summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--Android.mk16
-rw-r--r--blit.c169
-rw-r--r--blit.h5
-rwxr-xr-xrun.sh14
-rw-r--r--test_blit.c219
-rw-r--r--timer.c38
-rw-r--r--timer.h10
7 files changed, 471 insertions, 0 deletions
diff --git a/Android.mk b/Android.mk
new file mode 100644
index 0000000..f897ede
--- /dev/null
+++ b/Android.mk
@@ -0,0 +1,16 @@
+LOCAL_PATH := $(call my-dir)
+
+######################################################
+### blit_test ###
+######################################################
+include $(CLEAR_VARS)
+LOCAL_SRC_FILES := blit.c test_blit.c timer.c
+
+LOCAL_MODULE_TAGS := debug
+LOCAL_MODULE := blit_test
+
+LOCAL_MULTILIB := both
+LOCAL_MODULE_STEM_64 := $(LOCAL_MODULE)64
+LOCAL_MODULE_STEM_32 := $(LOCAL_MODULE)32
+
+include $(BUILD_EXECUTABLE)
diff --git a/blit.c b/blit.c
new file mode 100644
index 0000000..050dc82
--- /dev/null
+++ b/blit.c
@@ -0,0 +1,169 @@
+#include<stdio.h>
+#include<stddef.h>
+#include<stdint.h>
+#include<arm_neon.h>
+
+#define SK_B16_BITS 5
+#define SK_G16_BITS 6
+#define SK_R16_BITS 5
+
+#define SK_R16_SHIFT (SK_B16_BITS + SK_G16_BITS)
+#define SK_G16_SHIFT (SK_B16_BITS)
+
+#define SK_G16_MASK ((1 << SK_G16_BITS) - 1)
+#define SK_B16_MASK ((1 << SK_B16_BITS) - 1)
+
+#define SK_G16_MASK_IN_PLACE (SK_G16_MASK << SK_G16_SHIFT)
+
+
+static inline uint32_t SkExpand_rgb_16(unsigned c) {
+ return ((c & SK_G16_MASK_IN_PLACE) << 16) | (c & ~SK_G16_MASK_IN_PLACE);
+}
+
+static inline unsigned SkCompact_rgb_16(uint32_t c) {
+ return ((c >> 16) & SK_G16_MASK_IN_PLACE) | (c & ~SK_G16_MASK_IN_PLACE);
+}
+
+void SkRGB16_Opaque_Blitter_blitV_c(uint16_t* device, size_t deviceRB, int height, uint8_t alpha) {
+ unsigned scale5 = alpha >> 3;
+ uint32_t src32 = 256 * scale5;
+ scale5 = 32 - scale5;
+ do {
+ uint32_t dst32 = SkExpand_rgb_16(*device) * scale5;
+ *device = SkCompact_rgb_16((src32 + dst32) >> 5);
+ device = (uint16_t*)((char*)device + deviceRB);
+ } while (--height != 0);
+}
+
+#define LOAD_LANE_16(reg, n) \
+ reg = vld1q_lane_u16(device, reg, n); \
+ device = (uint16_t*)((char*)device + deviceRB);
+
+#define STORE_LANE_16(reg, n) \
+ vst1_lane_u16(dst, reg, n); \
+ dst = (uint16_t*)((char*)dst + deviceRB);
+
+void SkRGB16_Opaque_Blitter_blitV_neon(uint16_t* device, size_t deviceRB, int height, uint8_t alpha)
+{
+ unsigned scale = alpha >> 3;
+ uint32_t src32 = 256 * scale;
+ scale = 32 - scale;
+ if (height >= 8)
+ {
+ uint16_t* dst = device;
+
+ // prepare constants
+ uint16x8_t vdev = vdupq_n_u16(0);
+ uint16x8_t vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE);
+ uint16x8_t vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE);
+ uint32x4_t vsrc32 = vdupq_n_u32(src32);
+ uint32x4_t vscale5 = vdupq_n_u32((uint32_t)scale);
+
+ while (height >= 8){
+ LOAD_LANE_16(vdev, 0)
+ LOAD_LANE_16(vdev, 1)
+ LOAD_LANE_16(vdev, 2)
+ LOAD_LANE_16(vdev, 3)
+ LOAD_LANE_16(vdev, 4)
+ LOAD_LANE_16(vdev, 5)
+ LOAD_LANE_16(vdev, 6)
+ LOAD_LANE_16(vdev, 7)
+
+ // Expand_rgb_16
+ uint16x8x2_t vdst = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16));
