discover VK_KHR_shader_non_semantic_info, checked convolution imagestore (#5955)

.github/workflows/test-coverage.yml

@@ -252,7 +252,6 @@ jobs:
         cd build
         lcov -d ./src -c -o lcov.info
         lcov -r lcov.info '/usr/*' -o lcov.info
-        lcov -r lcov.info '*/install/*' -o lcov.info
         lcov -r lcov.info '*/build/*' -o lcov.info
         lcov --list lcov.info
 

src/gpu.cpp

@@ -334,6 +334,7 @@ public:
     int support_VK_KHR_sampler_ycbcr_conversion;
     int support_VK_KHR_shader_float16_int8;
     int support_VK_KHR_shader_float_controls;
+    int support_VK_KHR_shader_non_semantic_info;
     int support_VK_KHR_shader_subgroup_extended_types;
     int support_VK_KHR_shader_subgroup_rotate;
     int support_VK_KHR_storage_buffer_storage_class;
@@ -670,6 +671,7 @@ int GpuInfoPrivate::query_extensions()
     support_VK_KHR_sampler_ycbcr_conversion = 0;
     support_VK_KHR_shader_float16_int8 = 0;
     support_VK_KHR_shader_float_controls = 0;
+    support_VK_KHR_shader_non_semantic_info = 0;
     support_VK_KHR_shader_subgroup_extended_types = 0;
     support_VK_KHR_shader_subgroup_rotate = 0;
     support_VK_KHR_storage_buffer_storage_class = 0;
@@ -733,6 +735,8 @@ int GpuInfoPrivate::query_extensions()
             support_VK_KHR_shader_float16_int8 = exp.specVersion;
         else if (strcmp(exp.extensionName, "VK_KHR_shader_float_controls") == 0)
             support_VK_KHR_shader_float_controls = exp.specVersion;
+        else if (strcmp(exp.extensionName, "VK_KHR_shader_non_semantic_info") == 0)
+            support_VK_KHR_shader_non_semantic_info = exp.specVersion;
         else if (strcmp(exp.extensionName, "VK_KHR_shader_subgroup_extended_types") == 0)
             support_VK_KHR_shader_subgroup_extended_types = exp.specVersion;
         else if (strcmp(exp.extensionName, "VK_KHR_shader_subgroup_rotate") == 0)
@@ -1552,6 +1556,11 @@ int GpuInfo::support_VK_KHR_shader_float_controls() const
     return d->support_VK_KHR_shader_float_controls;
 }
 
+int GpuInfo::support_VK_KHR_shader_non_semantic_info() const
+{
+    return d->support_VK_KHR_shader_non_semantic_info;
+}
+
 int GpuInfo::support_VK_KHR_shader_subgroup_extended_types() const
 {
     return d->support_VK_KHR_shader_subgroup_extended_types;
@@ -2716,6 +2725,8 @@ VulkanDevice::VulkanDevice(int device_index)
         enabledExtensions.push_back("VK_KHR_shader_float16_int8");
     if (info.support_VK_KHR_shader_float_controls())
         enabledExtensions.push_back("VK_KHR_shader_float_controls");
+    if (info.support_VK_KHR_shader_non_semantic_info())
+        enabledExtensions.push_back("VK_KHR_shader_non_semantic_info");
     if (info.support_VK_KHR_shader_subgroup_extended_types())
         enabledExtensions.push_back("VK_KHR_shader_subgroup_extended_types");
     if (info.support_VK_KHR_shader_subgroup_rotate())

src/gpu.h

@@ -317,6 +317,7 @@ public:
     int support_VK_KHR_sampler_ycbcr_conversion() const;
     int support_VK_KHR_shader_float16_int8() const;
     int support_VK_KHR_shader_float_controls() const;
+    int support_VK_KHR_shader_non_semantic_info() const;
     int support_VK_KHR_shader_subgroup_extended_types() const;
     int support_VK_KHR_shader_subgroup_rotate() const;
     int support_VK_KHR_storage_buffer_storage_class() const;

src/layer/vulkan/shader/convolution.comp

@@ -202,13 +202,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack1to4.comp

@@ -202,13 +202,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack1to8.comp

@@ -220,13 +220,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack4.comp

@@ -233,13 +233,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack4to1.comp

@@ -202,13 +202,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack4to8.comp

@@ -348,13 +348,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack8.comp

@@ -348,13 +348,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack8to1.comp

@@ -204,13 +204,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st1(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st1(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st1(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
 

src/layer/vulkan/shader/convolution_pack8to4.comp

@@ -264,13 +264,16 @@ void main()
 
 #if NCNN_image_shader
     image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
-    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
-    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
+    if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
+    if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
+    if (gz2.y < psc(outc))
+    {
+        image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
+        if (gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
+        if (gy2.y < psc(outh)) image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
+        if (gy2.y < psc(outh) && gx2.y < psc(outw)) image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
+    }
 #else
     const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;