reconcile Dim4D and NchwDim into DimIndex (#3489)

Summary:
Pull Request resolved: #3489

1. Adapt SS-JIA's idea to represent `Dim4D` as a "negative index", and rename it as `DimIndex`
2. Merge `NchwDim`'s functionality with `Dim4D`.
3. Clean up `dim_at` call to assume only `DimIndex` as input.
4. Further clean up some usage of `uint` and convert them into `int`.

ghstack-source-id: 225521662

Reviewed By: SS-JIA

Differential Revision: D56778340

fbshipit-source-id: 68e1fd7d74b59fb89299123263009be096dd5c18

Author: yipjustin

backends/vulkan/runtime/graph/ops/glsl/permute.glsl

@@ -29,9 +29,9 @@ layout(set = 0, binding = 3) uniform PRECISION restrict Sizes {
 
 layout(set = 0, binding = 4) uniform PRECISION restrict Block {
   // output dims
-  uvec4 out_ndims;
+  ivec4 out_ndims;
   // x = output channels aligned to 4, y = input channels aligned to 4
-  uvec2 ch_info;
+  ivec2 ch_info;
 };
 
 layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;

backends/vulkan/runtime/graph/ops/impl/Cat.cpp

@@ -31,10 +31,10 @@ void add_cat_default_node(
   int64_t dim = graph.extract_scalar<int64_t>(dim_ref);
   vTensorPtr t_out = graph.get_tensor(out);
 
-  NchwDim nchw_dim = normalize_to_nchw_dim(*t_out, dim);
+  DimIndex dim_index = normalize_to_dim_index(*t_out, dim);
 
   // TODO: Find ways to factor out the similar code for width, height, and batch
-  if (nchw_dim == DimWidth) {
+  if (dim_index == kWidth4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -46,7 +46,7 @@ void add_cat_default_node(
       dst_offset.data[0] += range.data[0];
     }
 
-  } else if (nchw_dim == DimHeight) {
+  } else if (dim_index == kHeight4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -57,7 +57,7 @@ void add_cat_default_node(
           graph, input_ref, range, src_offset, dst_offset, out);
       dst_offset.data[1] += range.data[1];
     }
-  } else if (nchw_dim == DimBatch) {
+  } else if (dim_index == kBatch4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -68,19 +68,19 @@ void add_cat_default_node(
           graph, input_ref, range, src_offset, dst_offset, out);
       dst_offset.data[2] += range.data[2];
     }
-  } else if (nchw_dim == DimChannel) {
+  } else if (dim_index == kChannel4D) {
     int32_t src_offset = 0;
     int32_t dst_offset = 0;
 
     for (ValueRef input_ref : *input_list) {
       vTensorPtr t_in = graph.get_tensor(input_ref);
-      int32_t range = dim_at<Dim4D::Channel>(t_in->sizes());
+      int32_t range = dim_at(t_in->sizes(), kChannel4D);
       add_copy_channel_offset_node(
           graph, input_ref, range, src_offset, dst_offset, out);
       dst_offset += range;
     }
   } else {
-    VK_THROW("Unexpected value of nchw_dim=", nchw_dim);
+    VK_THROW("Unexpected value of dim_index=", dim_index);
   }
 }
 

backends/vulkan/runtime/graph/ops/impl/Copy.cpp

@@ -92,23 +92,23 @@ void add_copy_channel_offset_node(
   VK_CHECK_COND(t_out->dim() >= 3, "Dst dim should be at least 3");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Channel>(in_sizes) >= src_channel_offset + channel_range,
+      dim_at<kChannel4D>(in_sizes) >= src_channel_offset + channel_range,
       "Src channel (",
       src_channel_offset,
       ") and range (",
       channel_range,
       ") should be less than or equal to input tensor's channel size (",
-      dim_at<Dim4D::Channel>(in_sizes),
+      dim_at<kChannel4D>(in_sizes),
       ")");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Channel>(out_sizes) >= dst_channel_offset + channel_range,
+      dim_at<kChannel4D>(out_sizes) >= dst_channel_offset + channel_range,
       "Dst channel (",
       dst_channel_offset,
       ") and range (",
       channel_range,
       ") should be less than or equal to input tensor's channel size (",
-      dim_at<Dim4D::Channel>(out_sizes),
+      dim_at<kChannel4D>(out_sizes),
       ")");
 
