Update on "[ET-VK][Ops] aten.var.dim from scratch implementation"

Created the var.dim operator (which functionally supports var) from scratch

Differential Revision: [D75244137](https://our.internmc.facebook.com/intern/diff/D75244137/)

[ghstack-poisoned]

Author: morelos

backends/vulkan/runtime/graph/ops/glsl/var.glsl renamed to backends/vulkan/runtime/graph/ops/glsl/var_texture3d.glsl

@@ -43,8 +43,8 @@ layout(constant_id = 5) const int group_dim = 1;
 // work group will write into its assigned element in the shared array.
 #define MAX_NTHREADS 16
 
-shared vec4 shared_sum[MAX_NTHREADS];
-shared vec4 shared_sum_sq[MAX_NTHREADS];
+shared VEC4_T shared_sum[MAX_NTHREADS];
+shared VEC4_T shared_sum_sq[MAX_NTHREADS];
 shared int shared_count[MAX_NTHREADS];
 
 #include "indexing_utils.h"
@@ -53,9 +53,9 @@ int tid_to_smi(const ivec2 tid) {
   return tid.x + tid.y * NWORKERS;
 }
 
-vec4 calculate_variance(vec4 sum, vec4 sum_sq, int count) {
-  vec4 mean = sum / float(count);
-  vec4 variance = (sum_sq / float(count)) - (mean * mean);
+VEC4_T calculate_variance(VEC4_T sum, VEC4_T sum_sq, int count) {
+  VEC4_T mean = sum / float(count);
+  VEC4_T variance = (sum_sq / float(count)) - (mean * mean);
 
   if ((pc.unbiased != 0) && (count > 1)) {
     variance = variance * (float(count) / float(count - 1.0));
@@ -68,14 +68,14 @@ void reduce_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
   // shared memory index of this thread
   const int smi = tid_to_smi(tid);
 
-  vec4 sum = VEC4_T(0);
-  vec4 sum_sq = VEC4_T(0);
+  VEC4_T sum = VEC4_T(0);
+  VEC4_T sum_sq = VEC4_T(0);
   int count = 0;
 
   scan_pos[reduce_dim] = tid.x;
   for (int i = tid.x; i < tin_sizes[reduce_dim];
        i += NWORKERS, scan_pos[reduce_dim] += NWORKERS) {
-    vec4 val = load_texel(tin, scan_pos);
+    VEC4_T val = load_texel(tin, scan_pos);
     sum += val;
     sum_sq += val * val;
     count += 1;
@@ -109,7 +109,7 @@ void reduce_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
     const bool is_last_texel =
         scan_pos[packed_dim] == (tin_limits[packed_dim] - 1);
 
-    vec4 variance = calculate_variance(sum, sum_sq, count);
+    VEC4_T variance = calculate_variance(sum, sum_sq, count);
 
     // Explicitly set padding elements to 0
     if (is_last_texel && nspill > 0) {
@@ -141,16 +141,16 @@ void reduce_packed_dim(const ivec2 tid, ivec3 scan_pos) {
   // handled specially if it has padding elements.
   const int reduce_len = tin_sizes[packed_dim] - nspill;
 
-  vec4 sum = VEC4_T(0);
-  vec4 sum_sq = VEC4_T(0);
+  VEC4_T sum = VEC4_T(0);
+  VEC4_T sum_sq = VEC4_T(0);
   int count = 0;
 
   // Partially accumulate over elements i, i + NWORKERS, i + 2*NWORKERS, ... of
   // the reduction row
   scan_pos[reduce_dim] = tid.x;
   for (int i = tid.x * 4; i < reduce_len;
        i += NWORKERS * 4, scan_pos[reduce_dim] += NWORKERS) {
-    vec4 val = load_texel(tin, scan_pos);
+    VEC4_T val = load_texel(tin, scan_pos);
     sum += val;
     sum_sq += val * val;
     count += 4;
@@ -159,7 +159,7 @@ void reduce_packed_dim(const ivec2 tid, ivec3 scan_pos) {
   // element of the texel needs to be processed individually such that the
   // padding elements are ignored
   if (scan_pos[reduce_dim] == tin_limits[reduce_dim] - 1 && nspill > 0) {
-    const vec4 val = load_texel(tin, scan_pos);
+    const VEC4_T val = load_texel(tin, scan_pos);
     for (int i = 0; i < nspill; i++) {
       sum.x += val[i];
       sum_sq.x += val[i] * val[i];
@@ -198,7 +198,7 @@ void reduce_packed_dim(const ivec2 tid, ivec3 scan_pos) {
     }
 
     scan_pos[reduce_dim] = tid.x;
-    write_texel(tout, scan_pos, vec4(variance, 0, 0, 0));
+    write_texel(tout, scan_pos, VEC4_T(variance, 0, 0, 0));
   }
 }
 

backends/vulkan/runtime/graph/ops/glsl/var.yaml renamed to backends/vulkan/runtime/graph/ops/glsl/var_texture3d.yaml

@@ -4,7 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-var:
+var_texture3d:
   parameter_names_with_default_values:
     DTYPE: float
     STORAGE: texture3d
@@ -13,4 +13,4 @@ var:
       - VALUE: half
       - VALUE: float
   shader_variants:
-    - NAME: var
+    - NAME: var_texture3d

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

@@ -39,8 +39,9 @@ void add_var_buffer_node(
   int32_t reduce_dim = normalize(dim, ndim);
   reduce_dim = nchw_dim_to_whcn_dim(reduce_dim, ndim);
 
-  std::string kernel_name = "var_buffer";
+  std::string kernel_name = "var";
   kernel_name.reserve(kShaderNameReserve);
+  add_storage_type_suffix(kernel_name, graph.storage_type_of(out));
   add_dtype_suffix(kernel_name, graph.dtype_of(out));
 
   const uint32_t nworkers_per_group = 4;
@@ -56,7 +57,8 @@ void add_var_buffer_node(
   std::vector<PushConstantDataInfo> push_constants;
   int32_t unbiased_int = static_cast<int32_t>(unbiased);
   push_constants.emplace_back(
-      PushConstantDataInfo(&unbiased_int, sizeof(unbiased_int)));
+      PushConstantDataInfo(
+        &unbiased_int, sizeof(unbiased_int)));
 
   graph.execute_nodes().emplace_back(new DispatchNode(
       graph,
@@ -103,6 +105,7 @@ void add_var_texture_node(
 
   std::string kernel_name = "var";
   kernel_name.reserve(kShaderNameReserve);
+  add_storage_type_suffix(kernel_name, graph.storage_type_of(out));
   add_dtype_suffix(kernel_name, graph.dtype_of(out));
 
   // This should match the value of MAX_NTHREADS in the softmax shader.
@@ -131,7 +134,8 @@ void add_var_texture_node(
   std::vector<PushConstantDataInfo> push_constants;
   int32_t unbiased_int = static_cast<int32_t>(unbiased);
   push_constants.emplace_back(
-      PushConstantDataInfo(&unbiased_int, sizeof(unbiased_int)));
+      PushConstantDataInfo(
+        &unbiased_int, sizeof(unbiased_int)));
 
   graph.execute_nodes().emplace_back(new DispatchNode(
       graph,