remove splitk from exhaustive tuning (#3272)

Author: causten

CHANGELOG.md

@@ -14,7 +14,6 @@ Full documentation for MIGraphX is available at
 * Added a `--test` flag in migraphx-driver to validate the installation
 * Added support for ONNX Operator: Einsum
 * Added uint8 support in ONNX Operators
-* Enabled Split-k kernel configurations for performance improvements
 * Added fusion for group convolutions
 * Added rocMLIR conv3d support 
 * Added rocgdb to the Dockerfile
@@ -46,7 +45,6 @@ Full documentation for MIGraphX is available at
 * Added support for multi outputs in pointwise ops
 * Improve reduction fusion with reshape operators
 * Use the quantized output when an operator is used again
-* Enabled Split-k GEMM perf configs for rocMLIR based GEMM kernels for better performance on all Hardware 
 
 
 ### Fixes

src/targets/gpu/mlir.cpp

@@ -595,8 +595,7 @@ struct mlir_program
                             {"sym_name", sym_name},
                             {"kernel", std::string("mixr")},
                             {"arch", target_arch},
-                            {"num_cu", num_cu},
-                            {"enable_splitk_for_tuning", true}});
+                            {"num_cu", num_cu}});
         ops.add_region(std::move(region));
         insert(body, std::move(ops));
 

