Merge pull request #2160 from kwagyeman/kwabena/erode_dilate_speedup

imlib/binary: Improve erode/dilate performance using SIMD.
This commit is contained in:
Ibrahim Abdelkader 2024-04-06 15:13:47 +02:00 committed by GitHub
commit 0a2773ca9a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -910,18 +910,19 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
case PIXFORMAT_BINARY: {
buf.data = fb_alloc(IMAGE_BINARY_LINE_LEN_BYTES(img) * brows, FB_ALLOC_NO_HINT);
for (int y = 0, yy = img->h; y < yy; y++) {
for (int y = 0; y < img->h; y++) {
uint32_t *row_ptr = IMAGE_COMPUTE_BINARY_PIXEL_ROW_PTR(img, y);
uint32_t *buf_row_ptr = IMAGE_COMPUTE_BINARY_PIXEL_ROW_PTR(&buf, (y % brows));
int acc = 0;
for (int x = 0, xx = img->w; x < xx; x++) {
for (int x = 0; x < img->w; x++) {
int pixel = IMAGE_GET_BINARY_PIXEL_FAST(row_ptr, x);
IMAGE_PUT_BINARY_PIXEL_FAST(buf_row_ptr, x, pixel);
if (mask && (!image_get_mask_pixel(mask, x, y))) {
continue; // Short circuit.
}
if (x > ksize && x < img->w - ksize && y >= ksize && y < img->h - ksize) {
// faster
for (int j = -ksize; j <= ksize; j++) {
@ -966,7 +967,7 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
}
// Copy any remaining lines from the buffer image...
for (int y = IM_MAX(img->h - ksize, 0), yy = img->h; y < yy; y++) {
for (int y = IM_MAX(img->h - ksize, 0); y < img->h; y++) {
memcpy(IMAGE_COMPUTE_BINARY_PIXEL_ROW_PTR(img, y),
IMAGE_COMPUTE_BINARY_PIXEL_ROW_PTR(&buf, (y % brows)),
IMAGE_BINARY_LINE_LEN_BYTES(img));
@ -978,12 +979,17 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
case PIXFORMAT_GRAYSCALE: {
buf.data = fb_alloc(IMAGE_GRAYSCALE_LINE_LEN_BYTES(img) * brows, FB_ALLOC_NO_HINT);
for (int y = 0, yy = img->h; y < yy; y++) {
#if defined(ARM_MATH_DSP)
int32_t threshold_x4 = __USAT(threshold, 7) * 0x01010101;
uint32_t e_or_d_mask_x4 = e_or_d ? 0xFFFFFFFF : 0x00000000;
#endif
for (int y = 0; y < img->h; y++) {
uint8_t *row_ptr = IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(img, y);
uint8_t *buf_row_ptr = IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(&buf, (y % brows));
int acc = 0;
for (int x = 0, xx = img->w; x < xx; x++) {
for (int x = 0; x < img->w; x++) {
int pixel = IMAGE_GET_GRAYSCALE_PIXEL_FAST(row_ptr, x);
IMAGE_PUT_GRAYSCALE_PIXEL_FAST(buf_row_ptr, x, pixel);
@ -992,13 +998,47 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
}
if (x > ksize && x < img->w - ksize && y >= ksize && y < img->h - ksize) {
// faster
for (int j = -ksize; j <= ksize; j++) {
uint8_t *k_row_ptr = IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(img, y + j);
// subtract old left edge and add new right edge to sum
acc -= (IMAGE_GET_GRAYSCALE_PIXEL_FAST(k_row_ptr, x - ksize - 1) > 0);
acc += (IMAGE_GET_GRAYSCALE_PIXEL_FAST(k_row_ptr, x + ksize) > 0);
} // for j
if (0) {
#if defined(ARM_MATH_DSP)
// acc will never be larger than 121 (<= 127) for ksize <= 5.
