GELU Op implementation

tenstorrent · Oct 31, 2024 · 368fc3d · 368fc3d
1 parent caa7b90
commit 368fc3d
Show file tree

Hide file tree

Showing 10 changed files with 59 additions and 1 deletion.
diff --git a/include/ttmlir/Dialect/TTIR/IR/TTIROps.td b/include/ttmlir/Dialect/TTIR/IR/TTIROps.td
@@ -210,6 +210,13 @@ def TTIR_CosOp: TTIR_ElementwiseUnaryOp<"cos"> {
     }];
 }
 
+def TTIR_GeluOp: TTIR_ElementwiseUnaryOp<"gelu"> {
+  let summary = "Eltwise GELU op.";
+  let description = [{
+    Eltwise GELU operation.
+  }];
+}
+
 def TTIR_LogicalNotOp: TTIR_ElementwiseUnaryOp<"logical_not"> {
     let summary = "Eltwise logical not op.";
     let description = [{

diff --git a/include/ttmlir/Dialect/TTNN/IR/TTNNOps.td b/include/ttmlir/Dialect/TTNN/IR/TTNNOps.td
@@ -197,6 +197,13 @@ def TTNN_ExpOp : TTNN_ElementwiseUnaryOp<"exp"> {
     }];
 }
 
+def TTNN_GeluOp: TTNN_ElementwiseUnaryOp<"gelu"> {
+  let summary = "Eltwise GELU.";
+  let description = [{
+    Eltwise GELU operation.
+  }];
+}
+
 def TTNN_LogicalNotOp: TTNN_ElementwiseUnaryOp<"logical_not"> {
     let summary = "Eltwise logical not op.";
     let description = [{

diff --git a/include/ttmlir/Target/TTNN/program.fbs b/include/ttmlir/Target/TTNN/program.fbs
@@ -85,7 +85,8 @@ enum EltwiseOpType: uint32 {
   Minimum = 24,
   Ceil = 25,
   Sin = 26,
-  Cos = 27
+  Cos = 27,
+  Gelu = 28
 }
 
 table EltwiseOp {

diff --git a/lib/Conversion/TTIRToTTNN/TTIRToTTNN.cpp b/lib/Conversion/TTIRToTTNN/TTIRToTTNN.cpp
@@ -823,6 +823,7 @@ void populateTTIRToTTNNPatterns(MLIRContext *ctx, RewritePatternSet &patterns,
            ElementwiseOpConversionPattern<ttir::MinimumOp, ttnn::MinimumOp>,
            ElementwiseOpConversionPattern<ttir::NegOp, ttnn::NegOp>,
            ElementwiseOpConversionPattern<ttir::ReluOp, ttnn::ReluOp>,
+           ElementwiseOpConversionPattern<ttir::GeluOp, ttnn::GeluOp>,
            ElementwiseOpConversionPattern<ttir::SqrtOp, ttnn::SqrtOp>,
            ElementwiseOpConversionPattern<ttir::RsqrtOp, ttnn::RsqrtOp>,
            ElementwiseOpConversionPattern<ttir::SigmoidOp, ttnn::SigmoidOp>,

diff --git a/lib/Conversion/TTNNToEmitC/TTNNToEmitC.cpp b/lib/Conversion/TTNNToEmitC/TTNNToEmitC.cpp
@@ -619,6 +619,7 @@ void populateTTNNToEmitCPatterns(mlir::MLIRContext *ctx,
                DefaultOpConversionPattern<ttnn::LogicalNotOp>,
                DefaultOpConversionPattern<ttnn::NegOp>,
                DefaultOpConversionPattern<ttnn::ReluOp>,
+               DefaultOpConversionPattern<ttnn::GeluOp>,
                DefaultOpConversionPattern<ttnn::SqrtOp>,
                DefaultOpConversionPattern<ttnn::RsqrtOp>,
                DefaultOpConversionPattern<ttnn::SigmoidOp>,

diff --git a/lib/Target/TTNN/TTNNToFlatbuffer.cpp b/lib/Target/TTNN/TTNNToFlatbuffer.cpp
@@ -347,6 +347,8 @@ createEltwiseOp(FlatbufferObjectCache &cache, EltwiseOp op) {
     type = ::tt::target::ttnn::EltwiseOpType::Cos;
   } else if constexpr (std::is_same_v<EltwiseOp, SinOp>) {
     type = ::tt::target::ttnn::EltwiseOpType::Sin;
+  } else if constexpr (std::is_same_v<EltwiseOp, GeluOp>) {
+    type = ::tt::target::ttnn::EltwiseOpType::Gelu;
   } else {
     llvm_unreachable("unhandled EltwiseOp");
   }
@@ -661,6 +663,9 @@ emitTTNNOperation(FlatbufferObjectCache &cache, Operation *op,
   if (auto sinOp = dyn_cast<SinOp>(op); sinOp) {
     return createOperation(cache, createEltwiseOp(cache, sinOp), debugString);
   }
+  if (auto geluOp = dyn_cast<GeluOp>(op); geluOp) {
+    return createOperation(cache, createEltwiseOp(cache, geluOp), debugString);
+  }
 
