priorbox shader fix

src/layer/vulkan/priorbox_vulkan.cpp

@@ -30,6 +30,32 @@ PriorBox_vulkan::PriorBox_vulkan()
 
 int PriorBox_vulkan::create_pipeline(const Option& opt)
 {
+    const Mat& shape = bottom_shapes.empty() ? Mat() : bottom_shapes[0];
+
+    int elempack = 1;
+    if (shape.dims == 1) elempack = opt.use_shader_pack8 && shape.w % 8 == 0 ? 8 : shape.w % 4 == 0 ? 4 : 1;
+    if (shape.dims == 2) elempack = opt.use_shader_pack8 && shape.h % 8 == 0 ? 8 : shape.h % 4 == 0 ? 4 : 1;
+    if (shape.dims == 3) elempack = opt.use_shader_pack8 && shape.c % 8 == 0 ? 8 : shape.c % 4 == 0 ? 4 : 1;
+
+    size_t elemsize;
+    if (opt.use_fp16_storage)
+    {
+        elemsize = elempack * 2u;
+    }
+    else if (opt.use_fp16_packed)
+    {
+        elemsize = elempack == 1 ? 4u : elempack * 2u;
+    }
+    else
+    {
+        elemsize = elempack * 4u;
+    }
+
+    Mat shape_packed;
+    if (shape.dims == 1) shape_packed = Mat(shape.w / elempack, (void*)0, elemsize, elempack);
+    if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
+    if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
+
     // caffe style
     {
         int num_min_size = min_sizes.w;
@@ -40,7 +66,7 @@ int PriorBox_vulkan::create_pipeline(const Option& opt)
         if (flip)
             num_prior += num_min_size * num_aspect_ratio;
 
-        std::vector<vk_specialization_type> specializations(11);
+        std::vector<vk_specialization_type> specializations(11 + 2);
         specializations[0].i = flip;
         specializations[1].i = clip;
         specializations[2].f = offset;
@@ -52,6 +78,8 @@ int PriorBox_vulkan::create_pipeline(const Option& opt)
         specializations[8].i = num_max_size;
         specializations[9].i = num_aspect_ratio;
         specializations[10].i = num_prior;
+        specializations[11 + 0].i = shape_packed.w;
+        specializations[11 + 1].i = shape_packed.h;
 
         pipeline_priorbox = new Pipeline(vkdev);
         pipeline_priorbox->set_optimal_local_size_xyz();
@@ -65,12 +93,14 @@ int PriorBox_vulkan::create_pipeline(const Option& opt)
 
         int num_prior = num_sizes - 1 + num_ratios;
 
-        std::vector<vk_specialization_type> specializations(5);
+        std::vector<vk_specialization_type> specializations(5 + 2);
         specializations[0].i = clip;
         specializations[1].f = offset;
         specializations[2].i = num_sizes;
         specializations[3].i = num_ratios;
         specializations[4].i = num_prior;
+        specializations[5 + 0].i = shape_packed.w;
+        specializations[5 + 1].i = shape_packed.h;
 
         pipeline_priorbox_mxnet = new Pipeline(vkdev);
         pipeline_priorbox_mxnet->set_optimal_local_size_xyz();
@@ -107,12 +137,6 @@ int PriorBox_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector
 {
     int w = bottom_blobs[0].w;
     int h = bottom_blobs[0].h;
-    size_t elemsize = 4u;
-
-    if (opt.use_fp16_storage)
-    {
-        elemsize = 2u;
-    }
 
     if (bottom_blobs.size() == 1 && image_width == -233 && image_height == -233 && max_sizes.empty())
     {
@@ -129,8 +153,16 @@ int PriorBox_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector
 
         int num_prior = num_sizes - 1 + num_ratios;
 
+        int elempack = 4;
+
+        size_t elemsize = elempack * 4u;
+        if (opt.use_fp16_packed || opt.use_fp16_storage)
+        {
+            elemsize = elempack * 2u;
+        }
+
         VkMat& top_blob = top_blobs[0];
-        top_blob.create(4 * w * h * num_prior, elemsize, 1, opt.blob_vkallocator, opt.staging_vkallocator);
+        top_blob.create(4 * w * h * num_prior / elempack, elemsize, elempack, opt.blob_vkallocator, opt.staging_vkallocator);
         if (top_blob.empty())
             return -100;
 
