require GL_EXT_shader_16bit_storage only for fp16_storage, explicit type cast

src/CMakeLists.txt

@@ -112,7 +112,7 @@ macro(ncnn_add_layer class)
             add_custom_command(
                 OUTPUT ${SHADER_fp16s_SPV_HEX_FILE}
                 COMMAND ${GLSLANGVALIDATOR_EXECUTABLE}
-                ARGS -Dsfp=float16_t -Dsfpvec2=f16vec2 -Dsfpvec4=f16vec4 -Dsfpmat4=f16mat4 -Dafp=float -Dafpvec2=vec2 -Dafpvec4=vec4 -Dafpmat4=mat4 -DNCNN_fp16_storage=1 -V -s -e ${SHADER_fp16s_SRC_NAME_WE} --source-entrypoint main -x -o ${SHADER_fp16s_SPV_HEX_FILE} ${SHADER_SRC}
+                ARGS -Dsfp=float16_t -Dsfpvec2=f16vec2 -Dsfpvec4=f16vec4 -Dafp=float -Dafpvec2=vec2 -Dafpvec4=vec4 -Dafpmat4=mat4 -DNCNN_fp16_storage=1 -V -s -e ${SHADER_fp16s_SRC_NAME_WE} --source-entrypoint main -x -o ${SHADER_fp16s_SPV_HEX_FILE} ${SHADER_SRC}
                 DEPENDS ${SHADER_SRC}
                 COMMENT "Building SPIR-V module ${SHADER_fp16s_SRC_NAME_WE}.spv"
                 VERBATIM

src/layer/shader/absval.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfp(abs(v));
 }

src/layer/shader/absval_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfpvec4(abs(v));
 }

src/layer/shader/batchnorm.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,9 +49,9 @@ void main()
 
     if (p.dims == 1)
     {
-        afp v = bottom_top_blob_data[gx];
+        afp v = afp(bottom_top_blob_data[gx]);
 
-        v = b_data[gx] * v + a_data[gx];
+        v = afp(b_data[gx]) * v + afp(a_data[gx]);
 
         bottom_top_blob_data[gx] = sfp(v);
 
@@ -59,9 +62,9 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
-        v = b_data[gy] * v + a_data[gy];
+        v = afp(b_data[gy]) * v + afp(a_data[gy]);
 
         bottom_top_blob_data[gi] = sfp(v);
 
@@ -72,9 +75,9 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
-        v = b_data[gz] * v + a_data[gz];
+        v = afp(b_data[gz]) * v + afp(a_data[gz]);
 
         bottom_top_blob_data[gi] = sfp(v);
 

src/layer/shader/batchnorm_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,9 +49,9 @@ void main()
 
     if (p.dims == 1)
     {
-        afpvec4 v = bottom_top_blob_data[gx];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gx]);
 
-        v = b_data[gx] * v + a_data[gx];
+        v = afpvec4(b_data[gx]) * v + afpvec4(a_data[gx]);
 
         bottom_top_blob_data[gx] = sfpvec4(v);
 
@@ -59,9 +62,9 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
-        v = b_data[gy] * v + a_data[gy];
+        v = afpvec4(b_data[gy]) * v + afpvec4(a_data[gy]);
 
         bottom_top_blob_data[gi] = sfpvec4(v);
 
@@ -72,9 +75,9 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
-        v = b_data[gz] * v + a_data[gz];
+        v = afpvec4(b_data[gz]) * v + afpvec4(a_data[gz]);
 
         bottom_top_blob_data[gi] = sfpvec4(v);
 

src/layer/shader/binaryop.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -62,7 +65,7 @@ void main()
 
     const int gi = gz * p.outcstep + gy * p.outw + gx;
 
-    afp v1 = a_blob_data[gi];
+    afp v1 = afp(a_blob_data[gi]);
 
     afp res;
 
@@ -87,7 +90,7 @@ void main()
 
     if (p.adims == p.bdims)
     {
-        afp v2 = b_blob_data[gi];
+        afp v2 = afp(b_blob_data[gi]);
 
         if (op_type == 0) res = v1 + v2;
         if (op_type == 1) res = v1 - v2;

src/layer/shader/binaryop_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -62,7 +65,7 @@ void main()
 
     const int gi = gz * p.outcstep + gy * p.outw + gx;
 
-    afpvec4 v1 = a_blob_data[gi];
+    afpvec4 v1 = afpvec4(a_blob_data[gi]);
 
     afpvec4 res;
 
@@ -87,7 +90,7 @@ void main()
 
     if (p.adims == p.bdims)
     {
-        afpvec4 v2 = b_blob_data[gi];
+        afpvec4 v2 = afpvec4(b_blob_data[gi]);
 
         if (op_type == 0) res = v1 + v2;
         if (op_type == 1) res = v1 - v2;

src/layer/shader/clip.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -47,7 +50,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfp(clamp(v, afp(const_min), afp(const_max)));
 }

src/layer/shader/clip_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -47,7 +50,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfpvec4(clamp(v, afp(const_min), afp(const_max)));
 }

src/layer/shader/concat.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -53,39 +56,34 @@ void main()
     if (gx >= p.w || gy >= p.h || gz >= p.c)
         return;
 
-    int v_offset;
-    sfp v;
-
     if (p.dims == 1) // axis == 0
     {
-        v_offset = gx + p.offset;
-        v = bottom_blob_data[gx];
+        int v_offset = gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gx];
     }
     else if (p.dims == 2 && axis == 0)
     {
-        v_offset = (gy + p.offset) * p.outw + gx;
-        v = bottom_blob_data[gy * p.w + gx];
+        int v_offset = (gy + p.offset) * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gy * p.w + gx];
     }
     else if (p.dims == 2 && axis == 1)
     {
-        v_offset = gy * p.outw + gx + p.offset;
-        v = bottom_blob_data[gy * p.w + gx];
+        int v_offset = gy * p.outw + gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 0)
     {
-        v_offset = (gz + p.offset) * p.outcstep + gy * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = (gz + p.offset) * p.outcstep + gy * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 1)
     {
-        v_offset = gz * p.outcstep + (gy + p.offset) * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = gz * p.outcstep + (gy + p.offset) * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 2)
     {
-        v_offset = gz * p.outcstep + gy * p.outw + gx + p.offset;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = gz * p.outcstep + gy * p.outw + gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
-
-    top_blob_data[v_offset] = v;
 }

src/layer/shader/concat_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -53,39 +56,34 @@ void main()
     if (gx >= p.w || gy >= p.h || gz >= p.c)
         return;
 
