Adding bf16 data type workaround for max_pool2d op.

lib/Dialect/TTNN/IR/TTNNOps.cpp

@@ -90,13 +90,6 @@ ::mlir::LogicalResult mlir::tt::ttnn::MaxPool2dOp::verify() {
   ::mlir::RankedTensorType inputType = getInput().getType();
   ::llvm::ArrayRef<int64_t> inputShape = getInput().getType().getShape();
 
-  if (!inputType.getElementType().isBF16()) {
-    return emitOpError()
-           << "ttnn.max_pool2d currently only supports an input type of "
-              "bfloat16. Recieved "
-           << inputType.getElementType() << ".";
-  }
-
   if (getKernelHeight() > getInputHeight()) {
     return emitOpError() << "Kernel height " << getKernelHeight()
                          << " is greater than input height " << getInputHeight()
@@ -111,13 +104,13 @@ ::mlir::LogicalResult mlir::tt::ttnn::MaxPool2dOp::verify() {
 
   if (inputType.getRank() != 4) {
     return emitOpError()
-           << "Input tensor rank must be 4. Recieved input with rank "
+           << "Input tensor rank must be 4. Received input with rank "
            << inputType.getRank() << ". Shape: (" << inputShape << ").";
   }
 
   if (inputShape[0] != 1 || inputShape[1] != 1) {
     return emitOpError() << "Maxpool input must be in the form (1, 1, N*H*W, "
-                            "C). Recieved shape ("
+                            "C). Received shape ("
                          << inputShape << ").";
   }
 

lib/Dialect/TTNN/IR/TTNNWorkarounds.cpp

@@ -7,6 +7,7 @@
 #include "ttmlir/Utils.h"
 
 #include "llvm/ADT/SmallVector.h"
+#include <optional>
 
 namespace mlir::tt::ttnn::wa {
 
@@ -83,14 +84,17 @@ TTNNOperandsWorkarounds::createEmptyTTNNOperandsWorkarounds(Operation *op) {
 // tile layout, but the output of the operation is strictly in row-major layout.
 // In order to keep the output consistent with the input, the row-major
 // workaround is applied to both the input and output operands.
+// The input and output operands are expected to use the bf16 data type, so the
+// bf16 workaround is applied to both the input and output operands.
 TTNNOperandsWorkarounds
 TTNNOperandsWorkaroundsFactory::createMaxPool2DOpOperandsWorkarounds() {
-  wa::TTNNOperandWorkarounds rowMajorLayoutWorkaround =
-      wa::TTNNOperandWorkarounds(Layout::RowMajor);
+  wa::TTNNOperandWorkarounds rowMajorLayoutBF16Workaround;
+  rowMajorLayoutBF16Workaround.tensorLayoutWorkaround = Layout::RowMajor;
+  rowMajorLayoutBF16Workaround.tensorDataTypeWorkaround = DataType::BFloat16;
   return wa::TTNNOperandsWorkarounds::createEmptyTTNNOperandsWorkarounds()
-      .addInputOperandWorkaround(rowMajorLayoutWorkaround)
-      .addInputOperandWorkaround(rowMajorLayoutWorkaround)
-      .addOutputOperandWorkaround(rowMajorLayoutWorkaround);
+      .addInputOperandWorkaround(rowMajorLayoutBF16Workaround)
+      .addInputOperandWorkaround(rowMajorLayoutBF16Workaround)
+      .addOutputOperandWorkaround(rowMajorLayoutBF16Workaround);
 }
 
 // Factory method to create a set of workarounds for embedding operation
@@ -107,12 +111,13 @@ TTNNOperandsWorkaroundsFactory::createMaxPool2DOpOperandsWorkarounds() {
 TTNNOperandsWorkarounds
 TTNNOperandsWorkaroundsFactory::createEmbeddingOpOperandsWorkarounds() {
   // Create input and weight workarounds.
-  TTNNOperandWorkarounds inputWorkaround =
-      TTNNOperandWorkarounds(Layout::RowMajor);
+  TTNNOperandWorkarounds inputRowMajorInt32Workaround;
+  inputRowMajorInt32Workaround.tensorLayoutWorkaround = Layout::RowMajor;
+  inputRowMajorInt32Workaround.tensorDataTypeWorkaround = DataType::UInt32;
   TTNNOperandWorkarounds weightWorkaround =
       TTNNOperandWorkarounds(DataType::BFloat16);
   return TTNNOperandsWorkarounds::createEmptyTTNNOperandsWorkarounds(0, 0)
-      .addInputOperandWorkaround(inputWorkaround)
+      .addInputOperandWorkaround(inputRowMajorInt32Workaround)
       .addInputOperandWorkaround(weightWorkaround)
       .addInputOperandWorkaround(weightWorkaround)
       .addOutputOperandWorkaround(weightWorkaround);

