TYPE Complex32
-#define MIN_COMPLEX_32_MASK_TYPE VectorU8_1D
-#define MIN_COMPLEX_32_VECTOR_TYPE Vecto1D_32
-
-
-#endif
-
 #include "amlcomplexhead.h"
 
 class AML_PREFIX(Array2Complex64) {
@@ -12117,3 +11963,158 @@ AML_FUNCTION AML_PREFIX(ComplexN) operator/(const AmlNumber &lhs, const AML_PREF
 #endif
 
 #endif //MATH_LIB_A_MATH_LIB_H
+
+
+#if defined(USE_AVX512) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
+#define COMPLEX_DEFINITIONS
+
+#define MAX_COMPLEX_64_SIZE 8
+#define MAX_COMPLEX_64_TYPE Array8Complex64
+#define MAX_COMPLEX_64_MASK_TYPE VectorU8_8D
+#define MAX_COMPLEX_64_VECTOR_TYPE Vector8D_64
+#define IDEAL_COMPLEX_64_SIZE 8
+#define IDEAL_COMPLEX_64_TYPE Array8Complex64
+#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_8D
+#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector8D_64
+#define MIN_COMPLEX_64_SIZE 4
+#define MIN_COMPLEX_64_TYPE Array4Complex64
+#define MIN_COMPLEX_64_MASK_TYPE VectorU8_4D
+#define MIN_COMPLEX_64_VECTOR_TYPE Vector4D_64
+
+#define MAX_COMPLEX_32_SIZE 8
+#define MAX_COMPLEX_32_TYPE Array8Complex32
+#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
+#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
+#define IDEAL_COMPLEX_32_SIZE 8
+#define IDEAL_COMPLEX_32_TYPE Array8Complex32
+#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_8D
+#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector8D_32
+#define MIN_COMPLEX_32_SIZE 4
+#define MIN_COMPLEX_32_TYPE Array4Complex32
+#define MIN_COMPLEX_32_MASK_TYPE VectorU8_4D
+#define MIN_COMPLEX_32_VECTOR_TYPE Vector4D_64
+
+
+#elif defined(USE_AVX) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
+#define COMPLEX_DEFINITIONS
+
+#define MAX_COMPLEX_64_SIZE 8
+#define MAX_COMPLEX_64_TYPE Array8Complex64
+#define MAX_COMPLEX_64_MASK_TYPE VectorU8_8D
+#define MAX_COMPLEX_64_VECTOR_TYPE Vector8D_64
+#define IDEAL_COMPLEX_64_SIZE 4
+typedef Array4Complex64 IDEAL_COMPLEX_64_TYPE;
+typedef VectorU8_4D IDEAL_COMPLEX_64_MASK_TYPE;
+typedef Vector4D_64 IDEAL_COMPLEX_64_VECTOR_TYPE;
+#define MIN_COMPLEX_64_SIZE 2
+#define MIN_COMPLEX_64_TYPE Array2Complex64
+#define MIN_COMPLEX_64_MASK_TYPE VectorU8_2D
+#define MIN_COMPLEX_64_VECTOR_TYPE Vector2D_64
+
+#define MAX_COMPLEX_32_SIZE 8
+#define MAX_COMPLEX_32_TYPE Array8Complex32
+#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
+#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
+#define IDEAL_COMPLEX_32_SIZE 8
+#define IDEAL_COMPLEX_32_TYPE Array8Complex32
+#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_8D
+#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector8D_32
+#define MIN_COMPLEX_32_SIZE 4
+#define MIN_COMPLEX_32_TYPE Array4Complex32
+#define MIN_COMPLEX_32_MASK_TYPE VectorU8_4D
+#define MIN_COMPLEX_32_VECTOR_TYPE Vector4D_32
+
+
+#elif defined(USE_SSE) || defined(USE_NEON) || defined(USE_WASM_SIMD) && !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
+#define COMPLEX_DEFINITIONS
+
+#define MAX_COMPLEX_64_SIZE 4
+#define MAX_COMPLEX_64_TYPE Array4Complex64
+#define MAX_COMPLEX_64_MASK_TYPE VectorU8_4D
+#define MAX_COMPLEX_64_VECTOR_TYPE Vector4D_64
+#define IDEAL_COMPLEX_64_SIZE 2
+#define IDEAL_COMPLEX_64_TYPE Array2Complex64