-    int v_offset;
-    sfpvec4 v;
-
     if (p.dims == 1) // axis == 0
     {
-        v_offset = gx + p.offset;
-        v = bottom_blob_data[gx];
+        int v_offset = gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gx];
     }
     else if (p.dims == 2 && axis == 0)
     {
-        v_offset = (gy + p.offset) * p.outw + gx;
-        v = bottom_blob_data[gy * p.w + gx];
+        int v_offset = (gy + p.offset) * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gy * p.w + gx];
     }
     else if (p.dims == 2 && axis == 1)
     {
-        v_offset = gy * p.outw + gx + p.offset;
-        v = bottom_blob_data[gy * p.w + gx];
+        int v_offset = gy * p.outw + gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 0)
     {
-        v_offset = (gz + p.offset) * p.outcstep + gy * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = (gz + p.offset) * p.outcstep + gy * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 1)
     {
-        v_offset = gz * p.outcstep + (gy + p.offset) * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = gz * p.outcstep + (gy + p.offset) * p.outw + gx;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
     else if (p.dims == 3 && axis == 2)
     {
-        v_offset = gz * p.outcstep + gy * p.outw + gx + p.offset;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        int v_offset = gz * p.outcstep + gy * p.outw + gx + p.offset;
+        top_blob_data[v_offset] = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
     }
-
-    top_blob_data[v_offset] = v;
 }

src/layer/shader/concat_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -54,41 +57,41 @@ void main()
         return;
 
     ivec4 v_offset;
-    sfpvec4 v;
+    int gi;
 
     if (p.dims == 1) // axis == 0
     {
         v_offset = ivec4(gx * 4 + p.offset) + ivec4(0, 1, 2, 3);
-        v = bottom_blob_data[gx];
+        gi = gx;
     }
     else if (p.dims == 2 && axis == 0)
     {
         v_offset = (ivec4(gx * 4 + p.offset) + ivec4(0, 1, 2, 3)) * p.outw + gx;
-        v = bottom_blob_data[gy * p.w + gx];
+        gi = gy * p.w + gx;
     }
     else if (p.dims == 2 && axis == 1)
     {
         v_offset = (ivec4(gx * 4) + ivec4(0, 1, 2, 3)) * p.outw + gx + p.offset;
-        v = bottom_blob_data[gy * p.w + gx];
+        gi = gy * p.w + gx;
     }
     else if (p.dims == 3 && axis == 0)
     {
         v_offset = (ivec4(gz * 4 + p.offset) + ivec4(0, 1, 2, 3)) * p.outcstep + gy * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        gi = gz * p.cstep + gy * p.w + gx;
     }
     else if (p.dims == 3 && axis == 1)
     {
         v_offset = (ivec4(gz * 4) + ivec4(0, 1, 2, 3)) * p.outcstep + (gy + p.offset) * p.outw + gx;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        gi = gz * p.cstep + gy * p.w + gx;
     }
     else if (p.dims == 3 && axis == 2)
     {
         v_offset = (ivec4(gz * 4) + ivec4(0, 1, 2, 3)) * p.outcstep + gy * p.outw + gx + p.offset;
-        v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+        gi = gz * p.cstep + gy * p.w + gx;
     }
 
-    top_blob_data[v_offset.r] = v.r;
-    top_blob_data[v_offset.g] = v.g;
-    top_blob_data[v_offset.b] = v.b;
-    top_blob_data[v_offset.a] = v.a;
+    top_blob_data[v_offset.r] = bottom_blob_data[gi].r;
+    top_blob_data[v_offset.g] = bottom_blob_data[gi].g;
+    top_blob_data[v_offset.b] = bottom_blob_data[gi].b;
+    top_blob_data[v_offset.a] = bottom_blob_data[gi].a;
 }

src/layer/shader/convolution.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/convolution_1x1s1d1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/convolution_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -80,9 +83,9 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afp v = bottom_blob_data[v_offset + x * dilation_w];
+                afp v = afp(bottom_blob_data[v_offset + x * dilation_w]);
 
-                afpvec4 k = weight_data[w_offset + x];
+                afpvec4 k = afpvec4(weight_data[w_offset + x]);
 
                 sum += v * k;
             }

src/layer/shader/convolution_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -32,7 +35,12 @@ layout (local_size_z_id = 235) in;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+#else
 layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
+#endif
 layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
@@ -63,7 +71,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -80,9 +88,19 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afpvec4 v = bottom_blob_data[v_offset + x * dilation_w];
-
-                afpmat4 k = weight_data[w_offset + x];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);
+
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+                afpmat4 k = afpmat4(
+                    afpvec4(weight_data[(w_offset + x) * 4 + 0]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 1]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 2]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 3])
+                );
+#else
+                afpmat4 k = afpmat4(weight_data[w_offset + x]);
+#endif
 
                 sum += v * k;
             }

src/layer/shader/convolution_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {
@@ -80,9 +83,9 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afpvec4 v = bottom_blob_data[v_offset + x * dilation_w];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);
 
-                afpvec4 k = weight_data[w_offset + x];
+                afpvec4 k = afpvec4(weight_data[w_offset + x]);
 
                 sum += dot(v, k);
             }

src/layer/shader/convolutiondepthwise.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/convolutiondepthwise_group.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/convolutiondepthwise_group_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -89,9 +92,9 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afp v = bottom_blob_data[v_offset + x * dilation_w];
+                afp v = afp(bottom_blob_data[v_offset + x * dilation_w]);
 
-                afpvec4 k = weight_data[w_offset + x];
+                afpvec4 k = afpvec4(weight_data[w_offset + x]);
 
                 sum += v * k;
             }

src/layer/shader/convolutiondepthwise_group_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -33,7 +36,12 @@ layout (local_size_z_id = 235) in;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+#else
 layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
+#endif
 layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
@@ -64,7 +72,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -89,9 +97,19 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afpvec4 v = bottom_blob_data[v_offset + x * dilation_w];
-
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);
+
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+                afpmat4 k = afpmat4(
+                    afpvec4(weight_data[(w_offset + x) * 4 + 0]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 1]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 2]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 3])
+                );
+#else
                 afpmat4 k = weight_data[w_offset + x];
+#endif
 
                 sum += v * k;
             }

src/layer/shader/convolutiondepthwise_group_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {
@@ -89,9 +92,9 @@ void main()
         {
             for (int x = 0; x < kernel_w; x++)
             {
-                afpvec4 v = bottom_blob_data[v_offset + x * dilation_w];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);
 
-                afpvec4 k = weight_data[w_offset + x];
+                afpvec4 k = afpvec4(weight_data[w_offset + x]);
 
                 sum += dot(v, k);
             }

src/layer/shader/convolutiondepthwise_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,11 +67,11 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
-        sum = afpvec4(0.0);
+        sum = afpvec4(0.f);
     }
 
     // depth-wise convolution
@@ -79,9 +82,9 @@ void main()
     {
         for (int x = 0; x < kernel_w; x++)
         {
-            afpvec4 v = bottom_blob_data[v_offset + x * dilation_w];
+            afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);
 
-            afpvec4 k = weight_data[w_offset + x];
+            afpvec4 k = afpvec4(weight_data[w_offset + x]);
 
             sum += v * k;
         }

src/layer/shader/crop.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/crop_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/deconvolution.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/deconvolution_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -100,9 +103,9 @@ void main()
 
             for (int z = 0; z < p.c; z++)
             {
-                afp v = bottom_blob_data[v_offset];
+                afp v = afp(bottom_blob_data[v_offset]);
 
-                afpvec4 k = weight_data[w_offset];
+                afpvec4 k = afpvec4(weight_data[w_offset]);
 
                 sum += v * k;
 

src/layer/shader/deconvolution_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -32,7 +35,12 @@ layout (local_size_z_id = 235) in;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+#else
 layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
+#endif
 layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
@@ -63,7 +71,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -100,9 +108,19 @@ void main()
 
             for (int z = 0; z < p.c; z++)
             {
-                afpvec4 v = bottom_blob_data[v_offset];
-
-                afpmat4 k = weight_data[w_offset];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
+
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+                afpmat4 k = afpmat4(
+                    afpvec4(weight_data[(w_offset + x) * 4 + 0]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 1]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 2]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 3])
+                );
+#else
+                afpmat4 k = afpmat4(weight_data[w_offset]);
+#endif
 
                 sum += v * k;
 

src/layer/shader/deconvolution_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {
@@ -100,9 +103,9 @@ void main()
 
             for (int z = 0; z < p.c; z++)
             {
-                afpvec4 v = bottom_blob_data[v_offset];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
 
