Improve performance of simple convolution functions

octu0 · Apr 12, 2021 · e6d7fd5 · e6d7fd5
1 parent 7661853
commit e6d7fd5
Show file tree

Hide file tree

Showing 77 changed files with 17 additions and 16 deletions.
diff --git a/blurry.cpp b/blurry.cpp
@@ -890,7 +890,7 @@ Func edge_fn(Func input, Param<int32_t> width, Param<int32_t> height){
  Func gx = Func("gx");
  gx(x, y) = (in(x + 1, y) - in(x - 1, y)) / 2;
 
- Func edge = Func("edge");
+ Func edge = Func("edgedetect");
  Expr pow_gy = fast_pow(gy(x, y), 2);
  Expr pow_gx = fast_pow(gx(x, y), 2);
  Expr magnitude = pow_gy + pow_gx;
@@ -909,6 +909,7 @@ Func edge_fn(Func input, Param<int32_t> width, Param<int32_t> height){
  edge.compute_root()
  .parallel(ch);
 
+ in.compute_root();
  return edge;
 }
 

diff --git a/edge.go b/edge.go
@@ -2,16 +2,16 @@ package blurry
 
 /*
 #cgo CFLAGS: -I${SRCDIR}/include
-#cgo darwin LDFLAGS: -L${SRCDIR}/lib -lruntime_osx -ledge_osx -ldl -lm
-#cgo linux LDFLAGS: -L${SRCDIR}/lib -lruntime_linux -ledge_linux -ldl -lm
+#cgo darwin LDFLAGS: -L${SRCDIR}/lib -lruntime_osx -ledgedetect_osx -ldl -lm
+#cgo linux LDFLAGS: -L${SRCDIR}/lib -lruntime_linux -ledgedetect_linux -ldl -lm
 #include <stdlib.h>
 #include <string.h>
 
 #include "bridge.h"
 #ifdef __APPLE__
-#include "libedge_osx.h"
+#include "libedgedetect_osx.h"
 #elif __linux__
-#include "libedge_linux.h"
+#include "libedgedetect_linux.h"
 #endif
 
 int libedge(unsigned char *src, int32_t width, int32_t height, unsigned char *out) {
@@ -25,7 +25,7 @@ int libedge(unsigned char *src, int32_t width, int32_t height, unsigned char *ou
  return 1;
  }
 
- int ret = edge(in_rgba_buf, width, height, out_rgba_buf);
+ int ret = edgedetect(in_rgba_buf, width, height, out_rgba_buf);
  free_buf(in_rgba_buf);
  free_buf(out_rgba_buf);
  return ret;

diff --git a/include/libedge_osx.h → include/libedgedetect_linux.h b/include/libedge_osx.h → include/libedgedetect_linux.h
@@ -1,5 +1,5 @@
-#ifndef HALIDE__libedge_osx_h
-#define HALIDE__libedge_osx_h
+#ifndef HALIDE__libedgedetect_linux_h
+#define HALIDE__libedgedetect_linux_h
 #include <stdint.h>
 
 // Forward declarations of the types used in the interface
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int edge(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edge_buffer);
+int edgedetect(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edgedetect_buffer);
 
 HALIDE_FUNCTION_ATTRS
-int edge_argv(void **args);
+int edgedetect_argv(void **args);
 
 HALIDE_FUNCTION_ATTRS
-const struct halide_filter_metadata_t *edge_metadata();
+const struct halide_filter_metadata_t *edgedetect_metadata();
 
 #ifdef __cplusplus
 } // extern "C"

diff --git a/include/libedge_linux.h → include/libedgedetect_osx.h b/include/libedge_linux.h → include/libedgedetect_osx.h
@@ -1,5 +1,5 @@
-#ifndef HALIDE__libedge_linux_h
-#define HALIDE__libedge_linux_h
+#ifndef HALIDE__libedgedetect_osx_h
+#define HALIDE__libedgedetect_osx_h
 #include <stdint.h>
 
