Update on "[ET-VK] Simplifying conv1d op shader by changing it to process one output texel per thread."

This diff changes conv1d shader to process one output texel per thread, increasing GPU occupancy and improve performance.

Differential Revision: [D74097560](https://our.internmc.facebook.com/intern/diff/D74097560/)

[ghstack-poisoned]

Author: trivedivivek

backends/vulkan/runtime/graph/ops/glsl/conv1d.glsl

@@ -59,10 +59,6 @@ const lowp ivec4 bias_axis_map = unhash_axis_map(bias_layout);
 // This implementation performs N x out_C x out_L shader invocations, where each invocation
 // calculates the rolling kernel of the length dimension for each batch, i.e.,
 // computes out_L results.
-//
-// Note that we can rewrite this implementation as out_L * out_C * ceil(N / 4)
-// shader invocations, where each invocation computes 1 result. But that
-// performs worse.
 void main() {
   const ivec3 lpos = ivec3(gl_GlobalInvocationID);
 

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

@@ -520,7 +520,7 @@ void add_conv1d_node(
       // out channels
       static_cast<uint32_t>(out_channels),
       // out batches
-      graph.size_at<uint32_t>(-3, out)};
+      utils::div_up_4(graph.size_at<uint32_t>(-3, out))};
   const utils::uvec3 local_size = graph.create_local_wg_size(global_size);
 
   Kernel1dParams kernel_params = {