TellowKrinkle · May 9, 2021 04:47 · May 9, 2021
diff --git a/CPUWakeupSpeedTester.cpp b/CPUWakeupSpeedTester.cpp
@@ -0,0 +1,79 @@
+#include <string.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+extern "C" size_t test(size_t amt);
+
+template <size_t Len, size_t Iters>
+struct Tester {
+	size_t data[Iters][Len];
+
+	void run(useconds_t sleep_us, int sleep_count, size_t iterlen) {
+		for (auto& iter : data) {
+			for (int i = 0; i < sleep_count; i++)
+				usleep(sleep_us);
+			for (size_t i = 0; i < Len; i++) {
+				auto& pt = iter[i];
+				pt = test(iterlen);
+			}
+		}
+	}
+
+	void results() {
+		size_t min = SIZE_MAX;
+		static constexpr int HIST_MAX = 8;
+		static constexpr int HIST_RES = 4; // Bucket size: Min/HIST_RES
+		int histograms[Len][HIST_MAX] = {};
+
+		for (const auto& iter : data) {
+			for (auto pt : iter) {
+				if (min > pt) { min = pt; }
+			}
+		}
+
+		for (const auto& iter : data) {
+			for (size_t i = 0; i < Len; i++) {
+				auto pt = iter[i];
+				auto& hist = histograms[i];
+				auto quartile = (pt - min) / (min / HIST_RES);
+				if (quartile < HIST_MAX) {
+					hist[quartile]++;
+				}
+			}
+		}
+
+		printf("%8zu Baseline\n", min);
+		int digits[HIST_MAX] = {};
+		for (const auto& hist : histograms) {
+			for (int i = 0; i < HIST_MAX; i++) {
+				int v = hist[i];
+				int digitcount = 1;
+				while (v >= 10) {
+					digitcount++;
+					v /= 10;
+				}
+				if (digits[i] < digitcount) { digits[i] = digitcount; }
+			}
+		}
+		for (const auto& hist : histograms) {
+			printf("%2zd: ", (&hist - histograms));
+			for (int i = 0; i < HIST_MAX; i++) {
+				char fmt[10];
+				sprintf(fmt, "%%%dd ", digits[i]);
+				printf(fmt, hist[i]);
+			}
+			printf("\n");
+		}
+	}
+};
+
+int main(int argc, const char * argv[]) {
+	Tester<32, 1000> tester;
+	tester.run(4000, 1, 100000);
+	tester.results();
+	tester.run(1000, 1, 100000);
+	tester.results();
+	return 0;
+}
diff --git a/CPUWakeupSpeedTester.s b/CPUWakeupSpeedTester.s
@@ -0,0 +1,20 @@
+.intel_syntax
+.globl _test
+_test:
+	rdtsc
+	shl rdx, 32
+	or rax, rdx
+	mov rcx, rax
+loop:
+	add rdx, rdx
+	add rdx, rdx
+	add rdx, rdx
+	add rdx, rdx
+	dec rdi
+	jne loop
+
+	rdtsc
+	shl rdx, 32
+	or rax, rdx
+	sub rax, rcx
+	ret
No results found