runtime/lib/ttnn/operations/layout/typecast.cpp

@@ -6,6 +6,7 @@
 #include "tt/runtime/detail/ttnn.h"
 #include "tt/runtime/ttnn/operations/utils.h"
 #include "tt/runtime/ttnn/utils.h"
+#include <ttnn/operations/core/core.hpp>
 
 namespace tt::runtime::ttnn::operations::layout {
 
@@ -16,7 +17,11 @@ void run(const ::tt::target::ttnn::TypecastOp *op, ProgramContext &context) {
   ::ttnn::DataType targetDataType =
       ::tt::runtime::ttnn::utils::toTTNNDataType(op->dtype());
 
-  ::ttnn::Tensor out = ::ttnn::typecast(inputTensor, targetDataType);
+  bool isOnHost =
+      ::tt::runtime::ttnn::utils::isOnHost(inputTensor.storage_type());
+
+  ::ttnn::Tensor out = isOnHost ? ::ttnn::to_dtype(inputTensor, targetDataType)
+                                : ::ttnn::typecast(inputTensor, targetDataType);
   tensorPool.insert_or_assign(op->out()->global_id(), out);
 }
 

test/ttmlir/Dialect/TTNN/Transforms/Workarounds/max_pool2d_workaround.mlir

@@ -3,37 +3,41 @@
 #dram = #ttnn.buffer_type<dram>
 #system_desc = #tt.system_desc<[{role = host, target_triple = "x86_64-pc-linux-gnu"}], [{arch = <wormhole_b0>, grid = 8x8, l1_size = 1499136, num_dram_channels = 12, dram_channel_size = 1073741824, noc_l1_address_align_bytes = 16, pcie_address_align_bytes = 32, noc_dram_address_align_bytes = 32, l1_unreserved_base = 1024, erisc_l1_unreserved_base = 1024, dram_unreserved_base = 1024, dram_unreserved_end = 1073741824, physical_cores = {worker = [ 0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7,  1x0,  1x1,  1x2,  1x3,  1x4,  1x5,  1x6,  1x7,  2x0,  2x1,  2x2,  2x3,  2x4,  2x5,  2x6,  2x7,  3x0,  3x1,  3x2,  3x3,  3x4,  3x5,  3x6,  3x7,  4x0,  4x1,  4x2,  4x3,  4x4,  4x5,  4x6,  4x7,  5x0,  5x1,  5x2,  5x3,  5x4,  5x5,  5x6,  5x7,  6x0,  6x1,  6x2,  6x3,  6x4,  6x5,  6x6,  6x7,  7x0,  7x1,  7x2,  7x3,  7x4,  7x5,  7x6,  7x7] dram = [ 8x0,  9x0,  10x0,  8x1,  9x1,  10x1,  8x2,  9x2,  10x2,  8x3,  9x3,  10x3]}, supported_data_types = [<f32>, <f16>, <bf16>, <bfp_f8>, <bfp_bf8>, <bfp_f4>, <bfp_bf4>, <bfp_f2>, <bfp_bf2>, <u32>, <u16>, <u8>], supported_tile_sizes = [ 4x16,  16x16,  32x16,  4x32,  16x32,  32x32], num_cbs = 32}], [0], [3 : i32], [ 0x0x0x0]>
 #system_memory = #ttnn.buffer_type<system_memory>
-#ttnn_layout = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 16384 + d1 * 128 + d2, d3), <1x1>, memref<16384x32xbf16, #system_memory>>
-#ttnn_layout1 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 4096 + d1 * 64 + d2, d3), <1x1>, memref<4096x32xbf16, #system_memory>>
-#ttnn_layout2 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 16384 + d1 * 128 + d2, d3), <1x1>, memref<512x1x!tt.tile<32x32, bf16>, #dram>, <interleaved>>
-#ttnn_layout3 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 4096 + d1 * 64 + d2, d3), <1x1>, memref<128x1x!tt.tile<32x32, bf16>, #dram>, <interleaved>>
+#ttnn_layout = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 16384 + d1 * 128 + d2, d3), <1x1>, memref<16384x32xf32, #system_memory>>
