Removed enable_split_reader from Conv2dConfig (#27731)

### Ticket
#26430

### Problem description
After removing split_reader impact on `out_subblock_h/w` choice in
#26220, split_reader can be
enabled in cases where it is beneficial and the parameter can be
removed.

Currently split_reader is supported only if:
- Height sharded convolution
- not 1D depthwise convolution
- activation block size height [tiles] > 1 ( so there's something for
each reader to read)


Enabling should be chosen by heuristic that takes into account the
following things:
- local L1 transfer speed (activations are transferred locally L1->L1,
depends on chunk sizes being transferred, differs between dilated convs
and non-dilated convs)
- Noc link bandwith ( split reader works only for Hight sharded
convolutions currently and in HS convs only 1 core reads weights from
DRAM)
- speed of reading weights from DRAM for 1 core and speed of that same
core multicasting to the rest of the cores is aprox the same
- size of the activations block being read
- size of the weights block being read
- time for transferring activations is 2x more than transferring
weights. (exact number subject to change down the line)

### What's changed
- removed `enable_split_reader` from `ttnn.Conv2dConfig` 
- added said heuristic

## Model Performance Comparison: AVG DEVICE KERNEL SAMPLES/S

[main](https://github.com/tenstorrent/tt-metal/actions/runs/17569945562/job/49904184081)

[this_branch](https://github.com/tenstorrent/tt-metal/actions/runs/17557035559)
| Model | Main Branch | Current Branch | Difference | % Change |
|-------|-------------|----------------|------------|----------|
| falcon7b_decode_bfloat16-l1_sharded_seq1_kv_cache_len1024 | 578.1555 |
580.5385 | +2.3830 | 🟢 +0.41% |
| falcon7b_decode_bfloat16-l1_sharded_seq1_kv_cache_len128 | 640.3621 |
643.1583 | +2.7962 | 🟢 +0.44% |
| falcon7b_decode_bfloat16-l1_sharded_seq1_kv_cache_len2047 | 541.8032 |
544.0483 | +2.2451 | 🟢 +0.41% |
| falcon7b_prefill_bfloat16-dram_seq1024_kv_cache_len0 | 3112.8762 |
3109.6271 | -3.2491 | 🔴 -0.10% |
| falcon7b_prefill_bfloat16-dram_seq128_kv_cache_len0 | 2113.3771 |
2119.5114 | +6.1343 | 🟢 +0.29% |
| falcon7b_prefill_bfloat16-dram_seq2048_kv_cache_len0 | 2857.7001 |
2854.8933 | -2.8068 | 🔴 -0.10% |
| segformer_for_semantic_segmentation | 108.9917 | 109.1203 | +0.1286 |
🟢 +0.12% |
| tt_mnist128 | 909142.5649 | 910481.2035 | +1338.6386 | 🟢 +0.15% |
| sdxl_unet | 4.9903 | 4.9749 | -0.0154 | 🔴 -0.31% |
| sdxl_vae | 0.8346 | 0.8339 | -0.0007 | ⚪ -0.08% |
| stable_diffusion_1batch | 13.0289 | 13.0202 | -0.0087 | ⚪ -0.07% |
| ttnn_functional_Swin_s_1 (first) | 6.5262 | 6.5224 | -0.0038 | ⚪
-0.06% |
| ttnn_functional_Swin_v2_1 (first) | 7.4043 | 7.4038 | -0.0005 | ⚪
-0.01% |
| ttnn_functional_ttnn_vgg11_1 | 436.2223 | 438.8588 | +2.6365 | 🟢
+0.60% |
| ttnn_functional_ttnn_vgg16_1 | 352.6563 | 355.9618 | +3.3055 | 🟢
+0.94% |
| ttnn_functional_vgg_unet1 | 280.8845 | 280.2404 | -0.6441 | 🔴 -0.23% |
| ttnn_resnet50_batch_size16 | 5699.5534 | 5665.1853 | -34.3681 | 🔴
-0.60% |
| ttnn_roberta_8 | 180.1718 | 180.9504 | +0.7786 | 🟢 +0.43% |
| ttnn_sentence_bert8 | 461.7476 | 461.5519 | -0.1957 | ⚪ -0.04% |
| ttnn_ufld_v21 | 345.7357 | 344.7538 | -0.9819 | 🔴 -0.28% |
| ttnn_yolov10x1 | 50.6070 | 50.6093 | +0.0023 | ⚪ +0.00% |
| ttnn_yolov111 | 213.4715 | 217.6344 | +4.1629 | 🟢 +1.95% |
| unet-shallow_batch-1_groups-4 | 1585.2156 | 1585.7685 | +0.5529 | ⚪
+0.03% |
| mamba-2.8b_batch_32 | 19861.4416 | 19863.3156 | +1.8740 | ⚪ +0.01% |
| ttnn_efficientnetb01 | 94.8605 | 95.2415 | +0.3810 | 🟢 +0.40% |
| ttnn_functional_Swin_s_1 (second) | 6.5240 | 6.5243 | +0.0003 | ⚪
+0.00% |
| ttnn_functional_Swin_v2_1 (second) | 7.4053 | 7.4027 | -0.0026 | ⚪
-0.04% |
| ttnn_functional_mobilenetv210 | 3497.8757 | 3506.8598 | +8.9841 | 🟢
+0.26% |
| ttnn_functional_yolov4_1 | 98.4996 | 98.3081 | -0.1915 | 🔴 -0.19% |
| ttnn_functional_yolov71 | 121.9927 | 122.1801 | +0.1874 | 🟢 +0.15% |
| ttnn_functional_yolov8x1 | 69.7647 | 69.8535 | +0.0888 | 🟢 +0.13% |
| ttnn_functional_yolov9c1 | 31.9248 | 32.0006 | +0.0758 | 🟢 +0.24% |
| ttnn_vanilla_unet1 | 62.1010 | 62.1471 | +0.0461 | ⚪ +0.07% |
| ttnn_vovnet1 | 121.5876 | 121.5079 | -0.0797 | ⚪ -0.07% |
| ttnn_yolov12x1 | 14.3381 | 14.3413 | +0.0032 | ⚪ +0.02% |
| ttnn_yolov5x1 | 54.9410 | 54.9104 | -0.0306 | ⚪ -0.06% |
| ttnn_yolov6l1 | 76.9110 | 76.7928 | -0.1182 | 🔴 -0.15% |
| ttnn_yolov8s1 | 239.5090 | 240.5884 | +1.0794 | 🟢 +0.45% |
| vit-8 | 1538.0972 | 1539.6779 | +1.5807 | 🟢 +0.10% |


