Skip to content

Commit

Permalink
optmize resize bilinear
Browse files Browse the repository at this point in the history
  • Loading branch information
@nihui
nihui committed Dec 5, 2023
1 parent eea3fc9 commit dce2f45
Show file tree
Hide file tree
Showing 3 changed files with 4,450 additions and 4,302 deletions.
250 changes: 246 additions & 4 deletions src/mat_pixel_drawing.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
#include "mat.h"

#include <ctype.h>
#include <limits.h>

#include "platform.h"

Expand Down Expand Up @@ -1249,6 +1250,247 @@ void draw_text_c4(unsigned char* pixels, int w, int h, const char* text, int x,
return draw_text_c4(pixels, w, h, w * 4, text, x, y, fontpixelsize, color);
}

void resize_bilinear_font(const unsigned char* font_bitmap, unsigned char* resized_font_bitmap, int fontpixelsize)
{
const int INTER_RESIZE_COEF_BITS = 11;
const int INTER_RESIZE_COEF_SCALE = 1 << INTER_RESIZE_COEF_BITS;

const int srcw = 20;
const int srch = 40;
const int w = fontpixelsize;
const int h = fontpixelsize * 2;

double scale = (double)srcw / w;

int* buf = new int[w + h + w + h];

int* xofs = buf; //new int[w];
int* yofs = buf + w; //new int[h];

short* ialpha = (short*)(buf + w + h); //new short[w * 2];
short* ibeta = (short*)(buf + w + h + w); //new short[h * 2];

float fx;
float fy;
int sx;
int sy;

#define SATURATE_CAST_SHORT(X) (short)::std::min(::std::max((int)(X + (X >= 0.f ? 0.5f : -0.5f)), SHRT_MIN), SHRT_MAX);

for (int dx = 0; dx < w; dx++)
{
fx = (float)((dx + 0.5) * scale - 0.5);
sx = static_cast<int>(floor(fx));
fx -= sx;

if (sx < 0)
{
sx = 0;
fx = 0.f;
}
if (sx >= srcw - 1)
{
sx = srcw - 2;
fx = 1.f;
}

xofs[dx] = sx;

float a0 = (1.f - fx) * INTER_RESIZE_COEF_SCALE;
float a1 = fx * INTER_RESIZE_COEF_SCALE;

ialpha[dx * 2] = SATURATE_CAST_SHORT(a0);
ialpha[dx * 2 + 1] = SATURATE_CAST_SHORT(a1);
}

for (int dy = 0; dy < h; dy++)
{
fy = (float)((dy + 0.5) * scale - 0.5);
sy = static_cast<int>(floor(fy));
fy -= sy;

if (sy < 0)
{
sy = 0;
fy = 0.f;
}
if (sy >= srch - 1)
{
sy = srch - 2;
fy = 1.f;
}

yofs[dy] = sy;

float b0 = (1.f - fy) * INTER_RESIZE_COEF_SCALE;
float b1 = fy * INTER_RESIZE_COEF_SCALE;

ibeta[dy * 2] = SATURATE_CAST_SHORT(b0);
ibeta[dy * 2 + 1] = SATURATE_CAST_SHORT(b1);
}

#undef SATURATE_CAST_SHORT

// loop body
Mat rowsbuf0(w, (size_t)2u);
Mat rowsbuf1(w, (size_t)2u);
short* rows0 = (short*)rowsbuf0;
short* rows1 = (short*)rowsbuf1;

int prev_sy1 = -2;

for (int dy = 0; dy < h; dy++)
{
sy = yofs[dy];

if (sy == prev_sy1)
{
// reuse all rows
}
else if (sy == prev_sy1 + 1)
{
// hresize one row
short* rows0_old = rows0;
rows0 = rows1;
rows1 = rows0_old;
const unsigned char* S1 = font_bitmap + 10 * (sy + 1);

const short* ialphap = ialpha;
short* rows1p = rows1;
for (int dx = 0; dx < w; dx++)
{
sx = xofs[dx];
short a0 = ialphap[0];
short a1 = ialphap[1];