+#ttnn_layout1 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 4096 + d1 * 64 + d2, d3), <1x1>, memref<4096x32xf32, #system_memory>>
+#ttnn_layout2 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 16384 + d1 * 128 + d2, d3), <1x1>, memref<512x1x!tt.tile<32x32, f32>, #dram>, <interleaved>>
+#ttnn_layout3 = #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 4096 + d1 * 64 + d2, d3), <1x1>, memref<128x1x!tt.tile<32x32, f32>, #dram>, <interleaved>>
 module attributes {tt.device = #device, tt.system_desc = #system_desc} {
-  func.func @forward(%arg0: tensor<1x128x128x32xbf16, #ttnn_layout>) -> tensor<1x64x64x32xbf16, #ttnn_layout1> {
+  func.func @forward(%arg0: tensor<1x128x128x32xf32, #ttnn_layout>) -> tensor<1x64x64x32xf32, #ttnn_layout1> {
     %0 = "ttnn.get_device"() <{mesh_shape = #ttnn<mesh_shape 1x1>}> : () -> !tt.device<#device>
     // CHECK: %[[DEVICE_OP:.*]] = "ttnn.get_device"[[C:.*]]
-    %1 = "ttnn.to_layout"(%arg0, %0) <{dtype = #tt.supportedDataTypes<bf16>, layout = #ttnn.layout<tile>, memory_config = #ttnn.memory_config<#dram, <<512x1>>, <interleaved>>}> : (tensor<1x128x128x32xbf16, #ttnn_layout>, !tt.device<#device>) -> tensor<1x128x128x32xbf16, #ttnn_layout2>
-    %2 = "ttnn.reshape"(%1) <{shape = [1 : i32, 1 : i32, 16384 : i32, 32 : i32]}> : (tensor<1x128x128x32xbf16, #ttnn_layout2>) -> tensor<1x1x16384x32xbf16, #ttnn_layout2>
-    %3 = "ttnn.empty"(%0) <{dtype = #tt.supportedDataTypes<bf16>, layout = #ttnn.layout<tile>, memory_config = #ttnn.memory_config<#dram, <<128x1>>, <interleaved>>, shape = #ttnn.shape<1x1x4096x32>}> : (!tt.device<#device>) -> tensor<1x1x4096x32xbf16, #ttnn_layout3>
+    %1 = "ttnn.to_layout"(%arg0, %0) <{dtype = #tt.supportedDataTypes<f32>, layout = #ttnn.layout<tile>, memory_config = #ttnn.memory_config<#dram, <<512x1>>, <interleaved>>}> : (tensor<1x128x128x32xf32, #ttnn_layout>, !tt.device<#device>) -> tensor<1x128x128x32xf32, #ttnn_layout2>
+    %2 = "ttnn.reshape"(%1) <{shape = [1 : i32, 1 : i32, 16384 : i32, 32 : i32]}> : (tensor<1x128x128x32xf32, #ttnn_layout2>) -> tensor<1x1x16384x32xf32, #ttnn_layout2>
+    %3 = "ttnn.empty"(%0) <{dtype = #tt.supportedDataTypes<f32>, layout = #ttnn.layout<tile>, memory_config = #ttnn.memory_config<#dram, <<128x1>>, <interleaved>>, shape = #ttnn.shape<1x1x4096x32>}> : (!tt.device<#device>) -> tensor<1x1x4096x32xf32, #ttnn_layout3>
     // CHECK: %[[EMPTY_OP:.*]] = "ttnn.empty"(%[[DEVICE_OP]])
     // Check that the input operand is transformed into the row major layout.
     // CHECK-NEXT: %[[TO_LAYOUT_INPUT:.*]] = "ttnn.to_layout"
+    // CHECK-SAME: dtype = #tt.supportedDataTypes<bf16>
     // CHECK-SAME: layout = #ttnn.layout<row_major>
     // CHECK-SAME: memory_config = #ttnn.memory_config<#dram, <<16384x32>>, <interleaved>>
     // CHECK-SAME: -> tensor<1x1x16384x32xbf16,
     // Check that the output operand is transformed into the row major layout.