+#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_2D
+#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector2D_64
+#define MIN_COMPLEX_64_SIZE 1
+#define MIN_COMPLEX_64_TYPE Complex64
+#define MIN_COMPLEX_64_MASK_TYPE VectorU8_1D
+#define MIN_COMPLEX_64_VECTOR_TYPE Vecto1D_64
+
+
+#define MAX_COMPLEX_32_SIZE 8
+#define MAX_COMPLEX_32_TYPE Array8Complex32
+#define MAX_COMPLEX_32_MASK_TYPE VectorU8_8D
+#define MAX_COMPLEX_32_VECTOR_TYPE Vector8D_32
+#define IDEAL_COMPLEX_32_SIZE 4
+#define IDEAL_COMPLEX_32_TYPE Array4Complex32
+#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_4D
+#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector4D_32
+#define MIN_COMPLEX_32_SIZE 2
+#define MIN_COMPLEX_32_TYPE Array2Complex32
+#define MIN_COMPLEX_32_MASK_TYPE VectorU8_2D
+#define MIN_COMPLEX_32_VECTOR_TYPE Vector2D_32
+
+#elif defined(AML_USE_ARRAY_STRICT)
+
+#define MAX_COMPLEX_64_SIZE 2
+#define MAX_COMPLEX_64_TYPE Array2Complex64
+#define MAX_COMPLEX_64_MASK_TYPE VectorU8_2D
+#define MAX_COMPLEX_64_VECTOR_TYPE Vector2D_64
+#define IDEAL_COMPLEX_64_SIZE 2
+#define IDEAL_COMPLEX_64_TYPE Array2Complex64
+#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_2D
+#define IDEAL_COMPLEX_64_VECTOR_TYPE Vector2D_64
+#define MIN_COMPLEX_64_SIZE 2
+#define MIN_COMPLEX_64_TYPE Array2Complex64
+#define MIN_COMPLEX_64_MASK_TYPE VectorU8_2D
+#define MIN_COMPLEX_64_VECTOR_TYPE Vector2D_64
+
+
+#define MAX_COMPLEX_32_SIZE 2
+#define MAX_COMPLEX_32_TYPE Array2Complex32
+#define MAX_COMPLEX_32_MASK_TYPE VectorU8_2D
+#define MAX_COMPLEX_32_VECTOR_TYPE Vector2D_32
+#define IDEAL_COMPLEX_32_SIZE 2
+#define IDEAL_COMPLEX_32_TYPE Array2Complex32
+#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_2D
+#define IDEAL_COMPLEX_32_VECTOR_TYPE Vector2D_32
+#define MIN_COMPLEX_32_SIZE 2
+#define MIN_COMPLEX_32_TYPE Array2Complex32
+#define MIN_COMPLEX_32_MASK_TYPE VectorU8_2D
+#define MIN_COMPLEX_32_VECTOR_TYPE Vector2D_32
+
+
+#elif !defined(COMPLEX_DEFINITIONS) && !defined(USE_CUDA)
+#define COMPLEX_DEFINITIONS
+
+
+#define MAX_COMPLEX_64_SIZE 1
+#define MAX_COMPLEX_64_TYPE Complex64
+#define MAX_COMPLEX_64_MASK_TYPE VectorU8_1D
+#define MAX_COMPLEX_64_VECTOR_TYPE Vecto1D_64
+#define IDEAL_COMPLEX_64_SIZE 1
+#define IDEAL_COMPLEX_64_TYPE Complex64
+#define IDEAL_COMPLEX_64_MASK_TYPE VectorU8_1D
+#define IDEAL_COMPLEX_64_VECTOR_TYPE Vecto1D_64
+#define MIN_COMPLEX_64_SIZE 1
+#define MIN_COMPLEX_64_TYPE Complex64
+#define MIN_COMPLEX_64_MASK_TYPE VectorU8_1D
+#define MIN_COMPLEX_64_VECTOR_TYPE Vecto1D_64
+
+#define MAX_COMPLEX_32_SIZE 1
+#define MAX_COMPLEX_32_TYPE Complex32
+#define MAX_COMPLEX_32_MASK_TYPE VectorU8_1D
+#define MAX_COMPLEX_32_VECTOR_TYPE Vecto1D_32
+#define IDEAL_COMPLEX_32_SIZE 1
+#define IDEAL_COMPLEX_32_TYPE Complex32
+#define IDEAL_COMPLEX_32_MASK_TYPE VectorU8_1D
+#define IDEAL_COMPLEX_32_VECTOR_TYPE Vecto1D_32
+#define MIN_COMPLEX_32_SIZE 1
+#define MIN_COMPLEX_32_TYPE Complex32
+#define MIN_COMPLEX_32_MASK_TYPE VectorU8_1D
+#define MIN_COMPLEX_32_VECTOR_TYPE Vecto1D_32
+
+
+#endif
+
diff --git a/amlcomplex.h b/amlcomplex.h
@@ -536,6 +536,10 @@ class AML_POINTER_NAME(Complex) : public AML_TYPE_NAME(Complex) {
 		return index;
 	}
 