} else if ((x < img->w - ksize - 3) && (ksize <= 5) && (!mask)) {
int32_t acc_x4 = acc & 0xFF;
for (int j = -ksize; j <= ksize; j++) {
uint8_t *k_row_ptr = IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(img, y + j);
// subtract old left edge 4x
__USUB8(*((uint32_t *) (k_row_ptr + x - ksize - 1)), 0x01010101);
acc_x4 = __SSUB8(acc_x4, __SEL(0x01010101, 0));
// add new right edge to sum 4x
__USUB8(*((uint32_t *) (k_row_ptr + x + ksize)), 0x01010101);
acc_x4 = __SADD8(acc_x4, __SEL(0x01010101, 0));
}
// After the loop the lowest 8-bits contain a valid acc value. We
// compute the next 3 acc values by adding the sums up.
acc_x4 = __SADD8(acc_x4, acc_x4 << 8);
acc_x4 = __SADD8(acc_x4, acc_x4 << 16);
acc = acc_x4 >> 24;
// Compute all erode/dilate thresholds at once.
uint32_t pixel_x4 = *((uint32_t *) (row_ptr + x));
__SSUB8(e_or_d ? threshold_x4 : acc_x4, e_or_d ? acc_x4 : threshold_x4);
*((uint32_t *) (buf_row_ptr + x)) = __SEL(pixel_x4, e_or_d_mask_x4);
x += 3;
continue;
#endif
} else {
// faster
for (int j = -ksize; j <= ksize; j++) {
uint8_t *k_row_ptr = IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(img, y + j);
// subtract old left edge and add new right edge to sum
acc -= (IMAGE_GET_GRAYSCALE_PIXEL_FAST(k_row_ptr, x - ksize - 1) > 0);
acc += (IMAGE_GET_GRAYSCALE_PIXEL_FAST(k_row_ptr, x + ksize) > 0);
}
}
} else {
// slower way which checks boundaries per pixel
acc = e_or_d ? 0 : -1; // Don't count center pixel...
@ -1009,21 +1049,19 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
for (int k = -ksize; k <= ksize; k++) {
int x_k = IM_CLAMP(x + k, 0, (img->w - 1));
acc += IMAGE_GET_GRAYSCALE_PIXEL_FAST(k_row_ptr, x_k) > 0;
} // for k
} // for j
}
}
}
if (!e_or_d) {
// Preserve original pixel value... or clear it.
if (acc < threshold) {
IMAGE_PUT_GRAYSCALE_PIXEL_FAST(buf_row_ptr, x,
COLOR_GRAYSCALE_BINARY_MIN);
IMAGE_PUT_GRAYSCALE_PIXEL_FAST(buf_row_ptr, x, COLOR_GRAYSCALE_BINARY_MIN);
}
} else {
// Preserve original pixel value... or set it.
if (acc > threshold) {
IMAGE_PUT_GRAYSCALE_PIXEL_FAST(buf_row_ptr, x,
COLOR_GRAYSCALE_BINARY_MAX);
IMAGE_PUT_GRAYSCALE_PIXEL_FAST(buf_row_ptr, x, COLOR_GRAYSCALE_BINARY_MAX);
}
}
}
@ -1037,7 +1075,7 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
}
// Copy any remaining lines from the buffer image...
for (int y = IM_MAX(img->h - ksize, 0), yy = img->h; y < yy; y++) {
for (int y = IM_MAX(img->h - ksize, 0); y < img->h; y++) {
memcpy(IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(img, y),
IMAGE_COMPUTE_GRAYSCALE_PIXEL_ROW_PTR(&buf, (y % brows)),
IMAGE_GRAYSCALE_LINE_LEN_BYTES(img));
@ -1049,12 +1087,17 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
case PIXFORMAT_RGB565: {
buf.data = fb_alloc(IMAGE_RGB565_LINE_LEN_BYTES(img) * brows, FB_ALLOC_NO_HINT);
for (int y = 0, yy = img->h; y < yy; y++) {
#if defined(ARM_MATH_DSP)
int32_t threshold_x2 = __USAT(threshold, 15) * 0x00010001;
uint32_t e_or_d_mask_x2 = e_or_d ? 0xFFFFFFFF : 0x00000000;
#endif
for (int y = 0; y < img->h; y++) {
uint16_t *row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y);
uint16_t *buf_row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(&buf, (y % brows));
int acc = 0;
for (int x = 0, xx = img->w; x < xx; x++) {
for (int x = 0; x < img->w; x++) {
int pixel = IMAGE_GET_RGB565_PIXEL_FAST(row_ptr, x);
IMAGE_PUT_RGB565_PIXEL_FAST(buf_row_ptr, x, pixel);
@ -1063,12 +1106,45 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
}
if (x > ksize && x < img->w - ksize && y >= ksize && y < img->h - ksize) {
// faster
for (int j = -ksize; j <= ksize; j++) {
uint16_t *k_row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y + j);
// subtract old left column and add new right column
acc -= IMAGE_GET_RGB565_PIXEL_FAST(k_row_ptr, x - ksize - 1) > 0;
acc += IMAGE_GET_RGB565_PIXEL_FAST(k_row_ptr, x + ksize) > 0;
if (0) {
#if defined(ARM_MATH_DSP)
// acc will never be larger than 32,761 (<= 32767) for ksize <= 90.
