fix adaptive pooling vulkan, second try

src/layer/vulkan/shader/pooling_adaptive.comp

@@ -73,12 +73,12 @@ void main()
     afp res;
 
     // calculate adaptive kernel size
-    const int iw0 = int(floor((afp(psc(w)) / afp(psc(outw))) * afp(gx)));
-    const int iw1 = int(ceil((afp(psc(w)) / afp(psc(outw))) * afp(gx + 1)));
-    const int kernel_w = iw1 - iw0;
-    const int ih0 = int(floor((afp(psc(h)) / afp(psc(outh))) * afp(gy)));
-    const int ih1 = int(ceil((afp(psc(h)) / afp(psc(outh))) * afp(gy + 1)));
-    const int kernel_h = ih1 - ih0;
+    const int sx = psc(w) * gx / psc(outw);
+    const int ex = (psc(w) * (gx + 1) + psc(outw) - 1) / psc(outw);
+    const int kernel_w = ex - sx;
+    const int sy = psc(h) * gy / psc(outh);
+    const int ey = (psc(h) * (gy + 1) + psc(outh) - 1) / psc(outh);
+    const int kernel_h = ey - sy;
 
     if (pooling_type == 0)
     {
@@ -112,7 +112,6 @@ void main()
     if (pooling_type == 1)
     {
         float res_fp32 = 0.f;  // force accumulation in fp32
-        int area = 0;
 
 #if NCNN_image_shader
         for (int y = 0; y < kernel_h; y++)
@@ -120,7 +119,6 @@ void main()
             for (int x = 0; x < kernel_w; x++)
             {
                 res_fp32 += image3d_ld1(bottom_blob, ivec3(sx + x, sy + y, gz));
-                area += 1;
             }
         }
 #else
@@ -131,14 +129,13 @@ void main()
             for (int x = 0; x < kernel_w; x++)
             {
                 res_fp32 += buffer_ld1(bottom_blob_data, v_offset + x);
-                area += 1;
             }
 
             v_offset += psc(w);
         }
 #endif
 
-        res_fp32 /= float(area);
+        res_fp32 /= float(kernel_h * kernel_w);
         res = afp(res_fp32);  // cast to fp16 if possible
     }
 

src/layer/vulkan/shader/pooling_adaptive_pack4.comp

@@ -73,12 +73,12 @@ void main()
     afpvec4 res;
 
     // calculate adaptive kernel size
-    const int iw0 = int(floor((afp(psc(w)) / afp(psc(outw))) * afp(gx)));
-    const int iw1 = int(ceil((afp(psc(w)) / afp(psc(outw))) * afp(gx + 1)));
-    const int kernel_w = iw1 - iw0;
-    const int ih0 = int(floor((afp(psc(h)) / afp(psc(outh))) * afp(gy)));
-    const int ih1 = int(ceil((afp(psc(h)) / afp(psc(outh))) * afp(gy + 1)));
-    const int kernel_h = ih1 - ih0;
+    const int sx = psc(w) * gx / psc(outw);
+    const int ex = (psc(w) * (gx + 1) + psc(outw) - 1) / psc(outw);
+    const int kernel_w = ex - sx;
+    const int sy = psc(h) * gy / psc(outh);
+    const int ey = (psc(h) * (gy + 1) + psc(outh) - 1) / psc(outh);
+    const int kernel_h = ey - sy;
 
     if (pooling_type == 0)
     {
@@ -112,7 +112,6 @@ void main()
     else if (pooling_type == 1)
     {
         vec4 res_fp32 = vec4(0.f);  // force accumulation in fp32
-        int area = 0;
 
 #if NCNN_image_shader
         for (int y = 0; y < kernel_h; y++)
@@ -120,7 +119,6 @@ void main()
             for (int x = 0; x < kernel_w; x++)
             {
                 res_fp32 += image3d_ld4(bottom_blob, ivec3(sx + x, sy + y, gz));
-                area += 1;
             }
         }
 #else
@@ -131,14 +129,13 @@ void main()
             for (int x = 0; x < kernel_w; x++)
             {
                 res_fp32 += buffer_ld4(bottom_blob_data, v_offset + x);
-                area += 1;
             }
 
             v_offset += psc(w);
         }
 #endif
 
-        res_fp32 /= float(area);
+        res_fp32 /= float(kernel_h * kernel_w);
         res = afpvec4(res_fp32);  // cast to fp16 if possible
     }
 

src/layer/vulkan/shader/pooling_adaptive_pack8.comp

@@ -74,12 +74,12 @@ void main()
     afpvec8 res;
 
     // calculate adaptive kernel size
-    const int iw0 = int(floor((afp(psc(w)) / afp(psc(outw))) * afp(gx)));
-    const int iw1 = int(ceil((afp(psc(w)) / afp(psc(outw))) * afp(gx + 1)));
-    const int kernel_w = iw1 - iw0;
-    const int ih0 = int(floor((afp(psc(h)) / afp(psc(outh))) * afp(gy)));
-    const int ih1 = int(ceil((afp(psc(h)) / afp(psc(outh))) * afp(gy + 1)));
-    const int kernel_h = ih1 - ih0;
+    const int sx = psc(w) * gx / psc(outw);
+    const int ex = (psc(w) * (gx + 1) + psc(outw) - 1) / psc(outw);
+    const int kernel_w = ex - sx;
+    const int sy = psc(h) * gy / psc(outh);
+    const int ey = (psc(h) * (gy + 1) + psc(outh) - 1) / psc(outh);
+    const int kernel_h = ey - sy;
 
     if (pooling_type == 0)
     {
@@ -115,7 +115,6 @@ void main()
     else if (pooling_type == 1)
     {
         mat2x4 res_fp32 = mat2x4(vec4(0.f), vec4(0.f));  // force accumulation in fp32
-        int area = 0;
         
 #if NCNN_image_shader
         for (int y = 0; y < kernel_h; y++)
@@ -125,7 +124,6 @@ void main()
                 afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
                 res_fp32[0] += v[0];
                 res_fp32[1] += v[1];
-                area += 1;
             }
         }
 #else
@@ -138,13 +136,13 @@ void main()
                 afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
                 res_fp32[0] += v[0];
                 res_fp32[1] += v[1];
-                area += 1;
             }
 
             v_offset += psc(w);
         }
 #endif
 
+        int area = kernel_h * kernel_w;
         res_fp32[0] /= float(area);
         res_fp32[1] /= float(area);
         res = afpvec8(res_fp32);  // cast to fp16 if possible