// const unsigned char* S1p = S1 + sx;
unsigned char S1p0 = sx % 2 == 0 ? S1[sx/2] & 0x0f : (S1[sx/2] & 0xf0) >> 4;
unsigned char S1p1 = sx % 2 == 0 ? (S1[sx/2] & 0xf0) >> 4 : S1[sx/2 + 1] & 0x0f;
rows1p[dx] = (S1p0 * a0 + S1p1 * a1) * 17 >> 4;

ialphap += 2;
}
}
else
{
// hresize two rows
const unsigned char* S0 = font_bitmap + 10 * (sy);
const unsigned char* S1 = font_bitmap + 10 * (sy + 1);

const short* ialphap = ialpha;
short* rows0p = rows0;
short* rows1p = rows1;
for (int dx = 0; dx < w; dx++)
{
sx = xofs[dx];
short a0 = ialphap[0];
short a1 = ialphap[1];

// const unsigned char* S0p = S0 + sx;
// const unsigned char* S1p = S1 + sx;
unsigned char S0p0 = sx % 2 == 0 ? S0[sx/2] & 0x0f : (S0[sx/2] & 0xf0) >> 4;
unsigned char S1p0 = sx % 2 == 0 ? S1[sx/2] & 0x0f : (S1[sx/2] & 0xf0) >> 4;
unsigned char S0p1 = sx % 2 == 0 ? (S0[sx/2] & 0xf0) >> 4 : S0[sx/2 + 1] & 0x0f;
unsigned char S1p1 = sx % 2 == 0 ? (S1[sx/2] & 0xf0) >> 4 : S1[sx/2 + 1] & 0x0f;
rows0p[dx] = (S0p0 * a0 + S0p1 * a1) * 17 >> 4;
rows1p[dx] = (S1p0 * a0 + S1p1 * a1) * 17 >> 4;

ialphap += 2;
}
}

prev_sy1 = sy;

if (dy + 1 < h && yofs[dy + 1] == sy)
{
// vresize for two rows
short b0 = ibeta[0];
short b1 = ibeta[1];
short b2 = ibeta[2];
short b3 = ibeta[3];

short* rows0p = rows0;
short* rows1p = rows1;
unsigned char* Dp0 = resized_font_bitmap + w * (dy);
unsigned char* Dp1 = resized_font_bitmap + w * (dy + 1);

int dx = 0;
#if __ARM_NEON
int16x8_t _b0 = vdupq_n_s16(b0);
int16x8_t _b1 = vdupq_n_s16(b1);
int16x8_t _b2 = vdupq_n_s16(b2);
int16x8_t _b3 = vdupq_n_s16(b3);
int16x8_t _v2 = vdupq_n_s16(2);
for (; dx + 7 < w; dx += 8)
{
int16x8_t _r0 = vld1q_s16(rows0p);
int16x8_t _r1 = vld1q_s16(rows1p);
int16x8_t _acc0 = vaddq_s16(vqdmulhq_s16(_r0, _b0), vqdmulhq_s16(_r1, _b1));
int16x8_t _acc1 = vaddq_s16(vqdmulhq_s16(_r0, _b2), vqdmulhq_s16(_r1, _b3));
uint8x8_t _Dp0 = vqrshrun_n_s16(_acc0, 3);
uint8x8_t _Dp1 = vqrshrun_n_s16(_acc1, 3);
vst1_u8(Dp0, _Dp0);
vst1_u8(Dp1, _Dp1);
Dp0 += 8;
Dp1 += 8;
rows0p += 8;
rows1p += 8;
}
#endif // __ARM_NEON
for (; dx < w; dx++)
{
short s0 = *rows0p++;
short s1 = *rows1p++;