} else if ((x < img->w - ksize - 1) && (ksize <= 90) && (!mask)) {
int32_t acc_x2 = acc & 0xFFFF;
for (int j = -ksize; j <= ksize; j++) {
uint16_t *k_row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y + j);
// subtract old left edge 2x
__USUB16(*((uint32_t *) (k_row_ptr + x - ksize - 1)), 0x00010001);
acc_x2 = __SSUB16(acc_x2, __SEL(0x00010001, 0));
// add new right edge to sum 2x
__USUB16(*((uint32_t *) (k_row_ptr + x + ksize)), 0x00010001);
acc_x2 = __SADD16(acc_x2, __SEL(0x00010001, 0));
}
// After the loop the lowest 16-bits contain a valid acc value. We
// compute the next acc value by adding the sums up.
acc_x2 = __SADD16(acc_x2, acc_x2 << 16);
acc = acc_x2 >> 16;
// Compute all erode/dilate thresholds at once.
uint32_t pixel_x2 = *((uint32_t *) (row_ptr + x));
__SSUB16(e_or_d ? threshold_x2 : acc_x2, e_or_d ? acc_x2 : threshold_x2);
*((uint32_t *) (buf_row_ptr + x)) = __SEL(pixel_x2, e_or_d_mask_x2);
x += 1;
continue;
#endif
} else {
// faster
for (int j = -ksize; j <= ksize; j++) {
uint16_t *k_row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y + j);
// subtract old left column and add new right column
acc -= IMAGE_GET_RGB565_PIXEL_FAST(k_row_ptr, x - ksize - 1) > 0;
acc += IMAGE_GET_RGB565_PIXEL_FAST(k_row_ptr, x + ksize) > 0;
}
}
} else {
// need to check boundary conditions for each pixel
@ -1087,14 +1163,12 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
if (!e_or_d) {
// Preserve original pixel value... or clear it.
if (acc < threshold) {
IMAGE_PUT_RGB565_PIXEL_FAST(buf_row_ptr, x,
COLOR_RGB565_BINARY_MIN);
IMAGE_PUT_RGB565_PIXEL_FAST(buf_row_ptr, x, COLOR_RGB565_BINARY_MIN);
}
} else {
// Preserve original pixel value... or set it.
if (acc > threshold) {
IMAGE_PUT_RGB565_PIXEL_FAST(buf_row_ptr, x,
COLOR_RGB565_BINARY_MAX);
IMAGE_PUT_RGB565_PIXEL_FAST(buf_row_ptr, x, COLOR_RGB565_BINARY_MAX);
}
}
}
@ -1108,7 +1182,7 @@ static void imlib_erode_dilate(image_t *img, int ksize, int threshold, int e_or_
}
// Copy any remaining lines from the buffer image...
for (int y = IM_MAX(img->h - ksize, 0), yy = img->h; y < yy; y++) {
for (int y = IM_MAX(img->h - ksize, 0); y < img->h; y++) {
memcpy(IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y),
IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(&buf, (y % brows)),
IMAGE_RGB565_LINE_LEN_BYTES(img));