From 0a03261329addc6ea028f51b2ac064fbc5c26a00 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Branimir=20Karad=C5=BEi=C4=87?= <branimirkaradzic@gmail.com>
Date: Sun, 26 Feb 2017 18:30:16 -0800
Subject: [PATCH] Added SIMD rsqrt bench.

---
 tests/handle_bench.cpp |   9 +++-
 tests/simd_bench.cpp   | 115 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 123 insertions(+), 1 deletion(-)
 create mode 100644 tests/simd_bench.cpp

diff --git a/tests/handle_bench.cpp b/tests/handle_bench.cpp
index 6169bc4..e8d8bee 100644
--- a/tests/handle_bench.cpp
+++ b/tests/handle_bench.cpp
@@ -1,4 +1,8 @@
-#include <bx/bx.h>
+/*
+ * Copyright 2010-2017 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
+ */
+
 #include <bx/timer.h>
 #include <bx/handlealloc.h>
 #include <bx/maputil.h>
@@ -99,5 +103,8 @@ int main()
 		printf("HandleHashMap: %15f\n", double(elapsed) );
 	}
 
+	extern void simd_bench();
+	simd_bench();
+
 	return EXIT_SUCCESS;
 }
diff --git a/tests/simd_bench.cpp b/tests/simd_bench.cpp
new file mode 100644
index 0000000..205fef2
--- /dev/null
+++ b/tests/simd_bench.cpp
@@ -0,0 +1,115 @@
+/*
+ * Copyright 2010-2017 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx#license-bsd-2-clause
+ */
+
+#include <bx/simd_t.h>
+#include <bx/rng.h>
+#include <bx/crtimpl.h>
+#include <bx/timer.h>
+
+#include <stdio.h>
+
+static void flushCache()
+{
+	static uint32_t length = 1 << 26;
+	static uint8_t* input  = new uint8_t[length];
+	static uint8_t* output = new uint8_t[length];
+	bx::memCopy(output, input, length);
+}
+
+typedef bx::simd128_t (*SimdRsqrtFn)(bx::simd128_t _a);
+
+template<SimdRsqrtFn simdRsqrtFn>
+void simd_rsqrt_bench(bx::simd128_t* _dst, bx::simd128_t* _src, uint32_t _numVertices)
+{
+	for (uint32_t ii = 0, num = _numVertices/4; ii < num; ++ii)
+	{
+		bx::simd128_t* ptr = &_src[ii*4];
+		bx::simd128_t tmp0 = bx::simd_ld(ptr + 0);
+		bx::simd128_t tmp1 = bx::simd_ld(ptr + 1);
+		bx::simd128_t tmp2 = bx::simd_ld(ptr + 2);
+		bx::simd128_t tmp3 = bx::simd_ld(ptr + 3);
+		bx::simd128_t rsqrt0 = simdRsqrtFn(tmp0);
+		bx::simd128_t rsqrt1 = simdRsqrtFn(tmp1);
+		bx::simd128_t rsqrt2 = simdRsqrtFn(tmp2);
+		bx::simd128_t rsqrt3 = simdRsqrtFn(tmp3);
+
+		ptr = &_dst[ii*4];
+		bx::simd_st(ptr + 0, rsqrt0);
+		bx::simd_st(ptr + 1, rsqrt1);
+		bx::simd_st(ptr + 2, rsqrt2);
+		bx::simd_st(ptr + 3, rsqrt3);
+	}
+}
+
+void simd_bench()
+{
+	bx::CrtAllocator allocator;
+	bx::RngMwc rng;
+
+	const uint32_t numVertices = 1024*1024;
+
+	uint8_t* data = (uint8_t*)BX_ALIGNED_ALLOC(&allocator, 2*numVertices*sizeof(bx::simd128_t), 16);
+	bx::simd128_t* src = (bx::simd128_t*)data;
+	bx::simd128_t* dst = &src[numVertices];
+
+	for (uint32_t ii = 0; ii < numVertices; ++ii)
+	{
+		float* ptr = (float*)&src[ii];
+		randUnitSphere(ptr, &rng);
+		ptr[3] = 1.0f;
+	}
+
+	const uint32_t numIterations = 10;
+
+	{
+		int64_t elapsed = 0;
+		for (uint32_t test = 0; test < numIterations; ++test)
+		{
+			flushCache();
+			elapsed += -bx::getHPCounter();
+			simd_rsqrt_bench<bx::simd_rsqrt_est>(dst, src, numVertices);
+			elapsed += bx::getHPCounter();
+		}
+		printf("    simd_rsqrt_est: %15f\n", double(elapsed) );
+	}
+
+	{
+		int64_t elapsed = 0;
+		for (uint32_t test = 0; test < numIterations; ++test)
+		{
+			flushCache();
+			elapsed += -bx::getHPCounter();
+			simd_rsqrt_bench<bx::simd_rsqrt_nr>(dst, src, numVertices);
+			elapsed += bx::getHPCounter();
+		}
+		printf("     simd_rsqrt_nr: %15f\n", double(elapsed) );
+	}
+
+	{
+		int64_t elapsed = 0;
+		for (uint32_t test = 0; test < numIterations; ++test)
+		{
+			flushCache();
+			elapsed += -bx::getHPCounter();
+			simd_rsqrt_bench<bx::simd_rsqrt_carmack>(dst, src, numVertices);
+			elapsed += bx::getHPCounter();
+		}
+		printf("simd_rsqrt_carmack: %15f\n", double(elapsed) );
+	}
+
+	{
+		int64_t elapsed = 0;
+		for (uint32_t test = 0; test < numIterations; ++test)
+		{
+			flushCache();
+			elapsed += -bx::getHPCounter();
+			simd_rsqrt_bench<bx::simd_rsqrt>(dst, src, numVertices);
+			elapsed += bx::getHPCounter();
+		}
+		printf("        simd_rsqrt: %15f\n", double(elapsed) );
+	}
+
+	BX_ALIGNED_FREE(&allocator, data, 16);
+}