apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-21 07:40:37 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[Other] Remove useless macros (#1095)

Browse Source 
* Remove useless macros

* triger ci

* fix check error

* rename INTEGRATE_PADDLE2ONNX to ENABLE_PADDLE2ONNX
This commit is contained in:
Jason
2023-01-09 21:35:23 +08:00
committed by GitHub
parent 48bc7241cb
commit df20b2a02b
35 changed files with 510 additions and 559 deletions
Download Patch File Download Diff File Expand all files Collapse all files

124
fastdeploy/vision/visualize/segmentation_arm.cc
View File

@@ -12,8 +12,6 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#ifdef ENABLE_VISION_VISUALIZE
#include "fastdeploy/vision/visualize/segmentation_arm.h"
#ifdef __ARM_NEON
#include <arm_neon.h>
@@ -24,8 +22,9 @@ namespace vision {
static constexpr int _OMP_THREADS = 2;
static inline void QuantizeBlendingWeight8(
float weight, uint8_t* old_multi_factor, uint8_t* new_multi_factor) {
static inline void QuantizeBlendingWeight8(float weight,
uint8_t* old_multi_factor,
uint8_t* new_multi_factor) {
// Quantize the weight to boost blending performance.
// if 0.0 < w <= 1/8, w ~ 1/8=1/(2^3) shift right 3 mul 1, 7
// if 1/8 < w <= 2/8, w ~ 2/8=1/(2^3) shift right 3 mul 2, 6
@@ -39,34 +38,34 @@ static inline void QuantizeBlendingWeight8(
*old_multi_factor = (8 - weight_quantize);
}
cv::Mat VisSegmentationNEON(
const cv::Mat& im, const SegmentationResult& result,
float weight, bool quantize_weight) {
#ifndef __ARM_NEON
FDASSERT(false, "FastDeploy was not compiled with Arm NEON support!")
cv::Mat VisSegmentationNEON(const cv::Mat& im, const SegmentationResult& result,
float weight, bool quantize_weight) {
#ifndef __ARM_NEON
FDASSERT(false, "FastDeploy was not compiled with Arm NEON support!")
#else
int64_t height = result.shape[0];
int64_t width = result.shape[1];
auto vis_img = cv::Mat(height, width, CV_8UC3);
int32_t size = static_cast<int32_t>(height * width);
uint8_t *vis_ptr = static_cast<uint8_t*>(vis_img.data);
const uint8_t *label_ptr = static_cast<const uint8_t*>(result.label_map.data());
const uint8_t *im_ptr = static_cast<const uint8_t*>(im.data);
uint8_t* vis_ptr = static_cast<uint8_t*>(vis_img.data);
const uint8_t* label_ptr =
static_cast<const uint8_t*>(result.label_map.data());
const uint8_t* im_ptr = static_cast<const uint8_t*>(im.data);
if (!quantize_weight) {
uint8x16_t zerox16 = vdupq_n_u8(0);
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
for (int i = 0; i < size - 15; i += 16) {
uint8x16x3_t bgrx16x3 = vld3q_u8(im_ptr + i * 3); // 48 bytes
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
uint8x16_t ibx16 = bgrx16x3.val[0];
uint8x16_t igx16 = bgrx16x3.val[1];
uint8x16_t irx16 = bgrx16x3.val[2];
// e.g 0b00000001 << 7 -> 0b10000000 128;
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
uint8x16x3_t vbgrx16x3;
// Keep the pixels of input im if mask = 0
uint8x16_t cezx16 = vceqq_u8(labelx16, zerox16);
@@ -77,35 +76,34 @@ cv::Mat VisSegmentationNEON(
}
for (int i = size - 15; i < size; i++) {
uint8_t label = label_ptr[i];
vis_ptr[i * 3 + 0] = (label << 7);
vis_ptr[i * 3 + 1] = (label << 4);
vis_ptr[i * 3 + 2] = (label << 3);
vis_ptr[i * 3 + 0] = (label << 7);
vis_ptr[i * 3 + 1] = (label << 4);
vis_ptr[i * 3 + 2] = (label << 3);
}
// Blend the colors use OpenCV
cv::addWeighted(im, 1.0 - weight, vis_img, weight, 0, vis_img);
return vis_img;
}
// Quantize the weight to boost blending performance.
// After that, we can directly use shift instructions
// to blend the colors from input im and mask. Please
// to blend the colors from input im and mask. Please
// check QuantizeBlendingWeight8 for more details.
uint8_t old_multi_factor, new_multi_factor;
QuantizeBlendingWeight8(weight, &old_multi_factor,
&new_multi_factor);
QuantizeBlendingWeight8(weight, &old_multi_factor, &new_multi_factor);
if (new_multi_factor == 0) {
return im; // Only keep origin image.
}
return im; // Only keep origin image.
}
if (new_multi_factor == 8) {
// Only keep mask, no need to blending with origin image.
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