+ uint32x4_t vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst.val[0]), vscale5);
+ uint32x4_t vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst.val[1]), vscale5);
+
+ // Compact_rgb_16
+ vdst32_lo = vaddq_u32(vdst32_lo, vsrc32);
+ vdst32_hi = vaddq_u32(vdst32_hi, vsrc32);
+ vdst32_lo = vshrq_n_u32(vdst32_lo, 5);
+ vdst32_hi = vshrq_n_u32(vdst32_hi, 5);
+
+ uint16x4_t vtmp_lo = vmovn_u32(vdst32_lo) & vget_low_u16(vmaskq_ng16);
+ uint16x4_t vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vget_low_u16(vmaskq_g16);
+ uint16x4_t vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi);
+ vtmp_lo = vmovn_u32(vdst32_hi) & vget_low_u16(vmaskq_ng16);
+ vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vget_low_u16(vmaskq_g16);
+ uint16x4_t vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi);
+
+ STORE_LANE_16(vdst16_lo, 0)
+ STORE_LANE_16(vdst16_lo, 1)
+ STORE_LANE_16(vdst16_lo, 2)
+ STORE_LANE_16(vdst16_lo, 3)
+ STORE_LANE_16(vdst16_hi, 0)
+ STORE_LANE_16(vdst16_hi, 1)
+ STORE_LANE_16(vdst16_hi, 2)
+ STORE_LANE_16(vdst16_hi, 3)
+ height -= 8;
+ }
+ }
+ while (height != 0){
+ uint32_t dst32 = SkExpand_rgb_16(*device) * scale;
+ *device = SkCompact_rgb_16((src32 + dst32) >> 5);
+ device = (uint16_t*)((char*)device + deviceRB);
+ height--;
+ }
+}
+#undef LOAD_LANE_16
+#undef STORE_LANE_16
+
+void SkRGB16_Opaque_Blitter_blitH_c(uint16_t* device, int height, uint8_t alpha) {
+ unsigned scale5 = alpha >> 3;
+ uint32_t src32 = 256 * scale5;
+ scale5 = 32 - scale5;
+ do {
+ uint32_t dst32 = SkExpand_rgb_16(*device) * scale5;
+ *device++ = SkCompact_rgb_16((src32 + dst32) >> 5);
+ } while (--height != 0);
+}
+
+void SkRGB16_Opaque_Blitter_blitH_neon(uint16_t* device, int height, uint8_t alpha)
+{
+ unsigned scale = alpha >> 3;
+ uint32_t src32 = 256 * scale;
+ scale = 32 - scale;
+ if (height >= 8)
+ {
+ uint16_t* dst = device;
+
+ // prepare constants
+ uint16x8_t vdev = vdupq_n_u16(0);
+ uint16x8_t vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE);
+ uint16x8_t vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE);
+ uint32x4_t vsrc32 = vdupq_n_u32(src32);
+ uint32x4_t vscale5 = vdupq_n_u32((uint32_t)scale);
+
+ while (height >= 8){
+ vdev = vld1q_u16(device);
+
+ // Expand_rgb_16
+ uint16x8x2_t vdst = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16));
+ uint32x4_t vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst.val[0]), vscale5);
+ uint32x4_t vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst.val[1]), vscale5);
+
+ // Compact_rgb_16
+ vdst32_lo = vaddq_u32(vdst32_lo, vsrc32);
+ vdst32_hi = vaddq_u32(vdst32_hi, vsrc32);
+ vdst32_lo = vshrq_n_u32(vdst32_lo, 5);
+ vdst32_hi = vshrq_n_u32(vdst32_hi, 5);
+
+ uint16x4_t vtmp_lo = vmovn_u32(vdst32_lo) & vget_low_u16(vmaskq_ng16);
+ uint16x4_t vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vget_low_u16(vmaskq_g16);
+ uint16x4_t vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi);
+ vtmp_lo = vmovn_u32(vdst32_hi) & vget_low_u16(vmaskq_ng16);
+ vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vget_low_u16(vmaskq_g16);
