gtt-wc

1 files changed, 149 insertions, 0 deletions

diff --git a/tests/gem_gtt_speed.c b/tests/gem_gtt_speed.c
index 3d726c4e..ba8ed4f4 100644
--- a/tests/gem_gtt_speed.c
+++ b/tests/gem_gtt_speed.c
@@ -28,6 +28,7 @@
 
 #include "igt.h"
 #include "igt_x86.h"
+#include "igt_gt.h"
 #include <unistd.h>
 #include <stdlib.h>
 #include <stdint.h>
@@ -70,6 +71,84 @@ static void streaming_load(void *src, int len)
 
 	*(volatile __m128i *)src = tmp;
 }
+
+static uint8_t stage[4096] __attribute((__aligned__(64)));
+
+static void slow_streaming_store(void *dst, void *src, int len)
+{
+	__m128i *s = src, *d = dst;
+
+	while (len >= 64) {
+		__m128i t0 = _mm_stream_load_si128(s + 0);
+		__m128i t1 = _mm_stream_load_si128(s + 1);
+		__m128i t2 = _mm_stream_load_si128(s + 2);
+		__m128i t3 = _mm_stream_load_si128(s + 3);
+
+
+		_mm_store_si128(d + 0, t0);
+		_mm_store_si128(d + 1, t1);
+		_mm_store_si128(d + 2, t2);
+		_mm_store_si128(d + 3, t3);
+
+		len -= 64;
+		s += 4;
+		d += 4;
+	}
+
+	memcpy(d, s, len);
+}
+
+static void streaming_store(void *dst, void *src, int len)
+{
+	while (len > 64) {
+		__m128i *cache = (__m128i *)stage;
+		int slen = len > 4096 ? 4096 : len;
+
+		while (slen >= 64) {
+			__m128i *s = src;
+
+			__m128i t0 = _mm_stream_load_si128(s + 0);
+			__m128i t1 = _mm_stream_load_si128(s + 1);
+			__m128i t2 = _mm_stream_load_si128(s + 2);
+			__m128i t3 = _mm_stream_load_si128(s + 3);
+
+			_mm_store_si128(cache + 0, t0);
+			_mm_store_si128(cache + 1, t1);
+			_mm_store_si128(cache + 2, t2);
+			_mm_store_si128(cache + 3, t3);
+
+			src += 64;
+			slen -= 64;
+			cache += 4;
+		}
+
+		cache = (__m128i *)stage;
+		slen = len > 4096 ? 4096 : len;
+
+		__asm__ __volatile__("lfence" ::: "memory");
+
+		while (slen >= 64) {
+			__m128i *d = dst;
+
+			__m128i t0 = _mm_stream_load_si128(cache + 0);
+			__m128i t1 = _mm_stream_load_si128(cache + 1);
+			__m128i t2 = _mm_stream_load_si128(cache + 2);
+			__m128i t3 = _mm_stream_load_si128(cache + 3);
+
+			_mm_store_si128(d + 0, t0);
+			_mm_store_si128(d + 1, t1);
+			_mm_store_si128(d + 2, t2);
+			_mm_store_si128(d + 3, t3);
+
+			dst += 64;
+			slen -= 64;
+			cache += 4;
+		}
+
+		len -= len > 4096 ? 4096 : len;
+	}
+}
+
 static inline unsigned x86_64_features(void)
 {
 	return igt_x86_features();
@@ -84,6 +163,14 @@ static void streaming_load(void *src, int len)
 {
 	igt_assert(!"reached");
 }
+static void streaming_store(void *dst, void *src, int len)
+{
+	igt_assert(!"reached");
+}
+static void slow_streaming_store(void *dst, void *src, int len)
+{
+	igt_assert(!"reached");
+}
 #endif
 
 int main(int argc, char **argv)
@@ -112,6 +199,7 @@ int main(int argc, char **argv)
 		igt_info("Detected cpu faatures: %s\n",
 			 igt_x86_features_to_string(cpu, str));
 	}
+	igt_setup_clflush();
 
 	buf = malloc(size);
 	memset(buf, 0, size);
@@ -187,6 +275,45 @@ int main(int argc, char **argv)
 						 size/1024, elapsed(&start, &end, loop));
 				}
 
+				{
+					base = gem_mmap__cpu(fd, handle, 0,
+							     size,
+							     PROT_READ | PROT_WRITE);
+					gettimeofday(&start, NULL);
+					for (loop = 0; loop < 1000; loop++) {
+						ptr = base;
+						x = 0;
+
+						igt_clflush_range(base, size);
+						for (i = 0; i < size/sizeof(*ptr); i++)
+							x += ptr[i];
+
+						/* force overly clever gcc to actually compute x */
+						ptr[0] = x;
+
+					}
+					gettimeofday(&end, NULL);
+					munmap(base, size);
+					igt_info("Time to clflush + read %dk through a cached CPU map:	%7.3fµs\n",
+						 size/1024, elapsed(&start, &end, loop));
+				}
+
+				{
+					base = gem_mmap__cpu(fd, handle, 0,
+							     size,
+							     PROT_READ | PROT_WRITE);
+					gettimeofday(&start, NULL);
+					for (loop = 0; loop < 1000; loop++) {
+						igt_clflush_range(base, size);
+						memcpy(buf, base, size);
+
+					}
+					gettimeofday(&end, NULL);
+					munmap(base, size);
+					igt_info("Time to clflush + store %dk from a cached CPU map:	%7.3fµs\n",
+						 size/1024, elapsed(&start, &end, loop));
+				}
+
 				/* mmap write */
 				gettimeofday(&start, NULL);
 				for (loop = 0; loop < 1000; loop++) {
@@ -339,6 +466,28 @@ int main(int argc, char **argv)
 				}
 				igt_info("Time to stream %dk from a cached WC map:	%7.3fµs\n",
 					 size/1024, elapsed(&start, &end, loop));
+
+				{
+					uint32_t *base = gem_mmap__wc(fd, handle, 0, size, PROT_READ | PROT_WRITE);
+					gettimeofday(&start, NULL);
+					for (loop = 0; loop < 1000; loop++)
+						slow_streaming_store(buf, base, size);
+					gettimeofday(&end, NULL);
+					munmap(base, size);
+				}
+				igt_info("Time to (unoptimised) store %dk from a cached WC map:	%7.3fµs\n",
+					 size/1024, elapsed(&start, &end, loop));
+
+				{
+					uint32_t *base = gem_mmap__wc(fd, handle, 0, size, PROT_READ | PROT_WRITE);
+					gettimeofday(&start, NULL);
+					for (loop = 0; loop < 1000; loop++)
+						streaming_store(buf, base, size);
+					gettimeofday(&end, NULL);
+					munmap(base, size);
+				}
+				igt_info("Time to (optimised) store %dk from a cached WC map:	%7.3fµs\n",
+					 size/1024, elapsed(&start, &end, loop));
 			}
 		}