     // CHECK-NEXT: %[[TO_LAYOUT_OUTPUT_DPS:.*]] = "ttnn.to_layout"(%[[EMPTY_OP]], %[[DEVICE_OP]])
+    // CHECK-SAME: dtype = #tt.supportedDataTypes<bf16>
     // CHECK-SAME: layout = #ttnn.layout<row_major>
     // CHECK-SAME: memory_config = #ttnn.memory_config<#dram, <<4096x32>>, <interleaved>>
     // CHECK-SAME: -> tensor<1x1x4096x32xbf16,
-    %4 = "ttnn.max_pool2d"(%2, %3, %0) <{batch_size = 1 : si32, ceil_mode = false, channels = 32 : si32, dilation_height = 1 : si32, dilation_width = 1 : si32, input_height = 128 : si32, input_width = 128 : si32, kernel_height = 2 : si32, kernel_width = 2 : si32, padding_height = 0 : si32, padding_width = 0 : si32, stride_height = 2 : si32, stride_width = 2 : si32}> : (tensor<1x1x16384x32xbf16, #ttnn_layout2>, tensor<1x1x4096x32xbf16, #ttnn_layout3>, !tt.device<#device>) -> tensor<1x1x4096x32xbf16, #ttnn_layout3>
+    %4 = "ttnn.max_pool2d"(%2, %3, %0) <{batch_size = 1 : si32, ceil_mode = false, channels = 32 : si32, dilation_height = 1 : si32, dilation_width = 1 : si32, input_height = 128 : si32, input_width = 128 : si32, kernel_height = 2 : si32, kernel_width = 2 : si32, padding_height = 0 : si32, padding_width = 0 : si32, stride_height = 2 : si32, stride_width = 2 : si32}> : (tensor<1x1x16384x32xf32, #ttnn_layout2>, tensor<1x1x4096x32xf32, #ttnn_layout3>, !tt.device<#device>) -> tensor<1x1x4096x32xf32, #ttnn_layout3>
     // CHECK-NEXT: %[[MAX_POOL_2D_OP:.*]] = "ttnn.max_pool2d"(%[[TO_LAYOUT_INPUT]], %[[TO_LAYOUT_OUTPUT_DPS]], %[[DEVICE_OP]])
+    // Check that the output operand is transformed back into the tile and f32 data type.
     // CHECK-NEXT: %[[TO_LAYOUT_OUTPUT:.*]] = "ttnn.to_layout"(%[[MAX_POOL_2D_OP]], %[[DEVICE_OP]])
+    // CHECK-SAME: dtype = #tt.supportedDataTypes<f32>
     // CHECK-SAME: layout = #ttnn.layout<tile>
     // CHECK-SAME: memory_config = #ttnn.memory_config<#dram, <<128x1>>, <interleaved>>
-    // CHECK-SAME: -> tensor<1x1x4096x32xbf16
-    %5 = "ttnn.reshape"(%4) <{shape = [1 : i32, 64 : i32, 64 : i32, 32 : i32]}> : (tensor<1x1x4096x32xbf16, #ttnn_layout3>) -> tensor<1x64x64x32xbf16, #ttnn_layout3>
+    // CHECK-SAME: -> tensor<1x1x4096x32xf32
+    %5 = "ttnn.reshape"(%4) <{shape = [1 : i32, 64 : i32, 64 : i32, 32 : i32]}> : (tensor<1x1x4096x32xf32, #ttnn_layout3>) -> tensor<1x64x64x32xf32, #ttnn_layout3>
     // CHECK-NEXT: ttnn.reshape
-    %6 = "ttnn.to_layout"(%5) <{dtype = #tt.supportedDataTypes<bf16>, layout = #ttnn.layout<row_major>, memory_config = #ttnn.memory_config<#system_memory, <<4096x32>>>}> : (tensor<1x64x64x32xbf16, #ttnn_layout3>) -> tensor<1x64x64x32xbf16, #ttnn_layout1>
-    return %6 : tensor<1x64x64x32xbf16, #ttnn_layout1>
+    %6 = "ttnn.to_layout"(%5) <{dtype = #tt.supportedDataTypes<f32>, layout = #ttnn.layout<row_major>, memory_config = #ttnn.memory_config<#system_memory, <<4096x32>>>}> : (tensor<1x64x64x32xf32, #ttnn_layout3>) -> tensor<1x64x64x32xf32, #ttnn_layout1>
+    return %6 : tensor<1x64x64x32xf32, #ttnn_layout1>
   }
 }