@@ -177,6 +209,12 @@ int PriorBox_vulkan::forward(const std::vector<VkMat>& bottom_blobs, std::vector
     if (flip)
         num_prior += num_min_size * num_aspect_ratio;
 
+    size_t elemsize = 4u;
+    if (opt.use_fp16_storage)
+    {
+        elemsize = 2u;
+    }
+
     VkMat& top_blob = top_blobs[0];
     top_blob.create(4 * w * h * num_prior, 2, elemsize, 1, opt.blob_vkallocator, opt.staging_vkallocator);
     if (top_blob.empty())

src/layer/vulkan/shader/priorbox.comp

@@ -87,12 +87,7 @@ void main()
 
     afp min_size = buffer_ld1(min_sizes_data, gx);
 
-#if NCNN_fp16_packed
-    vec4 variances = vec4(variances_0, variances_1, variances_2, variances_3);
-#elif !NCNN_fp16_storage
-    // per component assignment makes qcom-adreno driver unhappy :(
-    sfpvec4 variances = sfpvec4(variances_0, variances_1, variances_2, variances_3);
-#endif
+    afpvec4 variances = afpvec4(variances_0, variances_1, variances_2, variances_3);
 
     // min size box
     box_w = box_h = min_size;
@@ -102,25 +97,10 @@ void main()
 
 #if NCNN_fp16_packed
     top_blob_data[v_offset] = vec4(box);
+    top_blob_data[var_offset] = vec4(variances);
 #else
     buffer_st4(top_blob_data, v_offset, box);
-#endif
-#if NCNN_fp16_packed || !NCNN_fp16_storage
-    top_blob_data[var_offset] = variances;
-#elif NCNN_fp16_storage
-    float variances_0_fp32 = variances_0;
-    float variances_1_fp32 = variances_1;
-    float variances_2_fp32 = variances_2;
-    float variances_3_fp32 = variances_3;
-    top_blob_data[var_offset].r = sfp(variances_0_fp32);
-    top_blob_data[var_offset].g = sfp(variances_1_fp32);
-    top_blob_data[var_offset].b = sfp(variances_2_fp32);
-    top_blob_data[var_offset].a = sfp(variances_3_fp32);
-#else
-    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);
+    buffer_st4(top_blob_data, var_offset, variances);
 #endif
 
     v_offset += 1;
@@ -138,26 +118,12 @@ void main()
 
 #if NCNN_fp16_packed
         top_blob_data[v_offset] = vec4(box);
+        top_blob_data[var_offset] = vec4(variances);
 #else
         buffer_st4(top_blob_data, v_offset, box);
+        buffer_st4(top_blob_data, var_offset, variances);
 #endif
-#if NCNN_fp16_packed || !NCNN_fp16_storage
-        top_blob_data[var_offset] = variances;
-#elif NCNN_fp16_storage
-        float variances_0_fp32 = variances_0;
-        float variances_1_fp32 = variances_1;
-        float variances_2_fp32 = variances_2;
-        float variances_3_fp32 = variances_3;
-        top_blob_data[var_offset].r = sfp(variances_0_fp32);
-        top_blob_data[var_offset].g = sfp(variances_1_fp32);
-        top_blob_data[var_offset].b = sfp(variances_2_fp32);
-        top_blob_data[var_offset].a = sfp(variances_3_fp32);
-#else
-        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);
-#endif
+
         v_offset += 1;
         var_offset += 1;
     }
@@ -175,25 +141,10 @@ void main()
 