   VK_CHECK_COND(channel_range >= 0, "Channel range must be non-negative");
@@ -121,11 +121,10 @@ void add_copy_channel_offset_node(
   kernel_name.reserve(kShaderNameReserve);
   add_dtype_suffix(kernel_name, *t_out);
 
-  int32_t out_channels = dim_at<Dim4D::Channel>(out_sizes);
+  int32_t out_channels = dim_at<kChannel4D>(out_sizes);
 
   // Copy one batch at a time.
-  for (int batch_idx = 0; batch_idx < dim_at<Dim4D::Batch>(in_sizes);
-       batch_idx++) {
+  for (int batch_idx = 0; batch_idx < dim_at<kBatch4D>(in_sizes); batch_idx++) {
     // Mapping the tensor NCHW coordinates into texture XYZ coordinates
     int32_t dst_first_z = dst_channel_offset / 4;
     int32_t dst_last_z = (dst_channel_offset + channel_range - 1) / 4;
@@ -139,8 +138,8 @@ void add_copy_channel_offset_node(
         0, 0, dst_first_z + batch_idx * api::utils::div_up(out_channels, 4)};
 
     uvec3 global_size{
-        dim_at<Dim4D::Width>(in_sizes),
-        dim_at<Dim4D::Height>(in_sizes),
+        api::utils::safe_downcast<uint32_t>(dim_at<kWidth4D>(in_sizes)),
+        api::utils::safe_downcast<uint32_t>(dim_at<kHeight4D>(in_sizes)),
         api::utils::safe_downcast<uint32_t>(dst_last_z - dst_first_z + 1)};
 
     uvec3 local_size = adaptive_work_group_size(global_size);

backends/vulkan/runtime/graph/ops/impl/Permute.cpp

@@ -17,8 +17,9 @@
 
 namespace vkcompute {
 
+using api::utils::ivec2;
 using api::utils::ivec3;
-using api::utils::uvec2;
+using api::utils::ivec4;
 using api::utils::uvec4;
 
 namespace {
@@ -53,7 +54,7 @@ void add_permute_node(
 
   check_args(*t_in, permute_dims, *t_out);
 
-  uvec4 out_dims{0u, 1u, 2u, 3u};
+  ivec4 out_dims{0, 1, 2, 3};
 
   int64_t out_dim = t_out->dim();
   std::vector<bool> seen(out_dim);
@@ -63,22 +64,22 @@ void add_permute_node(
         !seen[permute_dim], "Argument dim ", permute_dim, "  is repeated");
     seen[permute_dim] = true;
 
-    out_dims.data[(4u - out_dim) + i] = permute_dim + (4u - out_dim);
+    out_dims.data[(4u - out_dim) + i] = permute_dim + (4 - out_dim);
   }
 
   std::string kernel_name = "permute";
   kernel_name.reserve(kShaderNameReserve);
   add_dtype_suffix(kernel_name, *t_out);
 
-  uint32_t out_channels = dim_at<Dim4D::Channel>(t_out->sizes());
-  uint32_t in_channels = dim_at<Dim4D::Channel>(t_in->sizes());
+  int32_t out_channels = dim_at<kChannel4D>(t_out->sizes());
+  int32_t in_channels = dim_at<kChannel4D>(t_in->sizes());
 