-                afpvec4 k = weight_data[w_offset];
+                afpvec4 k = afpvec4(weight_data[w_offset]);
 
                 sum += dot(v, k);
 

src/layer/shader/deconvolutiondepthwise.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/deconvolutiondepthwise_group.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {

src/layer/shader/deconvolutiondepthwise_group_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -109,9 +112,9 @@ void main()
 
             for (int z = 0; z < channels_g; z++)
             {
-                afp v = bottom_blob_data[v_offset];
+                afp v = afp(bottom_blob_data[v_offset]);
 
-                afpvec4 k = weight_data[w_offset];
+                afpvec4 k = afpvec4(weight_data[w_offset]);
 
                 sum += v * k;
 

src/layer/shader/deconvolutiondepthwise_group_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -33,7 +36,12 @@ layout (local_size_z_id = 235) in;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+#else
 layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
+#endif
 layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
@@ -64,7 +72,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -109,9 +117,19 @@ void main()
 
             for (int z = 0; z < channels_g; z++)
             {
-                afpvec4 v = bottom_blob_data[v_offset];
-
-                afpmat4 k = weight_data[w_offset];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
+
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+                afpmat4 k = afpmat4(
+                    afpvec4(weight_data[(w_offset + x) * 4 + 0]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 1]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 2]),
+                    afpvec4(weight_data[(w_offset + x) * 4 + 3])
+                );
+#else
+                afpmat4 k = afpmat4(weight_data[w_offset]);
+#endif
 
                 sum += v * k;
 

src/layer/shader/deconvolutiondepthwise_group_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afp(bias_data[gz]);
     }
     else
     {
@@ -109,9 +112,9 @@ void main()
 
             for (int z = 0; z < channels_g; z++)
             {
-                afpvec4 v = bottom_blob_data[v_offset];
+                afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
 
-                afpvec4 k = weight_data[w_offset];
+                afpvec4 k = afpvec4(weight_data[w_offset]);
 
                 sum += dot(v, k);
 

src/layer/shader/deconvolutiondepthwise_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -64,7 +67,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gz];
+        sum = afpvec4(bias_data[gz]);
     }
     else
     {
@@ -101,9 +104,9 @@ void main()
             int v_offset = v_offset_0 + sy * p.w + sx;
             int w_offset = w_offset_0 + y * kernel_w + x;
 
-            afpvec4 v = bottom_blob_data[v_offset];
+            afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
 
-            afpvec4 k = weight_data[w_offset];
+            afpvec4 k = afpvec4(weight_data[w_offset]);
 
             sum += v * k;
         }

src/layer/shader/dropout.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     v *= afp(scale);
 

src/layer/shader/dropout_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     v *= afp(scale);
 

src/layer/shader/eltwise.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -52,8 +55,8 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v1 = bottom_blob1_data[gi];
-    afp v2 = bottom_blob2_data[gi];
+    afp v1 = afp(bottom_blob1_data[gi]);
+    afp v2 = afp(bottom_blob2_data[gi]);
 
     afp res;
 

src/layer/shader/eltwise_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -52,8 +55,8 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v1 = bottom_blob1_data[gi];
-    afpvec4 v2 = bottom_blob2_data[gi];
+    afpvec4 v1 = afpvec4(bottom_blob1_data[gi]);
+    afpvec4 v2 = afpvec4(bottom_blob2_data[gi]);
 
     afpvec4 res;
 

src/layer/shader/flatten.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/flatten_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -74,11 +77,8 @@ void main()
         v_offset = ivec4(z * 4 * p.cstep) + si4 * 4 + k4;
     }
 
-    sfpvec4 v;
-    v.r = bottom_blob_data[v_offset.r];
-    v.g = bottom_blob_data[v_offset.g];
-    v.b = bottom_blob_data[v_offset.b];
-    v.a = bottom_blob_data[v_offset.a];
-
-    top_blob_data[gx] = v;
+    top_blob_data[gx].r = bottom_blob_data[v_offset.r];
+    top_blob_data[gx].g = bottom_blob_data[v_offset.g];
+    top_blob_data[gx].b = bottom_blob_data[v_offset.b];
+    top_blob_data[gx].a = bottom_blob_data[v_offset.a];
 }

src/layer/shader/innerproduct.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -57,7 +60,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gx];
+        sum = afp(bias_data[gx]);
     }
     else
     {

src/layer/shader/innerproduct_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -57,20 +60,20 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gx];
+        sum = afpvec4(bias_data[gx]);
     }
     else
     {
-        sum = afpvec4(0.0);
+        sum = afpvec4(0.f);
     }
 
     int w_offset = gx * p.w;
 
     for (int i = 0; i < p.w; i++)
     {
-        afp v = bottom_blob_data[i];
+        afp v = afp(bottom_blob_data[i]);
 
-        afpvec4 k = weight_data[w_offset + i];
+        afpvec4 k = afpvec4(weight_data[w_offset + i]);
 
         sum += v * k;
     }

src/layer/shader/innerproduct_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -26,7 +29,12 @@ layout (local_size_z_id = 235) in;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+#else
 layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
+#endif
 layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
@@ -57,20 +65,30 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gx];
+        sum = afpvec4(bias_data[gx]);
     }
     else
     {
-        sum = afpvec4(0.0);
+        sum = afpvec4(0.f);
     }
 
     int w_offset = gx * p.w;
 
     for (int i = 0; i < p.w; i++)
     {
-        afpvec4 v = bottom_blob_data[i];
-
-        afpmat4 k = weight_data[w_offset + i];
+        afpvec4 v = afpvec4(bottom_blob_data[i]);
+
+#if NCNN_fp16_storage && !NCNN_fp16_arithmetic
+        // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
+        afpmat4 k = afpmat4(
+            afpvec4(weight_data[(w_offset + i) * 4 + 0]),
+            afpvec4(weight_data[(w_offset + i) * 4 + 1]),
+            afpvec4(weight_data[(w_offset + i) * 4 + 2]),
+            afpvec4(weight_data[(w_offset + i) * 4 + 3])
+        );
+#else
+        afpmat4 k = afpmat4(weight_data[w_offset + i]);
+#endif
 
         sum += v * k;
     }

src/layer/shader/innerproduct_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -57,7 +60,7 @@ void main()
 
     if (bias_term == 1)
     {
-        sum = bias_data[gx];
+        sum = afp(bias_data[gx]);
     }
     else
     {
@@ -68,9 +71,9 @@ void main()
 
     for (int i = 0; i < p.w; i++)
     {
-        afpvec4 v = bottom_blob_data[i];
+        afpvec4 v = afpvec4(bottom_blob_data[i]);
 
-        afpvec4 k = weight_data[w_offset + i];
+        afpvec4 k = afpvec4(weight_data[w_offset + i]);
 
         sum += dot(v, k);
     }

src/layer/shader/interp.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -54,8 +57,6 @@ void main()
     if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
-    afp res;
-
     if (resize_type == 1) // nearest
     {
         afpvec2 gxy = afpvec2(gx, gy);
@@ -67,7 +68,7 @@ void main()
 
         int v_offset = gz * p.cstep + sy * p.w + sx;
 
-        res = bottom_blob_data[v_offset];
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
     }
     else if (resize_type == 2) // bilinear
     {
@@ -94,17 +95,15 @@ void main()
         int v_offset_0 = gz * p.cstep + sy * p.w + sx;
         int v_offset_1 = gz * p.cstep + (sy + 1) * p.w + sx;
 