// Only keep mask, no need to blending with origin image.
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
for (int i = 0; i < size - 15; i += 16) {
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
// e.g 0b00000001 << 7 -> 0b10000000 128;
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
uint8x16x3_t vbgr16x3;
vbgr16x3.val[0] = mbx16;
vbgr16x3.val[1] = mgx16;
@@ -114,36 +112,36 @@ cv::Mat VisSegmentationNEON(
}
for (int i = size - 15; i < size; i++) {
uint8_t label = label_ptr[i];
vis_ptr[i * 3 + 0] = (label << 7);
vis_ptr[i * 3 + 1] = (label << 4);
vis_ptr[i * 3 + 2] = (label << 3);
}
return vis_img;
vis_ptr[i * 3 + 0] = (label << 7);
vis_ptr[i * 3 + 1] = (label << 4);
vis_ptr[i * 3 + 2] = (label << 3);
}
return vis_img;
}
uint8x16_t zerox16 = vdupq_n_u8(0);
uint8x16_t old_fx16 = vdupq_n_u8(old_multi_factor);
uint8x16_t new_fx16 = vdupq_n_u8(new_multi_factor);
// Blend the two colors together with quantize 'weight'.
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
// Blend the two colors together with quantize 'weight'.
#pragma omp parallel for proc_bind(close) num_threads(_OMP_THREADS)
for (int i = 0; i < size - 15; i += 16) {
uint8x16x3_t bgrx16x3 = vld3q_u8(im_ptr + i * 3); // 48 bytes
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
uint8x16_t labelx16 = vld1q_u8(label_ptr + i); // 16 bytes
uint8x16_t ibx16 = bgrx16x3.val[0];
uint8x16_t igx16 = bgrx16x3.val[1];
uint8x16_t irx16 = bgrx16x3.val[2];
// e.g 0b00000001 << 7 -> 0b10000000 128;
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
uint8x16_t mbx16 = vshlq_n_u8(labelx16, 7);
uint8x16_t mgx16 = vshlq_n_u8(labelx16, 4);
uint8x16_t mrx16 = vshlq_n_u8(labelx16, 3);
// Moving 7 bits to the right tends to result in zero,
// So, We choose to shift 3 bits to get an approximation
// So, We choose to shift 3 bits to get an approximation
uint8x16_t ibx16_mshr = vmulq_u8(vshrq_n_u8(ibx16, 3), old_fx16);
uint8x16_t igx16_mshr = vmulq_u8(vshrq_n_u8(igx16, 3), old_fx16);
uint8x16_t igx16_mshr = vmulq_u8(vshrq_n_u8(igx16, 3), old_fx16);
uint8x16_t irx16_mshr = vmulq_u8(vshrq_n_u8(irx16, 3), old_fx16);
uint8x16_t mbx16_mshr = vmulq_u8(vshrq_n_u8(mbx16, 3), new_fx16);
uint8x16_t mgx16_mshr = vmulq_u8(vshrq_n_u8(mgx16, 3), new_fx16);
uint8x16_t mrx16_mshr = vmulq_u8(vshrq_n_u8(mrx16, 3), new_fx16);
uint8x16_t mrx16_mshr = vmulq_u8(vshrq_n_u8(mrx16, 3), new_fx16);
uint8x16_t qbx16 = vqaddq_u8(ibx16_mshr, mbx16_mshr);
uint8x16_t qgx16 = vqaddq_u8(igx16_mshr, mgx16_mshr);
uint8x16_t qrx16 = vqaddq_u8(irx16_mshr, mrx16_mshr);
@@ -152,10 +150,10 @@ cv::Mat VisSegmentationNEON(
uint8x16_t abx16 = vandq_u8(cezx16, ibx16);
uint8x16_t agx16 = vandq_u8(cezx16, igx16);
uint8x16_t arx16 = vandq_u8(cezx16, irx16);
uint8x16x3_t vbgr16x3;
// Reset qx values to 0 if label is 0, then, keep mask values
// if label is not 0
uint8x16_t ncezx16 = vmvnq_u8(cezx16);
uint8x16x3_t vbgr16x3;
// Reset qx values to 0 if label is 0, then, keep mask values
// if label is not 0
uint8x16_t ncezx16 = vmvnq_u8(cezx16);
vbgr16x3.val[0] = vorrq_u8(abx16, vandq_u8(ncezx16, qbx16));
vbgr16x3.val[1] = vorrq_u8(agx16, vandq_u8(ncezx16, qgx16));
vbgr16x3.val[2] = vorrq_u8(arx16, vandq_u8(ncezx16, qrx16));
@@ -164,18 +162,16 @@ cv::Mat VisSegmentationNEON(
}
for (int i = size - 15; i < size; i++) {
uint8_t label = label_ptr[i];
vis_ptr[i * 3 + 0] = (im_ptr[i * 3 + 0] >> 3) * old_multi_factor
+ ((label << 7) >> 3) * new_multi_factor;
vis_ptr[i * 3 + 1] = (im_ptr[i * 3 + 1] >> 3) * old_multi_factor
+ ((label << 4) >> 3) * new_multi_factor;
vis_ptr[i * 3 + 2] = (im_ptr[i * 3 + 2] >> 3) * old_multi_factor
+ ((label << 3) >> 3) * new_multi_factor;
}
vis_ptr[i * 3 + 0] = (im_ptr[i * 3 + 0] >> 3) * old_multi_factor +
((label << 7) >> 3) * new_multi_factor;
vis_ptr[i * 3 + 1] = (im_ptr[i * 3 + 1] >> 3) * old_multi_factor +
((label << 4) >> 3) * new_multi_factor;
vis_ptr[i * 3 + 2] = (im_ptr[i * 3 + 2] >> 3) * old_multi_factor +
((label << 3) >> 3) * new_multi_factor;
}
return vis_img;
#endif
#endif
}
} // namespace vision
} // namespace fastdeploy
#endif
} // namespace fastdeploy