-  uint32_t out_c_aligned = api::utils::align_up(out_channels, 4u);
-  uint32_t in_c_aligned = api::utils::align_up(in_channels, 4u);
+  int32_t out_c_aligned = api::utils::align_up(out_channels, 4);
+  int32_t in_c_aligned = api::utils::align_up(in_channels, 4);
 
   const struct Block final {
-    uvec4 out_ndims;
-    uvec2 ch_info;
+    ivec4 out_ndims;
+    ivec2 ch_info;
   } params{
       out_dims,
       {out_c_aligned, in_c_aligned},

backends/vulkan/runtime/graph/ops/impl/Repeat.cpp

@@ -32,23 +32,23 @@ void check_args(
       "Input tensor dim size must be not greater than the repeat argument's size");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Width>(in.sizes()) * dim_at<Dim4D::Width>(repeats) ==
-          dim_at<Dim4D::Width>(out.sizes()),
+      dim_at<kWidth4D>(in.sizes()) * dim_at<kWidth4D>(repeats) ==
+          dim_at<kWidth4D>(out.sizes()),
       "Output's width doesn't match input's width * repeat count");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Height>(in.sizes()) * dim_at<Dim4D::Height>(repeats) ==
-          dim_at<Dim4D::Height>(out.sizes()),
+      dim_at<kHeight4D>(in.sizes()) * dim_at<kHeight4D>(repeats) ==
+          dim_at<kHeight4D>(out.sizes()),
       "Output's height doesn't match input's height * repeat count");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Channel>(in.sizes()) * dim_at<Dim4D::Channel>(repeats) ==
-          dim_at<Dim4D::Channel>(out.sizes()),
+      dim_at<kChannel4D>(in.sizes()) * dim_at<kChannel4D>(repeats) ==
+          dim_at<kChannel4D>(out.sizes()),
       "Output's channel doesn't match input's channel * repeat count");
 
   VK_CHECK_COND(
-      dim_at<Dim4D::Batch>(in.sizes()) * dim_at<Dim4D::Batch>(repeats) ==
-          dim_at<Dim4D::Batch>(out.sizes()),
+      dim_at<kBatch4D>(in.sizes()) * dim_at<kBatch4D>(repeats) ==
+          dim_at<kBatch4D>(out.sizes()),
       "Output's batch doesn't match input's batch * repeat count");
 }
 
@@ -70,13 +70,13 @@ void add_repeat_channel_node(
   const std::vector<int64_t>& in_sizes = t_in->sizes();
 
   int32_t in_width =
-      api::utils::safe_downcast<int32_t>(dim_at<Dim4D::Width>(in_sizes));
+      api::utils::safe_downcast<int32_t>(dim_at<kWidth4D>(in_sizes));
   int32_t in_height =
-      api::utils::safe_downcast<int32_t>(dim_at<Dim4D::Height>(in_sizes));
+      api::utils::safe_downcast<int32_t>(dim_at<kHeight4D>(in_sizes));
   int32_t in_channel =
-      api::utils::safe_downcast<int32_t>(dim_at<Dim4D::Channel>(in_sizes));
+      api::utils::safe_downcast<int32_t>(dim_at<kChannel4D>(in_sizes));
   int32_t in_batch =
-      api::utils::safe_downcast<int32_t>(dim_at<Dim4D::Batch>(in_sizes));
+      api::utils::safe_downcast<int32_t>(dim_at<kBatch4D>(in_sizes));
 
   int32_t out_channel = repeat_channel * in_channel;
 
@@ -142,11 +142,11 @@ void add_repeat_node(
   // dimension, we copy over the input texure to the output. In subsequent
   // dimensions, we read and write from the same tensor.
 