-        afp a0 = bottom_blob_data[v_offset_0];
-        afp a1 = bottom_blob_data[v_offset_0 + 1];
-        afp b0 = bottom_blob_data[v_offset_1];
-        afp b1 = bottom_blob_data[v_offset_1 + 1];
+        afp a0 = afp(bottom_blob_data[v_offset_0]);
+        afp a1 = afp(bottom_blob_data[v_offset_0 + 1]);
+        afp b0 = afp(bottom_blob_data[v_offset_1]);
+        afp b1 = afp(bottom_blob_data[v_offset_1 + 1]);
 
         afp fx = fxy.r;
         afp fy = fxy.g;
 
         afpvec2 ab = afpvec2(a0, b0) * (afp(1.f) - fx) + afpvec2(a1, b1) * fx;
-        res = ab.r * (afp(1.f) - fy) + ab.g * fy;
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfp(ab.r * (afp(1.f) - fy) + ab.g * fy);
     }
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfp(res);
 }

src/layer/shader/interp_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -54,7 +57,7 @@ void main()
     if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
-    afpvec4 res;
+//     afpvec4 res;
 
     if (resize_type == 1) // nearest
     {
@@ -67,7 +70,7 @@ void main()
 
         int v_offset = gz * p.cstep + sy * p.w + sx;
 
-        res = bottom_blob_data[v_offset];
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
     }
     else if (resize_type == 2) // bilinear
     {
@@ -94,10 +97,10 @@ void main()
         int v_offset_0 = gz * p.cstep + sy * p.w + sx;
         int v_offset_1 = gz * p.cstep + (sy + 1) * p.w + sx;
 
-        afpvec4 a0 = bottom_blob_data[v_offset_0];
-        afpvec4 a1 = bottom_blob_data[v_offset_0 + 1];
-        afpvec4 b0 = bottom_blob_data[v_offset_1];
-        afpvec4 b1 = bottom_blob_data[v_offset_1 + 1];
+        afpvec4 a0 = afpvec4(bottom_blob_data[v_offset_0]);
+        afpvec4 a1 = afpvec4(bottom_blob_data[v_offset_0 + 1]);
+        afpvec4 b0 = afpvec4(bottom_blob_data[v_offset_1]);
+        afpvec4 b1 = afpvec4(bottom_blob_data[v_offset_1 + 1]);
 
         afp fx = fxy.r;
         afp fy = fxy.g;
@@ -105,8 +108,6 @@ void main()
         afpvec4 a = a0 * (afp(1.f) - fx) + a1 * fx;
         afpvec4 b = b0 * (afp(1.f) - fx) + b1 * fx;
 
-        res = a * (afp(1.f) - fy) + b * fy;
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfpvec4(a * (afp(1.f) - fy) + b * fy);
     }
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfpvec4(res);
 }

src/layer/shader/lrn_norm.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -95,7 +98,7 @@ void main()
         scale = pow(afp(bias_constant) + alpha_div_size * sum, afp(-beta));
     }
 
-    afp v = bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx];
+    afp v = afp(bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx]);
 
     v *= scale;
 

src/layer/shader/lrn_norm_across_channel_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -75,7 +78,7 @@ void main()
 
     afpvec4 scale = pow(afp(bias_constant) + alpha_div_size * sum, afpvec4(-beta));
 
-    afpvec4 v = bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx]);
 
     v *= scale;
 

src/layer/shader/lrn_norm_within_channel_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -75,7 +78,7 @@ void main()
 
     afpvec4 scale = pow(afp(bias_constant) + alpha_div_size * sum, afpvec4(-beta));
 
-    afpvec4 v = bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gz * p.outcstep + gy * p.outw + gx]);
 
     v *= scale;
 

src/layer/shader/lrn_square_pad.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -62,7 +65,7 @@ void main()
         if (z >= 0 && z < p.c)
         {
             int v_offset = z * p.cstep + gy * p.w + gx;
-            afp v = bottom_blob_data[v_offset];
+            afp v = afp(bottom_blob_data[v_offset]);
             res = v * v;
         }
         else
@@ -78,7 +81,7 @@ void main()
         if (x >= 0 && x < p.w && y >= 0 && y < p.h)
         {
             int v_offset = gz * p.cstep + y * p.w + x;
-            afp v = bottom_blob_data[v_offset];
+            afp v = afp(bottom_blob_data[v_offset]);
             res = v * v;
         }
         else

src/layer/shader/lrn_square_pad_across_channel_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -62,7 +65,7 @@ void main()
     if (z >= 0 && z < p.c)
     {
         int v_offset = z * p.cstep + gy * p.w + gx;
-        afpvec4 v4 = bottom_blob_data[v_offset];
+        afpvec4 v4 = afpvec4(bottom_blob_data[v_offset]);
 
         int lane = (gz - pad_head) % 4;
 

src/layer/shader/lrn_square_pad_within_channel_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,7 +66,7 @@ void main()
     if (x >= 0 && x < p.w && y >= 0 && y < p.h)
     {
         int v_offset = gz * p.cstep + y * p.w + x;
-        afpvec4 v = bottom_blob_data[v_offset];
+        afpvec4 v = afpvec4(bottom_blob_data[v_offset]);
         res = v * v;
     }
     else

src/layer/shader/packing_1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -55,35 +58,25 @@ void main()
     {
         ivec4 x4 = ivec4(gx * 4) + ivec4(0, 1, 2, 3);
 
-        // prevent out of range access
-//         x4 = min(x4, p.w - 1);
-
         v_offset = x4;
     }
     else if (p.dims == 2)
     {
         ivec4 y4 = ivec4(gy * 4) + ivec4(0, 1, 2, 3);
 
-        // prevent out of range access
-//         y4 = min(y4, p.h - 1);
-
         v_offset = y4 * p.w + gx;
     }
     else // if (p.dims == 3)
     {
         ivec4 z4 = ivec4(gz * 4) + ivec4(0, 1, 2, 3);
 
-        // prevent out of range access
-        z4 = min(z4, p.c - 1);
-
         v_offset = z4 * p.cstep + ivec4(gy * p.w + gx);
     }
 
-    sfpvec4 v;
-    v.r = bottom_blob_data[v_offset.r];
-    v.g = bottom_blob_data[v_offset.g];
-    v.b = bottom_blob_data[v_offset.b];
-    v.a = bottom_blob_data[v_offset.a];
+    int gi = gz * p.outcstep + gy * p.outw + gx;
 
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    top_blob_data[gi].r = bottom_blob_data[v_offset.r];
+    top_blob_data[gi].g = bottom_blob_data[v_offset.g];
+    top_blob_data[gi].b = bottom_blob_data[v_offset.b];
+    top_blob_data[gi].a = bottom_blob_data[v_offset.a];
 }

src/layer/shader/packing_4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -70,10 +73,10 @@ void main()
         v_offset = z4 * p.outcstep + ivec4(gy * p.outw + gx);
     }
 
-    sfpvec4 v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+    int gi = gz * p.cstep + gy * p.w + gx;
 
-    top_blob_data[v_offset.r] = v.r;
-    top_blob_data[v_offset.g] = v.g;
-    top_blob_data[v_offset.b] = v.b;
-    top_blob_data[v_offset.a] = v.a;
+    top_blob_data[v_offset.r] = bottom_blob_data[gi].r;
+    top_blob_data[v_offset.g] = bottom_blob_data[gi].g;
+    top_blob_data[v_offset.b] = bottom_blob_data[gi].b;
+    top_blob_data[v_offset.a] = bottom_blob_data[gi].a;
 }

src/layer/shader/padding.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -56,8 +59,6 @@ void main()
     if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
-    sfp res;
-
     int x = gx - left;
     int y = gy - top;
 