+ uint16x4_t vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi);
+
+ vst1q_u16(device, vcombine_u16(vdst16_lo, vdst16_hi));
+ device += 8;
+ height -= 8;
+ }
+ }
+ while (height != 0){
+ uint32_t dst32 = SkExpand_rgb_16(*device) * scale;
+ *device++ = SkCompact_rgb_16((src32 + dst32) >> 5);
+ height--;
+ }
+}
+
diff --git a/blit.h b/blit.h
new file mode 100644
index 0000000..8e17699
--- /dev/null
+++ b/blit.h
@@ -0,0 +1,5 @@
+
+extern void SkRGB16_Opaque_Blitter_blitV_c(uint16_t* device, size_t deviceRB, int height, uint8_t alpha);
+extern void SkRGB16_Opaque_Blitter_blitV_neon(uint16_t* device, size_t deviceRB, int height, uint8_t alpha);
+extern void SkRGB16_Opaque_Blitter_blitH_c(uint16_t* device, int height, uint8_t alpha);
+extern void SkRGB16_Opaque_Blitter_blitH_neon(uint16_t* device, int height, uint8_t alpha);
diff --git a/run.sh b/run.sh
new file mode 100755
index 0000000..e119462
--- /dev/null
+++ b/run.sh
@@ -0,0 +1,14 @@
+#/bin/bash
+#gcc_path="/home/yang01/software/google/ndk/android-ndk-r10e/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/aarch64-linux-android-gcc"
+gcc_path="/data/armv8/lollipop/prebuilts/gcc/linux-x86/aarch64/aarch64-linux-android-4.9/bin/aarch64-linux-android-gcc"
+#sysroot_path="/home/yang01/software/google/ndk/android-ndk-r10e//platforms/android-21/arch-arm64/"
+sysroot_path="/data/armv8/lollipop/prebuilts/ndk/9/platforms/android-21/arch-arm64/"
+
+${gcc_path} -O2 --sysroot=${sysroot_path} -pie -o blit.c.o -c blit.c
+#${gcc_path} -O2 --sysroot=${sysroot_path} -pie -o blit_neon.c.o -c blit_neon.c
+${gcc_path} -O2 --sysroot=${sysroot_path} -pie -o timer.c.o -c timer.c
+${gcc_path} -O2 --sysroot=${sysroot_path} -pie -o test_blit.c.o -c test_blit.c
+${gcc_path} -O2 --sysroot=${sysroot_path} -pie blit.c.o timer.c.o test_blit.c.o -o test_blit
+
+adb push test_blit /data/local/tmp/
+adb shell /data/local/tmp/test_blit
diff --git a/test_blit.c b/test_blit.c
new file mode 100644
index 0000000..cb63d3e
--- /dev/null
+++ b/test_blit.c
@@ -0,0 +1,219 @@
+#include<stdio.h>
+#include<stdlib.h>
+#include<stddef.h>
+#include<stdint.h>
+#include<arm_neon.h>
+
+#include "timer.h"
+#include "blit.h"
+
+#define TEST_LENGTH_SAMPLES (10000)
+
+#define TEST_COUNT 500000
+
+static uint16_t testInput_u16[TEST_LENGTH_SAMPLES];
+
+//input and output
+static uint16_t* in0 = NULL;
+static uint16_t* in1 = NULL;
+static uint16_t* in2 = NULL;
+static uint16_t* in3 = NULL;
+
+static double time0 = 0;
+static double time1 = 0;
+
+void test_blitV_conformance()
+{
+
+ int i = 0;
+ int height = 0;
+ unsigned int sse0, sse1;
+
+
+ /* init input memory */
+ in0 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in0));
+ in1 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in1));
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ testInput_u16[i] = (uint16_t) i; //(rand()&0xffff) ;
+ }
+ fprintf (stdout, "-----------------------------------------------------------\n");
+ fprintf (stdout, "-------------------------------------------CONF TEST\n");
+ fprintf (stdout, "-----------------------------------------------------------\n");
+