*Dp0++ = (unsigned char)(((short)((b0 * s0) >> 16) + (short)((b1 * s1) >> 16) + 2) >> 2);
*Dp1++ = (unsigned char)(((short)((b2 * s0) >> 16) + (short)((b3 * s1) >> 16) + 2) >> 2);
}

ibeta += 4;
dy += 1;
}
else
{
// vresize
short b0 = ibeta[0];
short b1 = ibeta[1];

short* rows0p = rows0;
short* rows1p = rows1;
unsigned char* Dp = resized_font_bitmap + w * (dy);

int dx = 0;
#if __ARM_NEON
int16x8_t _b0 = vdupq_n_s16(b0);
int16x8_t _b1 = vdupq_n_s16(b1);
int16x8_t _v2 = vdupq_n_s16(2);
for (; dx + 7 < w; dx += 8)
{
int16x8_t _r0 = vld1q_s16(rows0p);
int16x8_t _r1 = vld1q_s16(rows1p);
int16x8_t _acc = vaddq_s16(vqdmulhq_s16(_r0, _b0), vqdmulhq_s16(_r1, _b1));
uint8x8_t _Dp = vqrshrun_n_s16(_acc, 3);
vst1_u8(Dp, _Dp);
Dp += 8;
rows0p += 8;
rows1p += 8;
}
#endif // __ARM_NEON
for (; dx < w; dx++)
{
short s0 = *rows0p++;
short s1 = *rows1p++;

*Dp++ = (unsigned char)(((short)((b0 * s0) >> 16) + (short)((b1 * s1) >> 16) + 2) >> 2);
}

ibeta += 2;
}
}

delete[] buf;
}

void draw_text_c1(unsigned char* pixels, int w, int h, int stride, const char* text, int x, int y, int fontpixelsize, unsigned int color)
{
const unsigned char* pen_color = (const unsigned char*)&color;
Expand All @@ -1275,7 +1517,7 @@ void draw_text_c1(unsigned char* pixels, int w, int h, int stride, const char* t
const unsigned char* font_bitmap = mono_font_data[ch - ' '];

// draw resized character
resize_bilinear_c1(font_bitmap, 20, 40, resized_font_bitmap, fontpixelsize, fontpixelsize * 2);
resize_bilinear_font(font_bitmap, resized_font_bitmap, fontpixelsize);

for (int j = cursor_y; j < cursor_y + fontpixelsize * 2; j++)
{
Expand Down Expand Up @@ -1336,7 +1578,7 @@ void draw_text_c2(unsigned char* pixels, int w, int h, int stride, const char* t
const unsigned char* font_bitmap = mono_font_data[font_bitmap_index];

// draw resized character
resize_bilinear_c1(font_bitmap, 20, 40, resized_font_bitmap, fontpixelsize, fontpixelsize * 2);
resize_bilinear_font(font_bitmap, resized_font_bitmap, fontpixelsize);

for (int j = cursor_y; j < cursor_y + fontpixelsize * 2; j++)
{
Expand Down Expand Up @@ -1398,7 +1640,7 @@ void draw_text_c3(unsigned char* pixels, int w, int h, int stride, const char* t
const unsigned char* font_bitmap = mono_font_data[font_bitmap_index];

// draw resized character
resize_bilinear_c1(font_bitmap, 20, 40, resized_font_bitmap, fontpixelsize, fontpixelsize * 2);
resize_bilinear_font(font_bitmap, resized_font_bitmap, fontpixelsize);

for (int j = cursor_y; j < cursor_y + fontpixelsize * 2; j++)
{
Expand Down Expand Up @@ -1460,7 +1702,7 @@ void draw_text_c4(unsigned char* pixels, int w, int h, int stride, const char* t
const unsigned char* font_bitmap = mono_font_data[ch - ' '];

// draw resized character
resize_bilinear_c1(font_bitmap, 20, 40, resized_font_bitmap, fontpixelsize, fontpixelsize * 2);
resize_bilinear_font(font_bitmap, resized_font_bitmap, fontpixelsize);

for (int j = cursor_y; j < cursor_y + fontpixelsize * 2; j++)
{
Expand Down
Loading

0 comments on commit dce2f45

Please sign in to comment.