   llvm_unreachable("unhandled op in emitTTNNOperation");
 }

diff --git a/runtime/lib/ttnn/operations/eltwise/unary.cpp b/runtime/lib/ttnn/operations/eltwise/unary.cpp
@@ -92,6 +92,10 @@ void run(const ::tt::target::ttnn::EltwiseOp *op, ProgramContext &context) {
     runEltwiseUnaryOP(op, tensorPool, ::ttnn::cos);
     break;
   }
+  case ::tt::target::ttnn::EltwiseOpType::Gelu: {
+    runEltwiseUnaryWithFastAndApproximateModeOP(op, tensorPool, ::ttnn::gelu);
+    break;
+  }
   case ::tt::target::ttnn::EltwiseOpType::LogicalNot: {
     runEltwiseUnaryOP(op, tensorPool, ::ttnn::logical_not);
     break;

diff --git a/test/ttmlir/Dialect/TTNN/eltwise/unary/gelu/simple_gelu.mlir b/test/ttmlir/Dialect/TTNN/eltwise/unary/gelu/simple_gelu.mlir
@@ -0,0 +1,11 @@
+// RUN: ttmlir-opt --ttir-to-ttnn-backend-pipeline %s | FileCheck %s
+#any_device = #tt.operand_constraint<dram|l1|scalar|tile|any_device|any_device_tile>
+module attributes {} {
+  func.func @forward(%arg0: tensor<64x128xf32>) -> tensor<64x128xf32> {
+    // CHECK: %[[C:.*]] = "ttnn.empty"[[C:.*]]
+    %0 = tensor.empty() : tensor<64x128xf32>
+    // CHECK: %[[C:.*]] = "ttnn.gelu"[[C:.*]]
+    %1 = "ttir.gelu"(%arg0, %0) <{operandSegmentSizes = array<i32: 1, 1>, operand_constraints = [#any_device, #any_device]}> : (tensor<64x128xf32>, tensor<64x128xf32>) -> tensor<64x128xf32>
+    return %1 : tensor<64x128xf32>
+  }
+}
diff --git a/test/ttmlir/Silicon/TTNN/perf_unit/test_perf_gelu.mlir b/test/ttmlir/Silicon/TTNN/perf_unit/test_perf_gelu.mlir
@@ -0,0 +1,13 @@
+// RUN: ttmlir-opt --ttir-to-ttnn-backend-pipeline="system-desc-path=%system_desc_path%" %s > %t.mlir
+// RUN: FileCheck %s --input-file=%t.mlir
+// RUN: ttmlir-translate --ttnn-to-flatbuffer %t.mlir > %t.ttnn
+#any_device = #tt.operand_constraint<dram|l1|scalar|tile|any_device|any_device_tile>
+#any_device_tile = #tt.operand_constraint<dram|l1|tile|any_device_tile>
+
+func.func @gelu(%arg0: tensor<64x128xf32>) -> tensor<64x128xf32> {
+  // CHECK: %[[C:.*]] = "ttnn.empty"[[C:.*]]
+  %0 = tensor.empty() : tensor<64x128xf32>
+  // CHECK: %[[C:.*]] = "ttnn.gelu"[[C:.*]]
+  %1 = "ttir.gelu"(%arg0, %0) <{operandSegmentSizes = array<i32: 1, 1>, operand_constraints = [#any_device, #any_device]}> : (tensor<64x128xf32>, tensor<64x128xf32>) -> tensor<64x128xf32>
+  return %1 : tensor<64x128xf32>
+}
diff --git a/test/ttmlir/Silicon/TTNN/simple_eltwise.mlir b/test/ttmlir/Silicon/TTNN/simple_eltwise.mlir
@@ -183,3 +183,11 @@ func.func @typecast(%arg0: tensor<64x128xf32>) -> tensor<64x128xbf16> {
   %1 = "ttir.typecast"(%arg0, %0) <{operandSegmentSizes = array<i32: 1, 1>, operand_constraints = [#any_device, #any_device]}> : (tensor<64x128xf32>, tensor<64x128xbf16>) -> tensor<64x128xbf16>
   return %1 : tensor<64x128xbf16>
 }
+
+func.func @gelu(%arg0: tensor<64x128xf32>) -> tensor<64x128xf32> {
+  // CHECK: %[[C:.*]] = "ttnn.empty"[[C:.*]]
+  %0 = tensor.empty() : tensor<64x128xf32>
+  // CHECK: %[[C:.*]] = "ttnn.gelu"[[C:.*]]
+  %1 = "ttir.gelu"(%arg0, %0) <{operandSegmentSizes = array<i32: 1, 1>, operand_constraints = [#any_device, #any_device]}> : (tensor<64x128xf32>, tensor<64x128xf32>) -> tensor<64x128xf32>
+  return %1 : tensor<64x128xf32>
+}