+ for (height = 1; height <= 128; height += 1)
+ {
+ printf ("height %d\n", height);
+ memcpy (in0, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ memcpy (in1, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in1));
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ if (in0[i]!=in1[i])
+ printf ("pre i %6d: %10d %10d \n", i, in0[i], in1[i]);
+ }
+ SkRGB16_Opaque_Blitter_blitV_c(in0, 4, height, 25);
+ SkRGB16_Opaque_Blitter_blitV_neon(in1, 4, height, 25);
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ if (in0[i]!=in1[i])
+ printf ("pos i %6d: %10d %10d \n", i, in0[i], in1[i]);
+ }
+ }
+ free (in0);
+ free (in1);
+}
+
+
+void test_blitV_performance()
+{
+
+ int i = 0;
+ int config[] = {1, 8, 18, 32, 76, 85, 120, 128, 512};
+ int num = 0;
+ int height = 0;
+ int loop = 0;
+ unsigned int sse0, sse1;
+
+ /* init input memory */
+ in0 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in0));
+ in1 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in1));
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ testInput_u16[i] = (uint16_t) i; //(rand()&0xffff) ;
+ }
+ fprintf (stdout, "-----------------------------------------------------------\n");
+ fprintf (stdout, "-------------------------------------blitV---------PERF TEST\n");
+ fprintf (stdout, "-----------------------------------------------------------\n");
+
+ for (num = 0; num < sizeof(config)/sizeof(*config); num++)
+ {
+ loop = TEST_COUNT;
+ height = config[num];
+
+ /* C */
+ memcpy (in0, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ reset_timer();
+ for (i = 0; i < loop ; i++)
+ SkRGB16_Opaque_Blitter_blitV_c(in0, 4, height, 25);
+ time0 = get_time();
+
+ /* NEON */
+ memcpy (in1, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ reset_timer();
+ for (i = 0; i < loop ; i++)
+ SkRGB16_Opaque_Blitter_blitV_neon(in1, 4, height, 25);
+ time1 = get_time();
+
+ printf ("height %5d: %20lf %20lf\n", height, time0, time1);
+ printf ("ratio: %lf \n", time0/time1);
+
+ }
+
+ free (in0);
+ free (in1);
+}
+void test_blitH_conformance()
+{
+
+ int i = 0;
+ int height = 0;
+ unsigned int sse0, sse1;
+
+
+ /* init input memory */
+ in0 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in0));
+ in1 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in1));
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ testInput_u16[i] = (uint16_t) i; //(rand()&0xffff) ;
+ }
+ fprintf (stdout, "-----------------------------------------------------------\n");
+ fprintf (stdout, "-------------------------------------------CONF TEST\n");
+ fprintf (stdout, "-----------------------------------------------------------\n");
+
+ for (height = 1; height <= 128; height += 1)
+ {
+ printf ("height %d\n", height);
+ memcpy (in0, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ memcpy (in1, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in1));
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ if (in0[i]!=in1[i])
+ printf ("pre i %6d: %10d %10d \n", i, in0[i], in1[i]);
+ }
+ SkRGB16_Opaque_Blitter_blitH_c(in0, height, 25);
+ SkRGB16_Opaque_Blitter_blitH_neon(in1, height, 25);