+	explicit AML_FUNCTION operator int() {
+		return index;
+	}
+
 	AML_FUNCTION void operator=(const AML_TYPE_NAME(Complex) newVal) {
 		c.c[0] = newVal.c.c[0];
 		c.c[1] = newVal.c.c[1];

diff --git a/benchmarks/Makefile b/benchmarks/Makefile
@@ -1,11 +1,11 @@
-ARGS = -Ofast -march=native -flto
+ARGS = -Ofast -march=native
 #ARGS =
 
 matrixMul: matrixBenchmark.cpp
 	g++ matrixBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o matrix_bench_gcc
 	clang++ matrixBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o matrix_bench_clang
 	/opt/intel/oneapi/compiler/latest/linux/bin/intel64/icpc matrixBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o matrix_bench_icc -no-multibyte-chars
-	/opt/nvidia/hpc_sdk/Linux_x86_64/2021/compilers/bin/pgc++ matrixBenchmark.cpp -tp=native -std=c++20 -O4 -o matrix_bench_pgc -lpthread -lbenchmark -DDEBUG
+	/opt/nvidia/hpc_sdk/Linux_x86_64/2021/compilers/bin/nvc++ matrixBenchmark.cpp -tp=native -std=c++20 -O4 -o matrix_bench_pgc -lpthread -lbenchmark -DDEBUG -fast
 	./matrix_bench_gcc
 	./matrix_bench_clang
 	./matrix_bench_icc
@@ -16,7 +16,7 @@ complex: complexBenchmark.cpp
 	g++ complexBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o complex_bench_gcc
 	clang++ complexBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o complex_bench_clang
 	/opt/intel/oneapi/compiler/latest/linux/bin/intel64/icpc complexBenchmark.cpp -lpthread -lbenchmark $(ARGS) -std=c++20 -Wall -Wextra -o complex_bench_icc -no-multibyte-chars
-	/opt/nvidia/hpc_sdk/Linux_x86_64/2021/compilers/bin/pgc++ complexBenchmark.cpp -tp=native -std=c++20 -O4 -o complex_bench_pgc -lpthread -lbenchmark -DDEBUG
+	/opt/nvidia/hpc_sdk/Linux_x86_64/2021/compilers/bin/nvc++ complexBenchmark.cpp -tp=native -std=c++20 -O4 -o complex_bench_pgc -lpthread -lbenchmark -DDEBUG
 	./complex_bench_gcc
 	./complex_bench_clang
 	./complex_bench_icc
-Original file line number
+Diff line change
@@ Expand Up / @@ -4,6 +4,7 @@ cmake-build-debug @@
     *.html
     *.js
     *.wasm
+    *.out
     samples/mandelbrot
@@ Expand Down @@