@@ -66,11 +67,11 @@ void main()
         if (x >= 0 && x < p.w && y >= 0 && y < p.h)
         {
             int v_offset = gz * p.cstep + y * p.w + x;
-            res = bottom_blob_data[v_offset];
+            top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
         }
         else
         {
-            res = sfp(value);
+            top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfp(value);
         }
     }
     else if (type == 1)
@@ -79,8 +80,6 @@ void main()
         y = clamp(y, 0, p.h - 1);
 
         int v_offset = gz * p.cstep + y * p.w + x;
-        res = bottom_blob_data[v_offset];
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
     }
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = res;
 }

src/layer/shader/padding_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -56,8 +59,6 @@ void main()
     if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
-    sfpvec4 res;
-
     int x = gx - left;
     int y = gy - top;
 
@@ -66,11 +67,11 @@ void main()
         if (x >= 0 && x < p.w && y >= 0 && y < p.h)
         {
             int v_offset = gz * p.cstep + y * p.w + x;
-            res = bottom_blob_data[v_offset];
+            top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
         }
         else
         {
-            res = sfpvec4(value);
+            top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfpvec4(value);
         }
     }
     else if (type == 1)
@@ -79,8 +80,6 @@ void main()
         y = clamp(y, 0, p.h - 1);
 
         int v_offset = gz * p.cstep + y * p.w + x;
-        res = bottom_blob_data[v_offset];
+        top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
     }
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = res;
 }

src/layer/shader/permute.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/permute_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -104,10 +107,10 @@ void main()
         }
     }
 
-    sfpvec4 v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+    int gi = gz * p.cstep + gy * p.w + gx;
 
-    top_blob_data[v_offset.r] = v.r;
-    top_blob_data[v_offset.g] = v.g;
-    top_blob_data[v_offset.b] = v.b;
-    top_blob_data[v_offset.a] = v.a;
+    top_blob_data[v_offset.r] = bottom_blob_data[gi].r;
+    top_blob_data[v_offset.g] = bottom_blob_data[gi].g;
+    top_blob_data[v_offset.b] = bottom_blob_data[gi].b;
+    top_blob_data[v_offset.a] = bottom_blob_data[gi].a;
 }

src/layer/shader/pooling.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/pooling_global.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/pooling_global_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/pooling_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/prelu.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -47,9 +50,9 @@ void main()
 
     if (p.dims == 1)
     {
-        afp v = bottom_top_blob_data[gx];
+        afp v = afp(bottom_top_blob_data[gx]);
 
-        afp slope = num_slope > 1 ? slope_blob_data[gx] : slope_blob_data[0];
+        afp slope = num_slope > 1 ? afp(slope_blob_data[gx]) : afp(slope_blob_data[0]);
 
         v = v < afp(0.f) ? v * slope : v;
 
@@ -62,9 +65,9 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
-        afp slope = num_slope > 1 ? slope_blob_data[gy] : slope_blob_data[0];
+        afp slope = num_slope > 1 ? afp(slope_blob_data[gy]) : afp(slope_blob_data[0]);
 
         v = v < afp(0.f) ? v * slope : v;
 
@@ -77,9 +80,9 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
-        afp slope = num_slope > 1 ? slope_blob_data[gz] : slope_blob_data[0];
+        afp slope = num_slope > 1 ? afp(slope_blob_data[gz]) : afp(slope_blob_data[0]);
 
         v = v < afp(0.f) ? v * slope : v;
 

src/layer/shader/prelu_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -47,9 +50,9 @@ void main()
 
     if (p.dims == 1)
     {
-        afpvec4 v = bottom_top_blob_data[gx];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gx]);
 
-        afpvec4 slope = num_slope > 1 ? slope_blob_data[gx] : afpvec4(slope_blob_data[0]);
+        afpvec4 slope = num_slope > 1 ? afpvec4(slope_blob_data[gx]) : afpvec4(slope_blob_data[0]);
 
         v = mix(v, v * slope, lessThan(v, afpvec4(0.f)));
 
@@ -62,9 +65,9 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
-        afpvec4 slope = num_slope > 1 ? slope_blob_data[gy] : afpvec4(slope_blob_data[0]);
+        afpvec4 slope = num_slope > 1 ? afpvec4(slope_blob_data[gy]) : afpvec4(slope_blob_data[0]);
 
         v = mix(v, v * slope, lessThan(v, afpvec4(0.f)));
 
@@ -77,9 +80,9 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
-        afpvec4 slope = num_slope > 1 ? slope_blob_data[gz] : afpvec4(slope_blob_data[0]);
+        afpvec4 slope = num_slope > 1 ? afpvec4(slope_blob_data[gz]) : afpvec4(slope_blob_data[0]);
 
         v = mix(v, v * slope, lessThan(v, afpvec4(0.f)));
 

src/layer/shader/priorbox.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -69,14 +72,12 @@ void main()
 
     afpvec4 image_norm = afp(1.f) / afpvec4(p.image_w, p.image_h, p.image_w, p.image_h);
 
-    sfpvec4 variance = sfpvec4(variances_0, variances_1, variances_2, variances_3);
-
     afpvec4 box;
 
     afp box_w;
     afp box_h;
 
-    afp min_size = min_sizes_data[gx];
+    afp min_size = afp(min_sizes_data[gx]);
 
     // min size box
     box_w = box_h = min_size;
@@ -84,14 +85,17 @@ void main()
     box = (center + afpvec4(-box_w, -box_h, box_w, box_h) * afp(0.5f)) * image_norm;
 
     top_blob_data[v_offset] = clip == 1 ? sfpvec4(clamp(box, afp(0.f), afp(1.f))) : sfpvec4(box);
-    top_blob_data[var_offset] = variance;
+    top_blob_data[var_offset].r = sfp(variances_0);
+    top_blob_data[var_offset].g = sfp(variances_1);
+    top_blob_data[var_offset].b = sfp(variances_2);
+    top_blob_data[var_offset].a = sfp(variances_3);
 
     v_offset += 1;
     var_offset += 1;
 
     if (num_max_size > 0)
     {
-        afp max_size = max_sizes_data[gx];
+        afp max_size = afp(max_sizes_data[gx]);
 
         // max size box
         box_w = box_h = sqrt(min_size * max_size);
@@ -99,7 +103,10 @@ void main()
         box = (center + afpvec4(-box_w, -box_h, box_w, box_h) * afp(0.5f)) * image_norm;
 
         top_blob_data[v_offset] = clip == 1 ? sfpvec4(clamp(box, afp(0.f), afp(1.f))) : sfpvec4(box);
-        top_blob_data[var_offset] = variance;
+        top_blob_data[var_offset].r = sfp(variances_0);
+        top_blob_data[var_offset].g = sfp(variances_1);
+        top_blob_data[var_offset].b = sfp(variances_2);
+        top_blob_data[var_offset].a = sfp(variances_3);
 
         v_offset += 1;
         var_offset += 1;
@@ -108,7 +115,7 @@ void main()
     // all aspect_ratios
     for (int pi = 0; pi < num_aspect_ratio; pi++)
     {
-        afp ar = aspect_ratios_data[pi];
+        afp ar = afp(aspect_ratios_data[pi]);
 
         box_w = min_size * sqrt(ar);
         box_h = min_size / sqrt(ar);
@@ -116,7 +123,10 @@ void main()
         box = (center + afpvec4(-box_w, -box_h, box_w, box_h) * afp(0.5f)) * image_norm;
 