### Checklist
- [X] [All post
commit](https://github.com/tenstorrent/tt-metal/actions/workflows/all-post-commit-workflows.yaml)
CI
[passed](https://github.com/tenstorrent/tt-metal/actions/runs/17713433723)
- [X] [Blackhole Post
commit](https://github.com/tenstorrent/tt-metal/actions/workflows/blackhole-post-commit.yaml)
CI
[passed](https://github.com/tenstorrent/tt-metal/actions/runs/17713434918)
- [X] [Model
regression](https://github.com/tenstorrent/tt-metal/actions/workflows/perf-models.yaml)
CI [same as
main](https://github.com/tenstorrent/tt-metal/actions/runs/17609536091)
- [X] [Device performance
regression](https://github.com/tenstorrent/tt-metal/actions/workflows/perf-device-models.yaml)
CI
[passes](https://github.com/tenstorrent/tt-metal/actions/runs/17676787104)
- [X] [(Single-card) Frequent model and ttnn
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/fast-dispatch-full-regressions-and-models.yaml)
CI
[passes](https://github.com/tenstorrent/tt-metal/actions/runs/17676793012)
- [X] [Nightly tt-metal L2
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/tt-metal-l2-nightly.yaml)
CI
[passes](https://github.com/tenstorrent/tt-metal/actions/runs/17609520723)

Author: astancovTT

models/demos/mobilenetv2/tests/perf/test_perf_mobilenetv2.py

@@ -14,7 +14,7 @@
 @pytest.mark.parametrize(
     "batch_size, expected_perf",
     [
-        [MOBILENETV2_BATCH_SIZE, 3436],
+        [MOBILENETV2_BATCH_SIZE, 3445],
     ],
 )
 def test_perf_device_mobilenetv2(batch_size, expected_perf):

models/demos/mobilenetv2/tt/common.py

@@ -23,7 +23,6 @@ def __init__(
         width_shard=False,
         act_blocks=32,
         enable_act_double_buffer=False,
-        enable_split_reader=False,
         reshard_if_not_optimal=True,
         activation_dtype=ttnn.bfloat8_b,
         shard_layout=ttnn.TensorMemoryLayout.HEIGHT_SHARDED,
@@ -42,7 +41,6 @@ def __init__(
         self.width_shard = width_shard
         self.act_blocks = act_blocks
         self.enable_act_double_buffer = enable_act_double_buffer
-        self.enable_split_reader = enable_split_reader
         self.reshard_if_not_optimal = reshard_if_not_optimal
         self.batch_size = batch_size
         self.shard_layout = shard_layout
@@ -64,9 +62,6 @@ def _initialize_conv_config(self):
             act_block_w_div=1,
             deallocate_activation=self.deallocate_activation,
             enable_act_double_buffer=self.enable_act_double_buffer,
-            enable_split_reader=True
-            if self.shard_layout == ttnn.TensorMemoryLayout.HEIGHT_SHARDED
-            else self.enable_split_reader,
             output_layout=self.output_layout,
             reallocate_halo_output=False,
             reshard_if_not_optimal=self.reshard_if_not_optimal,

models/demos/segformer/tt/common.py

@@ -48,7 +48,6 @@ def __call__(self, device, input_tensor):
             deallocate_activation=self.deallocate,
             reallocate_halo_output=True,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             output_layout=self.output_layout,
         )
         compute_config = ttnn.init_device_compute_kernel_config(