 // Forward declarations of the types used in the interface
@@ -41,13 +41,13 @@ extern "C" {
 #endif
 
 HALIDE_FUNCTION_ATTRS
-int edge(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edge_buffer);
+int edgedetect(struct halide_buffer_t *_src_buffer, int32_t _width, int32_t _height, struct halide_buffer_t *_edgedetect_buffer);
 
 HALIDE_FUNCTION_ATTRS
-int edge_argv(void **args);
+int edgedetect_argv(void **args);
 
 HALIDE_FUNCTION_ATTRS
-const struct halide_filter_metadata_t *edge_metadata();
+const struct halide_filter_metadata_t *edgedetect_metadata();
 
 #ifdef __cplusplus
 } // extern "C"

diff --git a/lib/libblockmozaic_linux.a b/lib/libblockmozaic_linux.a
diff --git a/lib/libblockmozaic_osx.a b/lib/libblockmozaic_osx.a
diff --git a/lib/libboxblur_linux.a b/lib/libboxblur_linux.a
diff --git a/lib/libboxblur_osx.a b/lib/libboxblur_osx.a
diff --git a/lib/libbrightness_linux.a b/lib/libbrightness_linux.a
diff --git a/lib/libbrightness_osx.a b/lib/libbrightness_osx.a
diff --git a/lib/libcanny_dilate_linux.a b/lib/libcanny_dilate_linux.a
diff --git a/lib/libcanny_dilate_osx.a b/lib/libcanny_dilate_osx.a
diff --git a/lib/libcanny_linux.a b/lib/libcanny_linux.a
diff --git a/lib/libcanny_morphology_close_linux.a b/lib/libcanny_morphology_close_linux.a
diff --git a/lib/libcanny_morphology_close_osx.a b/lib/libcanny_morphology_close_osx.a
diff --git a/lib/libcanny_morphology_open_linux.a b/lib/libcanny_morphology_open_linux.a
diff --git a/lib/libcanny_morphology_open_osx.a b/lib/libcanny_morphology_open_osx.a
diff --git a/lib/libcanny_osx.a b/lib/libcanny_osx.a
diff --git a/lib/libcloneimg_linux.a b/lib/libcloneimg_linux.a
diff --git a/lib/libcloneimg_osx.a b/lib/libcloneimg_osx.a
diff --git a/lib/libcontrast_linux.a b/lib/libcontrast_linux.a
diff --git a/lib/libcontrast_osx.a b/lib/libcontrast_osx.a
diff --git a/lib/libdilation_linux.a b/lib/libdilation_linux.a
diff --git a/lib/libdilation_osx.a b/lib/libdilation_osx.a
diff --git a/lib/libedge_osx.a b/lib/libedge_osx.a
diff --git a/lib/libedge_linux.a → lib/libedgedetect_linux.a b/lib/libedge_linux.a → lib/libedgedetect_linux.a
diff --git a/lib/libedgedetect_osx.a b/lib/libedgedetect_osx.a
diff --git a/lib/libemboss_linux.a b/lib/libemboss_linux.a
diff --git a/lib/libemboss_osx.a b/lib/libemboss_osx.a
diff --git a/lib/liberosion_linux.a b/lib/liberosion_linux.a
diff --git a/lib/liberosion_osx.a b/lib/liberosion_osx.a
diff --git a/lib/libgammacorrection_linux.a b/lib/libgammacorrection_linux.a
diff --git a/lib/libgammacorrection_osx.a b/lib/libgammacorrection_osx.a
diff --git a/lib/libgaussianblur_linux.a b/lib/libgaussianblur_linux.a
diff --git a/lib/libgaussianblur_osx.a b/lib/libgaussianblur_osx.a
diff --git a/lib/libgradient_linux.a b/lib/libgradient_linux.a
diff --git a/lib/libgradient_osx.a b/lib/libgradient_osx.a
diff --git a/lib/libgrayscale_linux.a b/lib/libgrayscale_linux.a