         top_blob_data[v_offset] = clip == 1 ? sfpvec4(clamp(box, afp(0.f), afp(1.f))) : sfpvec4(box);
-        top_blob_data[var_offset] = variance;
+        top_blob_data[var_offset].r = sfp(variances_0);
+        top_blob_data[var_offset].g = sfp(variances_1);
+        top_blob_data[var_offset].b = sfp(variances_2);
+        top_blob_data[var_offset].a = sfp(variances_3);
 
         v_offset += 1;
         var_offset += 1;
@@ -126,7 +136,10 @@ void main()
             box = (center + afpvec4(-box_h, -box_w, box_h, box_w) * afp(0.5f)) * image_norm;
 
             top_blob_data[v_offset] = clip == 1 ? sfpvec4(clamp(box, afp(0.f), afp(1.f))) : sfpvec4(box);
-            top_blob_data[var_offset] = variance;
+            top_blob_data[var_offset].r = sfp(variances_0);
+            top_blob_data[var_offset].g = sfp(variances_1);
+            top_blob_data[var_offset].b = sfp(variances_2);
+            top_blob_data[var_offset].a = sfp(variances_3);
 
             v_offset += 1;
             var_offset += 1;

src/layer/shader/priorbox_mxnet.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -58,7 +61,7 @@ void main()
     afpvec4 center = afpvec4(center_x, center_y, center_x, center_y);
 
     // ratio = 1, various sizes
-    afp size = min_sizes_data[gx];
+    afp size = afp(min_sizes_data[gx]);
     afp cw = size * afp(p.h) / afp(p.w) / afp(2);
     afp ch = size / afp(2);
 
@@ -69,7 +72,7 @@ void main()
     if (gx == num_sizes - 1)
     {
         // various ratios, size = min_size = size[0]
-        afp size = min_sizes_data[0];
+        afp size = afp(min_sizes_data[0]);
         for (int pi = 1; pi < num_ratios; pi++)
         {
             afp ratio = sqrt(afp(aspect_ratios_data[pi]));

src/layer/shader/relu.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     if (slope == 0)
         v = max(v, afp(0.f));

src/layer/shader/relu_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     if (slope == 0)
         v = max(v, afp(0.f));

src/layer/shader/reorg.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -58,7 +61,5 @@ void main()
 
     int v_offset = z * p.cstep + y * p.w + x;
 
-    sfp v = bottom_blob_data[v_offset];
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
 }

src/layer/shader/reorg_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -60,12 +63,10 @@ void main()
 
     ivec4 v_offset = z4 * p.cstep + y4 * p.w + x4;
 
-    sfpvec4 v;
-
-    v.r = bottom_blob_data[v_offset.r];
-    v.g = bottom_blob_data[v_offset.g];
-    v.b = bottom_blob_data[v_offset.b];
-    v.a = bottom_blob_data[v_offset.a];
+    int gi = gz * p.outcstep + gy * p.outw + gx;
 
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    top_blob_data[gi].r = bottom_blob_data[v_offset.r];
+    top_blob_data[gi].g = bottom_blob_data[v_offset.g];
+    top_blob_data[gi].b = bottom_blob_data[v_offset.b];
+    top_blob_data[gi].a = bottom_blob_data[v_offset.a];
 }

src/layer/shader/reorg_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -63,33 +66,25 @@ void main()
     ivec4 lane4 = z4 % 4;
 
     // v = v4[lane]
-    sfpvec4 v;
-
-    sfpvec4 v4;
-
-    v4 = bottom_blob_data[v_offset.r];
-    if (lane4.r == 0) v.r = v4.r;
-    else if (lane4.r == 1) v.r = v4.g;
-    else if (lane4.r == 2) v.r = v4.b;
-    else /* if (lane4.r == 3) */ v.r = v4.a;
-
-    v4 = bottom_blob_data[v_offset.g];
-    if (lane4.g == 0) v.g = v4.r;
-    else if (lane4.g == 1) v.g = v4.g;
-    else if (lane4.g == 2) v.g = v4.b;
-    else /* if (lane4.g == 3) */ v.g = v4.a;
-
-    v4 = bottom_blob_data[v_offset.b];
-    if (lane4.b == 0) v.b = v4.r;
-    else if (lane4.b == 1) v.b = v4.g;
-    else if (lane4.b == 2) v.b = v4.b;
-    else /* if (lane4.b == 3) */ v.b = v4.a;
-
-    v4 = bottom_blob_data[v_offset.a];
-    if (lane4.a == 0) v.a = v4.r;
-    else if (lane4.a == 1) v.a = v4.g;
-    else if (lane4.a == 2) v.a = v4.b;
-    else /* if (lane4.a == 3) */ v.a = v4.a;
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    if (lane4.r == 0) top_blob_data[gi].r = bottom_blob_data[v_offset.r].r;
+    else if (lane4.r == 1) top_blob_data[gi].r = bottom_blob_data[v_offset.r].g;
+    else if (lane4.r == 2) top_blob_data[gi].r = bottom_blob_data[v_offset.r].b;
+    else /* if (lane4.r == 3) */ top_blob_data[gi].r = bottom_blob_data[v_offset.r].a;
+
+    if (lane4.g == 0) top_blob_data[gi].g = bottom_blob_data[v_offset.g].r;
+    else if (lane4.g == 1) top_blob_data[gi].g = bottom_blob_data[v_offset.g].g;
+    else if (lane4.g == 2) top_blob_data[gi].g = bottom_blob_data[v_offset.g].b;
+    else /* if (lane4.g == 3) */ top_blob_data[gi].g = bottom_blob_data[v_offset.g].a;
+
+    if (lane4.b == 0) top_blob_data[gi].b = bottom_blob_data[v_offset.b].r;
+    else if (lane4.b == 1) top_blob_data[gi].b = bottom_blob_data[v_offset.b].g;
+    else if (lane4.b == 2) top_blob_data[gi].b = bottom_blob_data[v_offset.b].b;
+    else /* if (lane4.b == 3) */ top_blob_data[gi].b = bottom_blob_data[v_offset.b].a;
+
+    if (lane4.a == 0) top_blob_data[gi].a = bottom_blob_data[v_offset.a].r;
+    else if (lane4.a == 1) top_blob_data[gi].a = bottom_blob_data[v_offset.a].g;
+    else if (lane4.a == 2) top_blob_data[gi].a = bottom_blob_data[v_offset.a].b;
+    else /* if (lane4.a == 3) */ top_blob_data[gi].a = bottom_blob_data[v_offset.a].a;
 }

src/layer/shader/reshape.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -65,9 +68,19 @@ void main()
 
     int v_offset = z * p.cstep + y * p.w + x;
 
-    sfp v = bottom_blob_data[v_offset];
-
-    if (ndim == 1) top_blob_data[gx] = v;
-    if (ndim == 2) top_blob_data[gy * p.outw + gx] = v;
-    if (ndim == 3) top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    int gi;
+    if (ndim == 1)
+    {
+        gi = gx;
+    }
+    if (ndim == 2)
+    {
+        gi = gy * p.outw + gx;
+    }
+    if (ndim == 3)
+    {
+        gi = gz * p.outcstep + gy * p.outw + gx;
+    }
+
+    top_blob_data[gi] = bottom_blob_data[v_offset];
 }

src/layer/shader/reshape_pack1to4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -65,13 +68,22 @@ void main()
 
     ivec4 v_offset = z4 * p.cstep + y4 * p.w + x4;
 