models/demos/ttnn_resnet/tt/ttnn_functional_resnet50.py

@@ -150,7 +150,6 @@ def run_downsample_if_req(
         height_sharding=None,
         packer_l1_accum_enabled=True,
         enable_act_double_buffer=False,
-        enable_split_reader=False,
     ):
         if self.downsample:
             logger.debug(f"Running downsample")
@@ -180,7 +179,6 @@ def run_downsample_if_req(
                     if input_width < 56
                     else False,
                     enable_weights_double_buffer=True if input_width < 56 else False,
-                    enable_split_reader=enable_split_reader,
                     full_inner_dim=True,
                 ),
             }
@@ -217,7 +215,6 @@ def __call__(
         eltwise_binary_out_in_place=True,
         packer_l1_acc=True,
         enable_act_double_buffer=False,
-        enable_split_reader=False,
         ops_parallel_config=None,
         layer_module=None,
     ):
@@ -287,7 +284,6 @@ def __call__(
                 height_sharding,
                 packer_l1_accum_enabled=packer_l1_acc,
                 enable_act_double_buffer=False,
-                enable_split_reader=enable_split_reader,
             )
             if layer_module and layer_module == "layer4_module1":
                 if ops_parallel_config and "layer4_module1_downsample" not in ops_parallel_config:
@@ -331,7 +327,6 @@ def __call__(
                 reshard_if_not_optimal=reshard_if_not_optimal,
                 enable_act_double_buffer=enable_act_double_buffer,
                 enable_weights_double_buffer=True,
-                enable_split_reader=enable_split_reader,
                 full_inner_dim=True,
             ),
         }
@@ -439,7 +434,6 @@ def __call__(
                 height_sharding,
                 packer_l1_accum_enabled=packer_l1_acc,
                 enable_act_double_buffer=enable_act_double_buffer,
-                enable_split_reader=enable_split_reader,
             )
 
         assert ds_out is not None, "ds_out is None"
@@ -578,7 +572,6 @@ def __init__(
             deallocate_activation=dealloc_input,
             act_block_h_override=act_block_h_override,
             enable_act_double_buffer=is_wormhole_b0() or is_blackhole(),
-            enable_split_reader=True,
             shard_layout=ttnn.TensorMemoryLayout.HEIGHT_SHARDED,
             reshard_if_not_optimal=False,
             # otherwise act block h is not big enough for the reuse
@@ -812,7 +805,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             reshard_if_not_optimal=reshard,
             height_sharding=height_shard,
             enable_act_double_buffer=True,
-            enable_split_reader=True,
         )
 
         if is_first_run:
@@ -833,7 +825,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=False,
-            enable_split_reader=True,
             layer_module="layer1_module2",
         )
 
@@ -845,7 +836,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=False,
-            enable_split_reader=True,
             layer_module="layer1_module3",
         )
 
@@ -864,7 +854,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             reshard_if_not_optimal=reshard,
             height_sharding=height_shard,
             enable_act_double_buffer=True,
-            enable_split_reader=True,
             layer_module="layer2_module1",
         )
 
@@ -886,7 +875,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=True,
             layer_module="layer2_module2",
         )
 
@@ -898,7 +886,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=True,
             layer_module="layer2_module3",
         )
 
@@ -910,7 +897,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=True,
             layer_module="layer2_module4",
         )
 
@@ -931,7 +917,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             reshard_if_not_optimal=reshard,
             height_sharding=height_shard,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
         )
 
         if is_first_run:
@@ -952,7 +937,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
         )
 
         logger.debug(f"==== Running layer 3 module 3")
@@ -963,7 +947,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             layer_module="layer3_module3",
         )
 
@@ -975,7 +958,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             layer_module="layer3_module4",
         )
 
@@ -987,7 +969,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             layer_module="layer3_module5",
         )
 
@@ -1000,7 +981,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_width,
             eltwise_binary_out_in_place=True,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
         )
 
         reshard = is_blackhole() and self.batch_size == 20
@@ -1031,7 +1011,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             reshard_if_not_optimal=reshard,
             height_sharding=height_shard,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             ops_parallel_config=ops_parallel_config,
             layer_module="layer4_module1",
         )
@@ -1044,7 +1023,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             layer_module="layer4_module2",
         )
 
@@ -1056,7 +1034,6 @@ def run(self, input_tensor, device, ops_parallel_config) -> ttnn.Tensor:
             x_height,
             x_width,
             enable_act_double_buffer=True,
-            enable_split_reader=False,
             layer_module="layer4_module3",
         )
 

models/demos/ufld_v2/ttnn/common.py

@@ -46,7 +46,6 @@ def __init__(
             deallocate_activation=dealloc_act,
             enable_act_double_buffer=True if is_blk else False,
             enable_weights_double_buffer=True if is_blk else False,
-            enable_split_reader=True if not is_blk else False,
             reshard_if_not_optimal=True,
             activation=activation,
         )

models/demos/vanilla_unet/common.py

@@ -8,7 +8,7 @@
 
 from models.demos.vanilla_unet.reference.unet import UNet
 
-VANILLA_UNET_L1_SMALL_SIZE = (7 * 8192) + 1730
+VANILLA_UNET_L1_SMALL_SIZE = (7 * 8192) + 2592
 
 
 def load_torch_model(model_location_generator=None):