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ if (in0[i]!=in1[i])
+ printf ("pos i %6d: %10d %10d \n", i, in0[i], in1[i]);
+ }
+ }
+ free (in0);
+ free (in1);
+}
+
+
+void test_blitH_performance()
+{
+
+ int i = 0;
+ int config[] = {1, 8, 18, 32, 76, 85, 120, 128, 512};
+ int num = 0;
+ int height = 0;
+ int loop = 0;
+ unsigned int sse0, sse1;
+
+ /* init input memory */
+ in0 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in0));
+ in1 = (uint16_t*) malloc ( (TEST_LENGTH_SAMPLES) * sizeof (*in1));
+
+ for (i = 0; i < TEST_LENGTH_SAMPLES; i++)
+ {
+ testInput_u16[i] = (uint16_t) i; //(rand()&0xffff) ;
+ }
+ fprintf (stdout, "-----------------------------------------------------------\n");
+ fprintf (stdout, "--------------------------------------blitH--------PERF TEST\n");
+ fprintf (stdout, "-----------------------------------------------------------\n");
+
+ for (num = 0; num < sizeof(config)/sizeof(*config); num++)
+ {
+ loop = TEST_COUNT;
+ height = config[num];
+
+ /* C */
+ memcpy (in0, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ reset_timer();
+ for (i = 0; i < loop ; i++)
+ SkRGB16_Opaque_Blitter_blitH_c(in0, height, 25);
+ time0 = get_time();
+
+ /* NEON */
+ memcpy (in1, testInput_u16, TEST_LENGTH_SAMPLES * sizeof (*in0));
+ reset_timer();
+ for (i = 0; i < loop ; i++)
+ SkRGB16_Opaque_Blitter_blitH_neon(in1, height, 25);
+ time1 = get_time();
+
+ printf ("height %5d: %20lf %20lf\n", height, time0, time1);
+ printf ("ratio: %lf \n", time0/time1);
+
+ }
+
+ free (in0);
+ free (in1);
+}
+
+
+void main (void)
+{
+ test_blitV_conformance(); // run tests
+ test_blitV_performance(); // run tests
+ test_blitH_conformance(); // run tests
+ test_blitH_performance(); // run tests
+}
diff --git a/timer.c b/timer.c
new file mode 100644
index 0000000..6717a29
--- /dev/null
+++ b/timer.c
@@ -0,0 +1,38 @@
+#include <sys/resource.h>
+//#include <sys/time.h>
+#include <time.h>
+#include "timer.h"
+
+#define TIMER_TYPE CLOCK_PROCESS_CPUTIME_ID
+struct timespec start_time;
+
+void reset_timer()
+{
+ clock_gettime(TIMER_TYPE, &start_time);
+}
+
+double get_time()
+{
+ struct timespec t;
+ clock_gettime(TIMER_TYPE, &t);
+ return ((double)(t.tv_sec - start_time.tv_sec) + (double)(t.tv_nsec - start_time.tv_nsec) * 1.e-9);
+}
+
+#if 0
+void GetUserTime(struct timeval* time) {
+ struct rusage usage;
+ getrusage(RUSAGE_SELF, &usage);
+ memcpy(time, &usage.ru_utime, sizeof(*time));
+}
+
+double TimeDifference(const struct timeval * start,
+ const struct timeval * end) {
+ double start_time;
+ double end_time;
+ start_time = start->tv_sec + start->tv_usec * 1e-6;
+ end_time = end->tv_sec + end->tv_usec * 1e-6;
+
+ return end_time - start_time;
+}
+
+#endif
diff --git a/timer.h b/timer.h
new file mode 100644
index 0000000..44f0546
--- /dev/null
+++ b/timer.h
@@ -0,0 +1,10 @@
+#ifndef TIMER_H
+#define TIMER_H
+
+//#define TIMER_TYPE CLOCK_REALTIME
+#define TIMER_TYPE CLOCK_PROCESS_CPUTIME_ID
+
+void reset_timer();
+double get_time();
+
+#endif
generated by cgit v1.2.3 (git 2.25.1) at 2024-04-28 12:37:23 +0000