[ET-VK] Replace Uniform buffers with push constants for binary op (pytorch#7346)

Pull Request resolved: pytorch#7230

This diff replaces uniform buffers with push constants for binary op in the Vulkan backend of Executorch. The changes include updating the GLSL code to use push constants instead of uniform buffers and updating the C++ code to pass the sizes as push constants to the shader.
ghstack-source-id: 258575398
@exported-using-ghexport

Differential Revision: [D66853542](https://our.internmc.facebook.com/intern/diff/D66853542/)

Co-authored-by: Vivek Trivedi <[email protected]>

Author: pytorchbot

backends/vulkan/runtime/graph/ops/DispatchNode.h

@@ -46,12 +46,15 @@ class PushConstantDataInfo {
     payload_.attr = attr;
   }
 
-  explicit PushConstantDataInfo(const void* data, uint32_t dataLen)
+  explicit PushConstantDataInfo(
+      const void* data,
+      uint32_t dataLen,
+      uint32_t pushConstantLen = 0)
       : tensorUniformData(nullptr) {
     VK_CHECK_COND(
         dataLen <= 16, "Single push constant data size must be <= 16 bytes");
-    payload_.dataSize = dataLen;
-    memcpy(payload_.data, data, payload_.dataSize);
+    payload_.dataSize = pushConstantLen ? pushConstantLen : dataLen;
+    memcpy(payload_.data, data, dataLen);
   }
 
   /*

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

@@ -19,11 +19,6 @@ layout(std430) buffer;
 ${layout_declare_tensor(B, "w", "t_out", DTYPE, STORAGE)}
 ${layout_declare_tensor(B, "r", "t_in", DTYPE, STORAGE)}
 ${layout_declare_tensor(B, "r", "t_other", DTYPE, STORAGE)}
-${layout_declare_ubo(B, "ivec4", "out_sizes")}
-${layout_declare_ubo(B, "ivec4", "in_sizes")}
-${layout_declare_ubo(B, "ivec4", "other_sizes")}
-${layout_declare_ubo(B, "ivec2", "broadcast_params")}
-${layout_declare_ubo(B, "float", "alpha")}
 
 #include "broadcasting_utils.h"
 #include "indexing_utils.h"
@@ -40,6 +35,14 @@ const lowp ivec4 in_axis_map = unhash_axis_map(in_layout);
 ${layout_declare_spec_const(C, "int", "other_layout", "DEFAULT_LAYOUT")}
 const lowp ivec4 other_axis_map = unhash_axis_map(other_layout);
 
+layout(push_constant) uniform restrict Block {
+  ivec4 out_sizes;
+  ivec4 in_sizes;
+  ivec4 other_sizes;
+  ivec2 broadcast_params;
+  float alpha;
+};
+
 void main() {
   const ivec3 lpos = ivec3(gl_GlobalInvocationID);
   const ivec4 tidx = lpos_to_tidx(lpos, out_sizes, out_axis_map.w, packed_dim);

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

@@ -67,7 +67,10 @@ void add_binary_op_node(
     alpha_val = graph.extract_scalar<float>(alpha);
   }
 
-  const utils::ivec2 broadcast_params = create_broadcast_params(*t_in1, *t_in2);
+  const struct BinaryOpsParams {
+    const utils::ivec2 broadcast_params;
+    const float alpha_val;
+  } binary_ops_params{create_broadcast_params(*t_in1, *t_in2), alpha_val};
 
   std::string kernel_name("binary_");
   kernel_name.reserve(kShaderNameReserve);
@@ -83,16 +86,16 @@ void add_binary_op_node(
       {{out, vkapi::MemoryAccessType::WRITE},
        {{arg1, arg2}, vkapi::MemoryAccessType::READ}},
       // Shader params buffers
-      {t_out->sizes_ubo(),
-       t_in1->sizes_ubo(),
-       t_in2->sizes_ubo(),
-       graph.create_params_buffer(broadcast_params),
-       graph.create_params_buffer(alpha_val)},
+      {},
       // Specialization Constants
       {t_out->hashed_layout(), t_in1->hashed_layout(), t_in2->hashed_layout()},
       // Resizing Logic
       resize_binary_op_node,
-      {}));
+      {},
+      {{graph.sizes_pc_of(out),
+        graph.sizes_pc_of(arg1),
+        graph.sizes_pc_of(arg2),
+        PushConstantDataInfo(&binary_ops_params, sizeof(binary_ops_params))}}));
 }
 
 #define DEFINE_BINARY_OP_WITH_ALPHA_FN(op_name)                          \

backends/vulkan/test/vulkan_compute_api_test.cpp

@@ -1601,9 +1601,7 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
   auto addFn = VK_GET_OP_FN("aten.add.Tensor");
   addFn(graph, {a.value, b.value, kDummyValueRef, c});
 
-  // +2: alpha UBO, broadcast UBO for arithmetic shader
-  // +1: t.sizes_ubo() for arithmetic shader output c
-  expected_vma_allocation_count += 3;
+  // no new allocations if binary op uses push constants
   EXPECT_EQ(get_vma_allocation_count(), expected_vma_allocation_count);
 
   IOValueRef d = graph.add_input_tensor(
@@ -1624,17 +1622,16 @@ TEST(VulkanComputeGraphTest, test_simple_shared_objects_with_resize) {
   auto mulFn = VK_GET_OP_FN("aten.mul.Tensor");
   mulFn(graph, {c, d.value, e});
 
-  // +2: alpha UBO, broadcast UBO for arithmetic shader
-  // +1: t.sizes_ubo() for arithmetic shader output e
-  expected_vma_allocation_count += 3;
+  // no new allocations if binary op uses push constants
   EXPECT_EQ(get_vma_allocation_count(), expected_vma_allocation_count);
 
   IOValueRef out = {};
   out.value = e;
   out.staging = graph.set_output_tensor(out.value);
 
+  // +1: staging buffer input tensor
   // +1: staging buffer for the output tensor
-  expected_vma_allocation_count += 1;
+  expected_vma_allocation_count += 2;
   EXPECT_EQ(get_vma_allocation_count(), expected_vma_allocation_count);
 
   graph.prepare();