Improve MLIR symbol names (#2025)

src/targets/gpu/compile_gen.cpp

@@ -331,7 +331,7 @@ static std::vector<std::string> get_op_names(const module& m)
     {
         if(starts_with(ins.name(), "@"))
             continue;
-        if(ins.name() == "multibroadcast")
+        if(contains({"multibroadcast", "contiguous"}, ins.name()))
             continue;
         if(ins.name() == "pointwise")
         {

src/targets/gpu/mlir.cpp

@@ -50,6 +50,7 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/gpu/code_object_op.hpp>
 #include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/compile_gen.hpp>
 #include <migraphx/gpu/device_name.hpp>
 #include <migraphx/gpu/perfdb.hpp>
 #include <migraphx/gpu/tuning_config.hpp>
@@ -559,14 +560,7 @@ struct mlir_program
 
     static std::string get_symbol_name(const module& m)
     {
-        for(auto ins : iterator_for(m))
-        {
-            if(ins->name() == "convolution" or ins->name() == "dot")
-            {
-                return "mlir_" + ins->name();
-            }
-        }
-        return "main";
+        return "mlir_" + gen::generate_name_from_ops(m);
     }
 
     void parse(const module& m)

test/gpu/mlir.cpp

@@ -163,7 +163,7 @@ TEST_CASE(conv_add_relu)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_convolution_add_relu(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {arch = "", kernel = "mixr"} {
     %0 = migraphx.convolution(%arg2, %arg1) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
     %1 = migraphx.add(%0, %arg0) : (tensor<1x2x2x2xf32>, tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
     %2 = migraphx.relu(%1) : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
@@ -191,7 +191,7 @@ TEST_CASE(quant_dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @main(%arg0: tensor<1x5x4xi8>, %arg1: tensor<1x4x3xi8>, %arg2: tensor<1x5x3xi32>) -> tensor<1x5x3xi32> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_quant_dot_add(%arg0: tensor<1x5x4xi8>, %arg1: tensor<1x4x3xi8>, %arg2: tensor<1x5x3xi32>) -> tensor<1x5x3xi32> attributes {arch = "", kernel = "mixr"} {
     %0 = migraphx.quant_dot(%arg0, %arg1) : (tensor<1x5x4xi8>, tensor<1x4x3xi8>) -> tensor<1x5x3xi32>
     %1 = migraphx.add(%0, %arg2) : (tensor<1x5x3xi32>, tensor<1x5x3xi32>) -> tensor<1x5x3xi32>
     return %1 : tensor<1x5x3xi32>
@@ -218,7 +218,7 @@ TEST_CASE(dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_dot_add(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr"} {
     %0 = migraphx.dot(%arg0, %arg1) : (tensor<1x5x4xf32>, tensor<1x4x3xf32>) -> tensor<1x5x3xf32>
     %1 = migraphx.add(%0, %arg2) : (tensor<1x5x3xf32>, tensor<1x5x3xf32>) -> tensor<1x5x3xf32>
     return %1 : tensor<1x5x3xf32>
@@ -244,7 +244,7 @@ TEST_CASE(conv_int8_dequantize_quantize)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @main(%arg0: tensor<2x8x3x3xi8>, %arg1: tensor<1x8x4x4xi8>, %arg2: tensor<1x2x2x2xf32>, %arg3: tensor<1x2x2x2xi32>) -> tensor<1x2x2x2xi32> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_quant_convolution_dequantizelinear_quantizelinear(%arg0: tensor<2x8x3x3xi8>, %arg1: tensor<1x8x4x4xi8>, %arg2: tensor<1x2x2x2xf32>, %arg3: tensor<1x2x2x2xi32>) -> tensor<1x2x2x2xi32> attributes {arch = "", kernel = "mixr"} {
       %0 = migraphx.quant_convolution(%arg1, %arg0) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xi8>, tensor<2x8x3x3xi8>) -> tensor<1x2x2x2xi32>
       %1 = migraphx.dequantizelinear(%0, %arg2, %arg3) : (tensor<1x2x2x2xi32>, tensor<1x2x2x2xf32>, tensor<1x2x2x2xi32>) -> tensor<1x2x2x2xf32>
       %2 = migraphx.quantizelinear(%1, %arg2, %arg3) : (tensor<1x2x2x2xf32>, tensor<1x2x2x2xf32>, tensor<1x2x2x2xi32>) -> tensor<1x2x2x2xi32>
@@ -277,7 +277,7 @@ TEST_CASE(dot_convert)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>) -> tensor<1x5x3xf16> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_dot_convert(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>) -> tensor<1x5x3xf16> attributes {arch = "", kernel = "mixr"} {
     %0 = migraphx.dot(%arg0, %arg1) : (tensor<1x5x4xf32>, tensor<1x4x3xf32>) -> tensor<1x5x3xf32>
     %1 = migraphx.convert(%0) {target_type  =  1  :  i64} : (tensor<1x5x3xf32>) -> tensor<1x5x3xf16>
     return %1 : tensor<1x5x3xf16>
@@ -303,7 +303,7 @@ TEST_CASE(dot_where)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xi8>, %arg3: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr"} {
+  func.func @mlir_dot_where(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xi8>, %arg3: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr"} {
     %0 = migraphx.dot(%arg0, %arg1) : (tensor<1x5x4xf32>, tensor<1x4x3xf32>) -> tensor<1x5x3xf32>
     %1 = migraphx.where(%arg2, %0, %arg3) : (tensor<1x5x3xi8>, tensor<1x5x3xf32>, tensor<1x5x3xf32>) -> tensor<1x5x3xf32>
     return %1 : tensor<1x5x3xf32>