-    sfpvec4 v;
-    v.r = bottom_blob_data[v_offset.r];
-    v.g = bottom_blob_data[v_offset.g];
-    v.b = bottom_blob_data[v_offset.b];
-    v.a = bottom_blob_data[v_offset.a];
-
-    if (ndim == 1) top_blob_data[gx] = v;
-    if (ndim == 2) top_blob_data[gy * p.outw + gx] = v;
-    if (ndim == 3) top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    int gi;
+    if (ndim == 1)
+    {
+        gi = gx;
+    }
+    if (ndim == 2)
+    {
+        gi = gy * p.outw + gx;
+    }
+    if (ndim == 3)
+    {
+        gi = gz * p.outcstep + gy * p.outw + gx;
+    }
+
+    top_blob_data[gi].r = bottom_blob_data[v_offset.r];
+    top_blob_data[gi].g = bottom_blob_data[v_offset.g];
+    top_blob_data[gi].b = bottom_blob_data[v_offset.b];
+    top_blob_data[gi].a = bottom_blob_data[v_offset.a];
 }

src/layer/shader/reshape_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -85,35 +88,37 @@ void main()
         lane4 = z4 % 4;
     }
 
-    sfpvec4 v;
-
-    sfpvec4 v4;
-
-    v4 = bottom_blob_data[v_offset.r];
-    if (lane4.r == 0) v.r = v4.r;
-    else if (lane4.r == 1) v.r = v4.g;
-    else if (lane4.r == 2) v.r = v4.b;
-    else /* if (lane4.r == 3) */ v.r = v4.a;
-
-    v4 = bottom_blob_data[v_offset.g];
-    if (lane4.g == 0) v.g = v4.r;
-    else if (lane4.g == 1) v.g = v4.g;
-    else if (lane4.g == 2) v.g = v4.b;
-    else /* if (lane4.g == 3) */ v.g = v4.a;
-
-    v4 = bottom_blob_data[v_offset.b];
-    if (lane4.b == 0) v.b = v4.r;
-    else if (lane4.b == 1) v.b = v4.g;
-    else if (lane4.b == 2) v.b = v4.b;
-    else /* if (lane4.b == 3) */ v.b = v4.a;
-
-    v4 = bottom_blob_data[v_offset.a];
-    if (lane4.a == 0) v.a = v4.r;
-    else if (lane4.a == 1) v.a = v4.g;
-    else if (lane4.a == 2) v.a = v4.b;
-    else /* if (lane4.a == 3) */ v.a = v4.a;
-
-    if (ndim == 1) top_blob_data[gx] = v;
-    if (ndim == 2) top_blob_data[gy * p.outw + gx] = v;
-    if (ndim == 3) top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    int gi;
+    if (ndim == 1)
+    {
+        gi = gx;
+    }
+    if (ndim == 2)
+    {
+        gi = gy * p.outw + gx;
+    }
+    if (ndim == 3)
+    {
+        gi = gz * p.outcstep + gy * p.outw + gx;
+    }
+
+    if (lane4.r == 0) top_blob_data[gi].r = bottom_blob_data[v_offset.r].r;
+    else if (lane4.r == 1) top_blob_data[gi].r = bottom_blob_data[v_offset.r].g;
+    else if (lane4.r == 2) top_blob_data[gi].r = bottom_blob_data[v_offset.r].b;
+    else /* if (lane4.r == 3) */ top_blob_data[gi].r = bottom_blob_data[v_offset.r].a;
+
+    if (lane4.g == 0) top_blob_data[gi].g = bottom_blob_data[v_offset.g].r;
+    else if (lane4.g == 1) top_blob_data[gi].g = bottom_blob_data[v_offset.g].g;
+    else if (lane4.g == 2) top_blob_data[gi].g = bottom_blob_data[v_offset.g].b;
+    else /* if (lane4.g == 3) */ top_blob_data[gi].g = bottom_blob_data[v_offset.g].a;
+
+    if (lane4.b == 0) top_blob_data[gi].b = bottom_blob_data[v_offset.b].r;
+    else if (lane4.b == 1) top_blob_data[gi].b = bottom_blob_data[v_offset.b].g;
+    else if (lane4.b == 2) top_blob_data[gi].b = bottom_blob_data[v_offset.b].b;
+    else /* if (lane4.b == 3) */ top_blob_data[gi].b = bottom_blob_data[v_offset.b].a;
+
+    if (lane4.a == 0) top_blob_data[gi].a = bottom_blob_data[v_offset.a].r;
+    else if (lane4.a == 1) top_blob_data[gi].a = bottom_blob_data[v_offset.a].g;
+    else if (lane4.a == 2) top_blob_data[gi].a = bottom_blob_data[v_offset.a].b;
+    else /* if (lane4.a == 3) */ top_blob_data[gi].r = bottom_blob_data[v_offset.a].a;
 }

src/layer/shader/reshape_pack4to1.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -81,10 +84,10 @@ void main()
         v_offset = z4 * p.outcstep + y4 * p.outw + x4;
     }
 
-    sfpvec4 v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+    int gi = gz * p.cstep + gy * p.w + gx;
 
-    top_blob_data[v_offset.r] = v.r;
-    top_blob_data[v_offset.g] = v.g;
-    top_blob_data[v_offset.b] = v.b;
-    top_blob_data[v_offset.a] = v.a;
+    top_blob_data[v_offset.r] = bottom_blob_data[gi].r;
+    top_blob_data[v_offset.g] = bottom_blob_data[gi].g;
+    top_blob_data[v_offset.b] = bottom_blob_data[gi].b;
+    top_blob_data[v_offset.a] = bottom_blob_data[gi].a;
 }

src/layer/shader/scale.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -48,7 +51,7 @@ void main()
 
     if (p.dims == 1)
     {
-        afp v = bottom_top_blob_data[gx];
+        afp v = afp(bottom_top_blob_data[gx]);
 
         if (bias_term == 1)
             v = afp(scale_blob_data[gx]) * v + afp(bias_blob_data[gx]);
@@ -64,7 +67,7 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
         if (bias_term == 1)
             v = afp(scale_blob_data[gy]) * v + afp(bias_blob_data[gy]);
@@ -80,7 +83,7 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afp v = bottom_top_blob_data[gi];
+        afp v = afp(bottom_top_blob_data[gi]);
 
         if (bias_term == 1)
             v = afp(scale_blob_data[gz]) * v + afp(bias_blob_data[gz]);

src/layer/shader/scale_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -48,7 +51,7 @@ void main()
 
     if (p.dims == 1)
     {
-        afpvec4 v = bottom_top_blob_data[gx];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gx]);
 
         if (bias_term == 1)
             v = afpvec4(scale_blob_data[gx]) * v + afpvec4(bias_blob_data[gx]);
@@ -64,7 +67,7 @@ void main()
     {
         const int gi = gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
         if (bias_term == 1)
             v = afpvec4(scale_blob_data[gy]) * v + afpvec4(bias_blob_data[gy]);
@@ -80,7 +83,7 @@ void main()
     {
         const int gi = gz * p.cstep + gy * p.w + gx;
 
-        afpvec4 v = bottom_top_blob_data[gi];
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
         if (bias_term == 1)
             v = afpvec4(scale_blob_data[gz]) * v + afpvec4(bias_blob_data[gz]);

src/layer/shader/shufflechannel.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/shufflechannel_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -62,33 +65,25 @@ void main()
     ivec4 lane4 = z4 % 4;
 