test/ttmlir/Dialect/TTNN/pooling/simple_maxpool2d.mlir

@@ -1,9 +1,9 @@
 // RUN: ttmlir-opt --ttir-to-ttnn-backend-pipeline %s | FileCheck %s
 module attributes {} {
-  func.func @forward(%arg0: tensor<1x128x128x32xbf16>) -> tensor<1x64x64x32xbf16> {
-    %0 = tensor.empty() : tensor<1x64x64x32xbf16>
+  func.func @forward(%arg0: tensor<1x128x128x32xf32>) -> tensor<1x64x64x32xf32> {
+    %0 = tensor.empty() : tensor<1x64x64x32xf32>
     // CHECK: %[[C:.*]] = "ttnn.max_pool2d"[[C:.*]]
-    %1 = "ttir.max_pool2d"(%arg0, %0) <{kernel_height=2: si32, kernel_width=2: si32, stride_height=2: si32, stride_width=2: si32, dilation_height=1: si32, dilation_width=1: si32, ceil_mode=false, padding_left=0: si32, padding_right=0: si32, padding_top=0: si32, padding_bottom=0: si32}> : (tensor<1x128x128x32xbf16>, tensor<1x64x64x32xbf16>) -> tensor<1x64x64x32xbf16>
-    return %1 : tensor<1x64x64x32xbf16>
+    %1 = "ttir.max_pool2d"(%arg0, %0) <{kernel_height=2: si32, kernel_width=2: si32, stride_height=2: si32, stride_width=2: si32, dilation_height=1: si32, dilation_width=1: si32, ceil_mode=false, padding_left=0: si32, padding_right=0: si32, padding_top=0: si32, padding_bottom=0: si32}> : (tensor<1x128x128x32xf32>, tensor<1x64x64x32xf32>) -> tensor<1x64x64x32xf32>
+    return %1 : tensor<1x64x64x32xf32>
   }
 }

test/ttmlir/Silicon/TTNN/embedding/simple_embedding.mlir

@@ -2,11 +2,11 @@
 // RUN: FileCheck %s --input-file=%t.mlir
 // RUN: ttmlir-translate --ttnn-to-flatbuffer %t.mlir > %t.ttnn
 module attributes {} {
-  func.func @forward(%arg0: tensor<32x32xbf16>, %arg1: tensor<512x128xbf16>) -> tensor<32x32x128xbf16> {
+  func.func @forward(%arg0: tensor<32x32xf32>, %arg1: tensor<512x128xf32>) -> tensor<32x32x128xf32> {
     // CHECK: %[[C:.*]] = "ttnn.empty"[[C:.*]]
-    %0 = tensor.empty() : tensor<32x32x128xbf16>
+    %0 = tensor.empty() : tensor<32x32x128xf32>
     // CHECK: %[[C:.*]] = "ttnn.embedding"[[C:.*]]
-    %1 = "ttir.embedding"(%arg0, %arg1, %0) : (tensor<32x32xbf16>, tensor<512x128xbf16>, tensor<32x32x128xbf16>) -> tensor<32x32x128xbf16>
-    return %1 : tensor<32x32x128xbf16>
+    %1 = "ttir.embedding"(%arg0, %arg1, %0) : (tensor<32x32xf32>, tensor<512x128xf32>, tensor<32x32x128xf32>) -> tensor<32x32x128xf32>
+    return %1 : tensor<32x32x128xf32>
   }
 }

test/ttmlir/Silicon/TTNN/pooling/simple_pooling.mlir

@@ -2,8 +2,8 @@
 // RUN: FileCheck %s --input-file=%t.mlir
 // RUN: ttmlir-translate --ttnn-to-flatbuffer %t.mlir > %t.ttnn
 module attributes {} {
-  func.func @forward(%arg0: tensor<1x32x128x128xbf16>) -> tensor<1x32x64x64xbf16> {
-    %0 = tensor.empty() : tensor<1x32x64x64xbf16>
+  func.func @forward(%arg0: tensor<1x32x128x128xf32>) -> tensor<1x32x64x64xf32> {
+    %0 = tensor.empty() : tensor<1x32x64x64xf32>
     // CHECK: "ttnn.permute"
     // CHECK-SAME: permutation = array<i64: 0, 2, 3, 1>
     // CHECK: "ttnn.max_pool2d"
@@ -16,7 +16,7 @@ module attributes {} {
         window_strides = array<i64: 1, 1, 2, 2>,
         base_dilations = array<i64: 1, 1, 1, 1>,
         window_dilations = array<i64: 1, 1, 1, 1>,
-        padding = array<i64: 0, 0, 0, 0, 0, 0, 0, 0>}> : (tensor<1x32x128x128xbf16>, tensor<1x32x64x64xbf16>) -> tensor<1x32x64x64xbf16>
-    return %1 : tensor<1x32x64x64xbf16>
+        padding = array<i64: 0, 0, 0, 0, 0, 0, 0, 0>}> : (tensor<1x32x128x128xf32>, tensor<1x32x64x64xf32>) -> tensor<1x32x64x64xf32>
+    return %1 : tensor<1x32x64x64xf32>
   }
 }