 #if NCNN_fp16_packed
         top_blob_data[v_offset] = vec4(box);
+        top_blob_data[var_offset] = vec4(variances);
 #else
         buffer_st4(top_blob_data, v_offset, box);
-#endif
-#if NCNN_fp16_packed || !NCNN_fp16_storage
-        top_blob_data[var_offset] = variances;
-#elif NCNN_fp16_storage
-        float variances_0_fp32 = variances_0;
-        float variances_1_fp32 = variances_1;
-        float variances_2_fp32 = variances_2;
-        float variances_3_fp32 = variances_3;
-        top_blob_data[var_offset].r = sfp(variances_0_fp32);
-        top_blob_data[var_offset].g = sfp(variances_1_fp32);
-        top_blob_data[var_offset].b = sfp(variances_2_fp32);
-        top_blob_data[var_offset].a = sfp(variances_3_fp32);
-#else
-        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);
+        buffer_st4(top_blob_data, var_offset, variances);
 #endif
 
         v_offset += 1;
@@ -206,25 +157,10 @@ void main()
 
 #if NCNN_fp16_packed
             top_blob_data[v_offset] = vec4(box);
+            top_blob_data[var_offset] = vec4(variances);
 #else
             buffer_st4(top_blob_data, v_offset, box);
-#endif
-#if NCNN_fp16_packed || !NCNN_fp16_storage
-            top_blob_data[var_offset] = variances;
-#elif NCNN_fp16_storage
-            float variances_0_fp32 = variances_0;
-            float variances_1_fp32 = variances_1;
-            float variances_2_fp32 = variances_2;
-            float variances_3_fp32 = variances_3;
-            top_blob_data[var_offset].r = sfp(variances_0_fp32);
-            top_blob_data[var_offset].g = sfp(variances_1_fp32);
-            top_blob_data[var_offset].b = sfp(variances_2_fp32);
-            top_blob_data[var_offset].a = sfp(variances_3_fp32);
-#else
-            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);
+            buffer_st4(top_blob_data, var_offset, variances);
 #endif
 
             v_offset += 1;

src/layer/vulkan/shader/priorbox_mxnet.comp

@@ -66,29 +66,27 @@ void main()
 
     // ratio = 1, various sizes
     afp size = buffer_ld1(min_sizes_data, gx);
-    afp cw = size * afp(psc(h)) / afp(psc(w)) / afp(2);
-    afp ch = size / afp(2);
+    afp cw = size * afp(0.5f) * afp(psc(h)) / afp(psc(w));
+    afp ch = size * afp(0.5f);
 
     afpvec4 box = center + afpvec4(-cw, -ch, cw, ch);
     box = clip == 1 ? clamp(box, afp(0.f), afp(1.f)) : box;
 
     buffer_st4(top_blob_data, v_offset, box);
 
-    if (gx == num_sizes - 1)
+    if (gx == 0)
     {
         // various ratios, size = min_size = size[0]
-        afp size = buffer_ld1(min_sizes_data, 0);
         for (int pi = 1; pi < num_ratios; pi++)
         {
             afp v = buffer_ld1(aspect_ratios_data, pi);
-            afp ratio = sqrt(v);
-            afp cwr = size * afp(psc(h)) / afp(psc(w)) * ratio / afp(2);
-            afp chr = size / ratio / afp(2);
+            afp cwr = cw * sqrt(v);
+            afp chr = ch / sqrt(v);
 
             afpvec4 box = center + afpvec4(-cwr, -chr, cwr, chr);
             box = clip == 1 ? clamp(box, afp(0.f), afp(1.f)) : box;
 
-            buffer_st4(top_blob_data, v_offset + pi, box);
+            buffer_st4(top_blob_data, v_offset + num_sizes - 1 + pi, box);
         }
     }
 }

src/pipeline.cpp

@@ -347,7 +347,7 @@ int Pipeline::create_pipeline(VkShaderModule shader_module, const char* entry_na
     VkResult ret = vkCreateComputePipelines(vkdev->vkdevice(), 0, 1, &computePipelineCreateInfo, 0, &pipeline);
     if (ret != VK_SUCCESS)
     {
-        fprintf(stderr, "vkCreateComputePipelines failed %d\n", ret);
+        fprintf(stderr, "vkCreateComputePipelines failed %d %s\n", ret, entry_name);
         return -1;
     }