     // v = v4[lane]
-    sfpvec4 v;
-
-    sfpvec4 v4;
-
-    v4 = bottom_blob_data[v_offset.r];
-    if (lane4.r == 0) v.r = v4.r;
-    else if (lane4.r == 1) v.r = v4.g;
-    else if (lane4.r == 2) v.r = v4.b;
-    else /* if (lane4.r == 3) */ v.r = v4.a;
-
-    v4 = bottom_blob_data[v_offset.g];
-    if (lane4.g == 0) v.g = v4.r;
-    else if (lane4.g == 1) v.g = v4.g;
-    else if (lane4.g == 2) v.g = v4.b;
-    else /* if (lane4.g == 3) */ v.g = v4.a;
-
-    v4 = bottom_blob_data[v_offset.b];
-    if (lane4.b == 0) v.b = v4.r;
-    else if (lane4.b == 1) v.b = v4.g;
-    else if (lane4.b == 2) v.b = v4.b;
-    else /* if (lane4.b == 3) */ v.b = v4.a;
-
-    v4 = bottom_blob_data[v_offset.a];
-    if (lane4.a == 0) v.a = v4.r;
-    else if (lane4.a == 1) v.a = v4.g;
-    else if (lane4.a == 2) v.a = v4.b;
-    else /* if (lane4.a == 3) */ v.a = v4.a;
-
-    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    if (lane4.r == 0) top_blob_data[gi].r = bottom_blob_data[v_offset.r].r;
+    else if (lane4.r == 1) top_blob_data[gi].r = bottom_blob_data[v_offset.r].g;
+    else if (lane4.r == 2) top_blob_data[gi].r = bottom_blob_data[v_offset.r].b;
+    else /* if (lane4.r == 3) */ top_blob_data[gi].r = bottom_blob_data[v_offset.r].a;
+
+    if (lane4.g == 0) top_blob_data[gi].g = bottom_blob_data[v_offset.g].r;
+    else if (lane4.g == 1) top_blob_data[gi].g = bottom_blob_data[v_offset.g].g;
+    else if (lane4.g == 2) top_blob_data[gi].g = bottom_blob_data[v_offset.g].b;
+    else /* if (lane4.g == 3) */ top_blob_data[gi].g = bottom_blob_data[v_offset.g].a;
+
+    if (lane4.b == 0) top_blob_data[gi].b = bottom_blob_data[v_offset.b].r;
+    else if (lane4.b == 1) top_blob_data[gi].b = bottom_blob_data[v_offset.b].g;
+    else if (lane4.b == 2) top_blob_data[gi].b = bottom_blob_data[v_offset.b].b;
+    else /* if (lane4.b == 3) */ top_blob_data[gi].b = bottom_blob_data[v_offset.b].a;
+
+    if (lane4.a == 0) top_blob_data[gi].a = bottom_blob_data[v_offset.a].r;
+    else if (lane4.a == 1) top_blob_data[gi].a = bottom_blob_data[v_offset.a].g;
+    else if (lane4.a == 2) top_blob_data[gi].a = bottom_blob_data[v_offset.a].b;
+    else /* if (lane4.a == 3) */ top_blob_data[gi].a = bottom_blob_data[v_offset.a].a;
 }

src/layer/shader/sigmoid.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     v = afp(1.f) / (afp(1.f) + exp(-v));
 

src/layer/shader/sigmoid_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     v = afp(1.f) / (afp(1.f) + exp(-v));
 

src/layer/shader/softmax_div_sum.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -53,8 +56,8 @@ void main()
 
     if (p.dims == 1) // axis == 0
     {
-        afp sum = sum_workspace_data[0];
-        afp v = bottom_top_blob_data[gx];
+        afp sum = afp(sum_workspace_data[0]);
+        afp v = afp(bottom_top_blob_data[gx]);
         bottom_top_blob_data[gx] = sfp(v / sum);
         return;
     }
@@ -62,8 +65,8 @@ void main()
     if (p.dims == 2 && axis == 0)
     {
         int gi = gy * p.w + gx;
-        afp sum = sum_workspace_data[gx];
-        afp v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gx]);
+        afp v = afp(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfp(v / sum);
         return;
     }
@@ -71,8 +74,8 @@ void main()
     if (p.dims == 2 && axis == 1)
     {
         int gi = gy * p.w + gx;
-        afp sum = sum_workspace_data[gy];
-        afp v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gy]);
+        afp v = afp(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfp(v / sum);
         return;
     }
@@ -80,8 +83,8 @@ void main()
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gy * p.w + gx];
-        afp v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gy * p.w + gx]);
+        afp v = afp(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfp(v / sum);
         return;
     }
@@ -89,8 +92,8 @@ void main()
     if (p.dims == 3 && axis == 1)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gz * p.w + gx];
-        afp v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gz * p.w + gx]);
+        afp v = afp(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfp(v / sum);
         return;
     }
@@ -98,8 +101,8 @@ void main()
     if (p.dims == 3 && axis == 2)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gz * p.h + gy];
-        afp v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gz * p.h + gy]);
+        afp v = afp(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfp(v / sum);
         return;
     }

src/layer/shader/softmax_div_sum_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -53,8 +56,8 @@ void main()
 
     if (p.dims == 1) // axis == 0
     {
-        afp sum = sum_workspace_data[0];
-        afpvec4 v = bottom_top_blob_data[gx];
+        afp sum = afp(sum_workspace_data[0]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gx]);
         bottom_top_blob_data[gx] = sfpvec4(v / sum);
         return;
     }
@@ -62,8 +65,8 @@ void main()
     if (p.dims == 2 && axis == 0)
     {
         int gi = gy * p.w + gx;
-        afp sum = sum_workspace_data[gx];
-        afpvec4 v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gx]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfpvec4(v / sum);
         return;
     }
@@ -71,8 +74,8 @@ void main()
     if (p.dims == 2 && axis == 1)
     {
         int gi = gy * p.w + gx;
-        afp sum = sum_workspace_data[gy];
-        afpvec4 v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gy]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfpvec4(v / sum);
         return;
     }
@@ -80,8 +83,8 @@ void main()
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gy * p.w + gx];
-        afpvec4 v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gy * p.w + gx]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfpvec4(v / sum);
         return;
     }
@@ -89,8 +92,8 @@ void main()
     if (p.dims == 3 && axis == 1)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gz * p.w + gx];
-        afpvec4 v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gz * p.w + gx]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfpvec4(v / sum);
         return;
     }
@@ -98,8 +101,8 @@ void main()
     if (p.dims == 3 && axis == 2)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        afp sum = sum_workspace_data[gz * p.h + gy];
-        afpvec4 v = bottom_top_blob_data[gi];
+        afp sum = afp(sum_workspace_data[gz * p.h + gy]);
+        afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
         bottom_top_blob_data[gi] = sfpvec4(v / sum);
         return;
     }

src/layer/shader/softmax_exp_sub_max.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/softmax_exp_sub_max_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/softmax_reduce_max.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/softmax_reduce_max_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/softmax_reduce_sum.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/softmax_reduce_sum_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 

src/layer/shader/tanh.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfp(tanh(v));
 }

src/layer/shader/tanh_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -44,7 +47,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     bottom_top_blob_data[gi] = sfpvec4(tanh(v));
 }

src/layer/shader/unaryop.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afp v = bottom_top_blob_data[gi];
+    afp v = afp(bottom_top_blob_data[gi]);
 
     afp res;
 

src/layer/shader/unaryop_pack4.comp

@@ -14,7 +14,10 @@
 
 #version 450
 
-#if NCNN_fp16_storage || NCNN_fp16_arithmetic
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+#endif
+#if NCNN_fp16_arithmetic
 #extension GL_AMD_gpu_shader_half_float: require
 #endif
 
@@ -46,7 +49,7 @@ void main()
 
     const int gi = gz * p.cstep + gy * p.w + gx;
 
-    afpvec4 v = bottom_top_blob_data[gi];
+    afpvec4 v = afpvec4(bottom_top_blob_data[gi]);
 
     afpvec4 res;