-  if (int64_t channel_repeat = dim_at<Dim4D::Channel>(repeats);
+  if (int64_t channel_repeat = dim_at<kChannel4D>(repeats);
       channel_repeat == 1) {
     // If no repeat, short-cut to a direct copy
-    api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
-    api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
+    api::utils::ivec3 src_offset{0, 0, 0};
+    api::utils::ivec3 dst_offset{0, 0, 0};
 
     add_copy_offset_node(graph, in, running_range, src_offset, dst_offset, out);
 
@@ -156,12 +156,11 @@ void add_repeat_node(
 
   // TODO: refactor width, height, and batch into a common helper function.
   // Width
-  if (int64_t width_repeat = dim_at<Dim4D::Width>(repeats); width_repeat > 1) {
-    api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
+  if (int64_t width_repeat = dim_at<kWidth4D>(repeats); width_repeat > 1) {
+    api::utils::ivec3 src_offset{0, 0, 0};
 
     for (int i = 1; i < width_repeat; ++i) {
-      api::utils::ivec3 dst_offset = api::utils::make_ivec3(
-          {i * dim_at<Dim4D::Width>(in_sizes), 0, 0}, false);
+      api::utils::ivec3 dst_offset{i * dim_at<kWidth4D>(in_sizes), 0, 0};
 
       add_copy_offset_node(
           graph, out, running_range, src_offset, dst_offset, out);
@@ -171,13 +170,11 @@ void add_repeat_node(
   }
 
   // Height
-  if (int64_t height_repeat = dim_at<Dim4D::Height>(repeats);
-      height_repeat > 1) {
-    api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
+  if (int64_t height_repeat = dim_at<kHeight4D>(repeats); height_repeat > 1) {
+    api::utils::ivec3 src_offset{0, 0, 0};
 
     for (int i = 1; i < height_repeat; ++i) {
-      api::utils::ivec3 dst_offset = api::utils::make_ivec3(
-          {0, i * dim_at<Dim4D::Height>(in_sizes), 0}, false);
+      api::utils::ivec3 dst_offset = {0, i * dim_at<kHeight4D>(in_sizes), 0};
 
       add_copy_offset_node(
           graph, out, running_range, src_offset, dst_offset, out);
@@ -187,12 +184,11 @@ void add_repeat_node(
   }
 
   // Batch
-  if (int64_t batch_repeat = dim_at<Dim4D::Batch>(repeats); batch_repeat > 1) {
-    api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
+  if (int64_t batch_repeat = dim_at<kBatch4D>(repeats); batch_repeat > 1) {
+    api::utils::ivec3 src_offset{0, 0, 0};
 
     for (int i = 1; i < batch_repeat; ++i) {
-      api::utils::ivec3 dst_offset =
-          api::utils::make_ivec3({0, 0, i * running_range.data[2]}, false);
+      api::utils::ivec3 dst_offset = {0, 0, i * running_range.data[2]};
 
       add_copy_offset_node(
           graph, out, running_range, src_offset, dst_offset, out);

backends/vulkan/runtime/graph/ops/impl/Slice.cpp

@@ -43,8 +43,7 @@ void add_slice_tensor_out_node(
 
   dim = normalize(dim, t_in->dim());
 
-  // Create a dim value as in the underlying dim is 4-dimension.
-  int64_t nchw_dim = dim + (4 - t_in->dim());
+  DimIndex dim_index = normalize_to_dim_index(*t_in, dim);
 
   std::optional<int64_t> opt_start =
       graph.extract_optional_scalar<int64_t>(opt_start_ref);
@@ -61,7 +60,7 @@ void add_slice_tensor_out_node(
   VK_CHECK_COND((0 <= start) && (start < in_sizes[dim]));
   VK_CHECK_COND((0 <= end) && (end <= in_sizes[dim]));
 
-  if (nchw_dim == 1) {
+  if (dim_index == kChannel4D) {
     // slice by channel
     std::string kernel_name = "slice_channel";
     kernel_name.reserve(kShaderNameReserve);
@@ -93,17 +92,17 @@ void add_slice_tensor_out_node(
     // GPU's coordinate is in x, y, z
     int64_t gpu_dim = -1;
     int64_t stride = 1;
-    if (nchw_dim == 3) {
+    if (dim_index == kWidth4D) {
       gpu_dim = 0; // width: x dimension in gpu
       VK_CHECK_COND(out_sizes[dim] == (1 + (end - start - 1) / step));
-    } else if (nchw_dim == 2) {
+    } else if (dim_index == kHeight4D) {
       gpu_dim = 1; // height: y dimension
       VK_CHECK_COND(out_sizes[dim] == (1 + (end - start - 1) / step));
-    } else if (nchw_dim == 0) {
+    } else if (dim_index == kBatch4D) {
       gpu_dim = 2; // batch: z dimension
 
