QConv: Use buffer storage for weights

Summary: TSIA!

Differential Revision: D88701731

Author: SS-JIA

backends/vulkan/runtime/graph/ops/glsl/conv2d_q8_utils.glslh

@@ -90,7 +90,7 @@ ivec4 load_weight_block(
     const int Kw,
     const int Kh,
     const int OC4) {
-#ifdef PACKED_INT8_WEIGHTS_BUFFER
+#ifdef WEIGHT_BUFFER
   const int block_x = oc4 * Kw + kx;
   const int block_y = ky * IC4 + ic4;
   return t_packed_int8_weight[block_y * (Kw * OC4) + block_x];

backends/vulkan/runtime/graph/ops/glsl/conv2d_q8ta_q8csw_q8to.yaml

@@ -13,7 +13,9 @@ conv2d_q8ta_q8csw_q8to:
     combination:
       parameter_names: [IO_STORAGE, WEIGHT_STORAGE]
       combos:
+        - parameter_values: [buffer, buffer]
         - parameter_values: [buffer, texture2d]
+        - parameter_values: [texture3d, buffer]
         - parameter_values: [texture3d, texture2d]
     DTYPE:
       - VALUE: float

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

@@ -442,11 +442,7 @@ ValueRef prepack_quantized_conv2d_weight(
 
   std::vector<int64_t> packed_weight_sizes{output_height, output_width};
 
-  utils::StorageType storage_type = utils::kTexture2D;
-  uint32_t max_extent = graph.context()->adapter_ptr()->max_texture2d_dim();
-  if (output_width > max_extent * 4 || output_height > max_extent) {
-    storage_type = utils::kBuffer;
-  }
+  utils::StorageType storage_type = utils::kBuffer;
 
   ValueRef packed_weight = graph.add_tensor(
       packed_weight_sizes,