#include <sys/types.h>
#include <sys/mman.h>
-typedef struct {
- unsigned char x;
- unsigned char y;
- unsigned char z;
- unsigned char w;
-} uchar4;
-
-typedef struct {
- float x;
- float y;
- float z;
- float w;
-} float4;
-
-
-#define MAX_RADIUS 25
-#define WIDTH 512
-#define HEIGHT 512
-
-uchar4 bufIn[WIDTH * HEIGHT];
-uchar4 bufOut[WIDTH * HEIGHT];
-uchar4 bufTmp[WIDTH * HEIGHT];
-
-static float gaussian[MAX_RADIUS * 2 + 1];
-static int height = 512;
-static int width = 512;
-static int radius = MAX_RADIUS;
-
-
-
-static void horiz(uchar4 *output, const uchar4 *inputBuf, uint32_t x, uint32_t y) {
- const uchar4 *input = inputBuf + (y * width);
- float4 blurredPixel = {0,0,0,0};
- float4 currentPixel;
-
- for(int r = -radius; r <= radius; r ++) {
- // Stepping left and right away from the pixel
- int validW = x + r;
- // Clamp to zero and width max() isn't exposed for ints yet
- if(validW < 0) {
- validW = 0;
- }
- if(validW > WIDTH - 1) {
- validW = WIDTH - 1;
- }
- //int validW = rsClamp(w + r, 0, width - 1);
-
- float weight = gaussian[r + radius];
- currentPixel.x = (float)(input[validW].x);
- currentPixel.y = (float)(input[validW].y);
- currentPixel.z = (float)(input[validW].z);
- //currentPixel.w = (float)(input->a);
-
- blurredPixel.x += currentPixel.x * weight;
- blurredPixel.y += currentPixel.y * weight;
- blurredPixel.z += currentPixel.z * weight;
- }
-
- output->x = (uint8_t)blurredPixel.x;
- output->y = (uint8_t)blurredPixel.y;
- output->z = (uint8_t)blurredPixel.z;
-}
-
-
-static void vert(uchar4 *output, const uchar4 *inputBuf, uint32_t x, uint32_t y) {
- const uchar4 *input = inputBuf + x;
-
- float4 blurredPixel = {0,0,0,0};
-
- float4 currentPixel;
- for(int r = -radius; r <= radius; r ++) {
- int validH = y + r;
- // Clamp to zero and width
- if(validH < 0) {
- validH = 0;
- }
- if(validH > HEIGHT - 1) {
- validH = HEIGHT - 1;
- }
-
- const uchar4 *i = input + validH * WIDTH;
-
- float weight = gaussian[r + radius];
-
- currentPixel.x = (float)(i->x);
- currentPixel.y = (float)(i->y);
- currentPixel.z = (float)(i->z);
-
- blurredPixel.x += currentPixel.x * weight;
- blurredPixel.y += currentPixel.y * weight;
- blurredPixel.z += currentPixel.z * weight;
- }
-
- //output->xyz = convert_uchar3(blurredPixel.xyz);
- output->x = (uint8_t)blurredPixel.x;
- output->y = (uint8_t)blurredPixel.y;
- output->z = (uint8_t)blurredPixel.z;
-}
+#ifdef __ARM_NEON__
+#include <arm_neon.h>
+#endif
typedef long long nsecs_t;
+static nsecs_t gTime;
+float data_f[1024 * 128];
static nsecs_t system_time()
{
return nsecs_t(t.tv_sec)*1000000000LL + t.tv_nsec;
}
-int fp_test(int argc, char** argv) {
- for (int ct=0; ct < (sizeof(gaussian)/4); ct++) {
- gaussian[ct] = 1.f;
- }
- memset(bufIn, 0, sizeof(bufIn));
+static void startTime()
+{
+ gTime = system_time();
+}
+
+static void endTime(const char *str, double ops)
+{
+ nsecs_t t = system_time() - gTime;
+ double ds = ((double)t) / 1e9;
+ printf("Test: %s, %f Mops\n", str, ops / ds / 1e6);
+}
- nsecs_t t1 = system_time();
- for (int y = 0; y < HEIGHT; y++) {
- for (int x = 0; x < WIDTH; x++) {
- horiz(&bufTmp[x + y * WIDTH], bufIn, x, y);
+static void test_mad() {
+ for(int i=0; i<1020; i++) {
+ data_f[i] = i;
+ }
+
+ startTime();
+
+ float total = 0;
+ // Do ~1 billion ops
+ for (int ct=0; ct < (1000 * (1000 / 20)); ct++) {
+ for (int i=0; i < 1000; i++) {
+ data_f[i] = (data_f[i] * 0.02f +
+ data_f[i+1] * 0.04f +
+ data_f[i+2] * 0.05f +
+ data_f[i+3] * 0.1f +
+ data_f[i+4] * 0.2f +
+ data_f[i+5] * 0.2f +
+ data_f[i+6] * 0.1f +
+ data_f[i+7] * 0.05f +
+ data_f[i+8] * 0.04f +
+ data_f[i+9] * 0.02f + 1.f);
}
}
- for (int y = 0; y < HEIGHT; y++) {
- for (int x = 0; x < WIDTH; x++) {
- vert(&bufOut[x + y * WIDTH], bufTmp, x, y);
+ endTime("scalar mad", 1e9);
+}
+
+
+#ifdef __ARM_NEON__
+
+static void test_fma() {
+ for(int i=0; i<1020 * 4; i++) {
+ data_f[i] = i;
+ }
+ float32x4_t c0_02 = vdupq_n_f32(0.02f);
+ float32x4_t c0_04 = vdupq_n_f32(0.04f);
+ float32x4_t c0_05 = vdupq_n_f32(0.05f);
+ float32x4_t c0_10 = vdupq_n_f32(0.1f);
+ float32x4_t c0_20 = vdupq_n_f32(0.2f);
+ float32x4_t c1_00 = vdupq_n_f32(1.0f);
+
+ startTime();
+
+ float total = 0;
+ // Do ~1 billion ops
+ for (int ct=0; ct < (1000 * (1000 / 80)); ct++) {
+ for (int i=0; i < 1000; i++) {
+ float32x4_t t;
+ t = vmulq_f32(vld1q_f32((float32_t *)&data_f[i]), c0_02);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+4]), c0_04);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+8]), c0_05);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+12]), c0_10);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+16]), c0_20);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+20]), c0_20);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+24]), c0_10);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+28]), c0_05);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+32]), c0_04);
+ t = vmlaq_f32(t, vld1q_f32((float32_t *)&data_f[i+36]), c0_02);
+ t = vaddq_f32(t, c1_00);
+ vst1q_f32((float32_t *)&data_f[i], t);
}
}
- nsecs_t t2 = system_time();
+ endTime("neon fma", 1e9);
+}
+#endif
+
+int fp_test(int argc, char** argv) {
+ test_mad();
- printf("FP Test time %i ms\n", (int)((t2 - t1) / 1000000) );
+#ifdef __ARM_NEON__
+ test_fma();
+#endif
return 0;
}
+
+
+
OSDN Git Service