       // Due to channel packing, each batch value is span over stride planes
-      int64_t n_channels = dim_at<Dim4D::Channel>(in_sizes);
+      int64_t n_channels = dim_at(in_sizes, kChannel4D);
       stride = api::utils::div_up<int64_t>(n_channels, 4ll);
     } else {
       VK_THROW("Unexpected ncwh_dim!");

backends/vulkan/runtime/graph/ops/impl/Split.cpp

@@ -29,7 +29,7 @@ void add_split_with_sizes_default_node(
 
   ValueListPtr out_list = graph.get_value_list(out_list_ref);
 
-  NchwDim nchw_dim = normalize_to_nchw_dim(*t_in, dim);
+  DimIndex dim_index = normalize_to_dim_index(*t_in, dim);
 
   VK_CHECK_COND(out_list->size() == split_sizes.size());
 
@@ -39,10 +39,10 @@ void add_split_with_sizes_default_node(
 
     vTensorPtr t_out = graph.get_tensor(out_ref);
     VK_CHECK_COND(check_memory_layout_is(*t_out, api::kChannelsPacked));
-    VK_CHECK_COND(dim_at(*t_out, nchw_dim) == split_size);
+    VK_CHECK_COND(dim_at(*t_out, dim_index) == split_size);
   }
 
-  if (nchw_dim == DimWidth) {
+  if (dim_index == kWidth4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -55,7 +55,7 @@ void add_split_with_sizes_default_node(
 
       src_offset.data[0] += range.data[0];
     }
-  } else if (nchw_dim == DimHeight) {
+  } else if (dim_index == kHeight4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -66,7 +66,7 @@ void add_split_with_sizes_default_node(
 
       src_offset.data[1] += range.data[1];
     }
-  } else if (nchw_dim == DimBatch) {
+  } else if (dim_index == kBatch4D) {
     api::utils::ivec3 src_offset = api::utils::make_ivec3({0, 0, 0}, false);
     api::utils::ivec3 dst_offset = api::utils::make_ivec3({0, 0, 0}, false);
 
@@ -77,13 +77,13 @@ void add_split_with_sizes_default_node(
 
       src_offset.data[2] += range.data[2];
     }
-  } else if (nchw_dim == DimChannel) {
+  } else if (dim_index == kChannel4D) {
     int32_t src_offset = 0;
     int32_t dst_offset = 0;
 
     for (ValueRef out_ref : *out_list) {
       vTensorPtr t_out = graph.get_tensor(out_ref);
-      int32_t range = dim_at<Dim4D::Channel>(t_out->sizes());
+      int32_t range = dim_at<kChannel4D>(t_out->sizes());
       add_copy_channel_offset_node(
           graph, in, range, src_offset, dst_offset, out_ref);
       src_offset += range;
@@ -122,8 +122,8 @@ void add_split_tensor_node(
   int64_t dim = graph.extract_scalar<int64_t>(dim_ref);
 
   vTensorPtr t_in = graph.get_tensor(in);
-  NchwDim nchw_dim = normalize_to_nchw_dim(*t_in, dim);
-  int64_t size = dim_at(*t_in, nchw_dim);
+  DimIndex dim_index = normalize_to_dim_index(*t_in, dim);
+  int64_t size = dim_at(*t_in, dim_index);
   std::vector<int64_t> split_sizes(size / split_size, split_size);
 
   add_split_with_sizes_default_node(graph, in, split_sizes, dim, out);