diff --git a/lib/libgrayscale_osx.a b/lib/libgrayscale_osx.a
diff --git a/lib/libhighpass_linux.a b/lib/libhighpass_linux.a
diff --git a/lib/libhighpass_osx.a b/lib/libhighpass_osx.a
diff --git a/lib/libinvert_linux.a b/lib/libinvert_linux.a
diff --git a/lib/libinvert_osx.a b/lib/libinvert_osx.a
diff --git a/lib/liblaplacian_linux.a b/lib/liblaplacian_linux.a
diff --git a/lib/liblaplacian_osx.a b/lib/liblaplacian_osx.a
diff --git a/lib/libmatch_template_ncc_linux.a b/lib/libmatch_template_ncc_linux.a
diff --git a/lib/libmatch_template_ncc_osx.a b/lib/libmatch_template_ncc_osx.a
diff --git a/lib/libmatch_template_sad_linux.a b/lib/libmatch_template_sad_linux.a
diff --git a/lib/libmatch_template_sad_osx.a b/lib/libmatch_template_sad_osx.a
diff --git a/lib/libmatch_template_ssd_linux.a b/lib/libmatch_template_ssd_linux.a
diff --git a/lib/libmatch_template_ssd_osx.a b/lib/libmatch_template_ssd_osx.a
diff --git a/lib/libmatch_template_zncc_linux.a b/lib/libmatch_template_zncc_linux.a
diff --git a/lib/libmatch_template_zncc_osx.a b/lib/libmatch_template_zncc_osx.a
diff --git a/lib/libmorphology_close_linux.a b/lib/libmorphology_close_linux.a
diff --git a/lib/libmorphology_close_osx.a b/lib/libmorphology_close_osx.a
diff --git a/lib/libmorphology_gradient_linux.a b/lib/libmorphology_gradient_linux.a
diff --git a/lib/libmorphology_gradient_osx.a b/lib/libmorphology_gradient_osx.a
diff --git a/lib/libmorphology_open_linux.a b/lib/libmorphology_open_linux.a
diff --git a/lib/libmorphology_open_osx.a b/lib/libmorphology_open_osx.a
diff --git a/lib/libprepare_ncc_template_linux.a b/lib/libprepare_ncc_template_linux.a
diff --git a/lib/libprepare_ncc_template_osx.a b/lib/libprepare_ncc_template_osx.a
diff --git a/lib/libprepare_zncc_template_linux.a b/lib/libprepare_zncc_template_linux.a
diff --git a/lib/libprepare_zncc_template_osx.a b/lib/libprepare_zncc_template_osx.a
diff --git a/lib/libprepared_match_template_ncc_linux.a b/lib/libprepared_match_template_ncc_linux.a
diff --git a/lib/libprepared_match_template_ncc_osx.a b/lib/libprepared_match_template_ncc_osx.a
diff --git a/lib/libprepared_match_template_zncc_linux.a b/lib/libprepared_match_template_zncc_linux.a
diff --git a/lib/libprepared_match_template_zncc_osx.a b/lib/libprepared_match_template_zncc_osx.a
diff --git a/lib/librotate0_linux.a b/lib/librotate0_linux.a
diff --git a/lib/librotate0_osx.a b/lib/librotate0_osx.a
diff --git a/lib/librotate180_linux.a b/lib/librotate180_linux.a
diff --git a/lib/librotate180_osx.a b/lib/librotate180_osx.a
diff --git a/lib/librotate270_linux.a b/lib/librotate270_linux.a
diff --git a/lib/librotate270_osx.a b/lib/librotate270_osx.a
diff --git a/lib/librotate90_linux.a b/lib/librotate90_linux.a
diff --git a/lib/librotate90_osx.a b/lib/librotate90_osx.a
diff --git a/lib/libsobel_linux.a b/lib/libsobel_linux.a
diff --git a/lib/libsobel_osx.a b/lib/libsobel_osx.a