test/gpu/mlir.cpp

@@ -158,7 +158,7 @@ TEST_CASE(conv)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x2x2x2xf32, 8x4x2x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_convolution(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x2x2x2xf32, 8x4x2x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.convolution %arg1, %arg0 {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : <1x8x4x4xf32, 128x16x4x1>, <2x8x3x3xf32, 72x9x3x1> -> <1x2x2x2xf32, 8x4x2x1>
     return %0 : !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>
   }
@@ -181,7 +181,7 @@ TEST_CASE(conv_nhwc)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x1x24x8>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x1x32x8>) -> !migraphx.shaped<1x2x2x2xf32, 8x1x4x2> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_convolution(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x1x24x8>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x1x32x8>) -> !migraphx.shaped<1x2x2x2xf32, 8x1x4x2> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.convolution %arg1, %arg0 {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : <1x8x4x4xf32, 128x1x32x8>, <2x8x3x3xf32, 72x1x24x8> -> <1x2x2x2xf32, 8x1x4x2>
     return %0 : !migraphx.shaped<1x2x2x2xf32, 8x1x4x2>
   }
@@ -204,7 +204,7 @@ TEST_CASE(conv_add_relu)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution_add_relu(%arg0: !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>, %arg1: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg2: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x2x2x2xf32, 8x4x2x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_convolution_add_relu(%arg0: !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>, %arg1: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg2: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x2x2x2xf32, 8x4x2x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.convolution %arg2, %arg1 {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : <1x8x4x4xf32, 128x16x4x1>, <2x8x3x3xf32, 72x9x3x1> -> <1x2x2x2xf32, 8x4x2x1>
     %1 = migraphx.add %0, %arg0 : <1x2x2x2xf32, 8x4x2x1>, <1x2x2x2xf32, 8x4x2x1> -> <1x2x2x2xf32, 8x4x2x1>
     %2 = migraphx.relu %1 : <1x2x2x2xf32, 8x4x2x1> -> <1x2x2x2xf32, 8x4x2x1>
@@ -233,7 +233,7 @@ TEST_CASE(conv_reshape_dim_minus_one)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution_reshape(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x4x1x2xf32, 8x2x2x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_convolution_reshape(%arg0: !migraphx.shaped<2x8x3x3xf32, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xf32, 128x16x4x1>) -> !migraphx.shaped<1x4x1x2xf32, 8x2x2x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.convolution %arg1, %arg0 {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : <1x8x4x4xf32, 128x16x4x1>, <2x8x3x3xf32, 72x9x3x1> -> <1x2x2x2xf32, 8x4x2x1>
     %1 = migraphx.reshape %0 {dims  =  [1,  4,  1,  2]} : <1x2x2x2xf32, 8x4x2x1> -> <1x4x1x2xf32, 8x2x2x1>
     return %1 : !migraphx.shaped<1x4x1x2xf32, 8x2x2x1>
@@ -258,7 +258,7 @@ TEST_CASE(quant_dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_quant_dot_add(%arg0: !migraphx.shaped<1x5x4xi8, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xi8, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xi32, 15x3x1>) -> !migraphx.shaped<1x5x3xi32, 15x3x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_quant_dot_add(%arg0: !migraphx.shaped<1x5x4xi8, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xi8, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xi32, 15x3x1>) -> !migraphx.shaped<1x5x3xi32, 15x3x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.quant_dot %arg0, %arg1 : <1x5x4xi8, 20x4x1>, <1x4x3xi8, 12x3x1> -> <1x5x3xi32, 15x3x1>
     %1 = migraphx.add %0, %arg2 : <1x5x3xi32, 15x3x1>, <1x5x3xi32, 15x3x1> -> <1x5x3xi32, 15x3x1>
     return %1 : !migraphx.shaped<1x5x3xi32, 15x3x1>
@@ -285,7 +285,7 @@ TEST_CASE(dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_add(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xf32, 15x3x1>) -> !migraphx.shaped<1x5x3xf32, 15x3x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_dot_add(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xf32, 15x3x1>) -> !migraphx.shaped<1x5x3xf32, 15x3x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.dot %arg0, %arg1 : <1x5x4xf32, 20x4x1>, <1x4x3xf32, 12x3x1> -> <1x5x3xf32, 15x3x1>
     %1 = migraphx.add %0, %arg2 : <1x5x3xf32, 15x3x1>, <1x5x3xf32, 15x3x1> -> <1x5x3xf32, 15x3x1>
     return %1 : !migraphx.shaped<1x5x3xf32, 15x3x1>
@@ -311,7 +311,7 @@ TEST_CASE(conv_int8_dequantize_quantize)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_quant_convolution_dequantizelinear_quantizelinear(%arg0: !migraphx.shaped<2x8x3x3xi8, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xi8, 128x16x4x1>, %arg2: !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>, %arg3: !migraphx.shaped<1x2x2x2xi32, 8x4x2x1>) -> !migraphx.shaped<1x2x2x2xi32, 8x4x2x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_quant_convolution_dequantizelinear_quantizelinear(%arg0: !migraphx.shaped<2x8x3x3xi8, 72x9x3x1>, %arg1: !migraphx.shaped<1x8x4x4xi8, 128x16x4x1>, %arg2: !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>, %arg3: !migraphx.shaped<1x2x2x2xi32, 8x4x2x1>) -> !migraphx.shaped<1x2x2x2xi32, 8x4x2x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
       %0 = migraphx.quant_convolution %arg1, %arg0 {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : <1x8x4x4xi8, 128x16x4x1>, <2x8x3x3xi8, 72x9x3x1> -> <1x2x2x2xi32, 8x4x2x1>
       %1 = migraphx.dequantizelinear %0, %arg2, %arg3 : <1x2x2x2xi32, 8x4x2x1>, <1x2x2x2xf32, 8x4x2x1>, !migraphx.shaped<1x2x2x2xi32, 8x4x2x1> -> <1x2x2x2xf32, 8x4x2x1>
       %2 = migraphx.quantizelinear %1, %arg2, %arg3 : <1x2x2x2xf32, 8x4x2x1>, <1x2x2x2xf32, 8x4x2x1>, !migraphx.shaped<1x2x2x2xi32, 8x4x2x1> -> <1x2x2x2xi32, 8x4x2x1>
@@ -344,7 +344,7 @@ TEST_CASE(dot_convert)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_convert(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>) -> !migraphx.shaped<1x5x3xf16, 15x3x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_dot_convert(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>) -> !migraphx.shaped<1x5x3xf16, 15x3x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.dot %arg0, %arg1 : <1x5x4xf32, 20x4x1>, <1x4x3xf32, 12x3x1> -> <1x5x3xf32, 15x3x1>
     %1 = migraphx.convert %0 {target_type  =  1  :  i64} : <1x5x3xf32, 15x3x1> to <1x5x3xf16, 15x3x1>
     return %1 : !migraphx.shaped<1x5x3xf16, 15x3x1>
@@ -370,7 +370,7 @@ TEST_CASE(dot_where)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_where(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xi8, 15x3x1>, %arg3: !migraphx.shaped<1x5x3xf32, 15x3x1>) -> !migraphx.shaped<1x5x3xf32, 15x3x1> attributes {arch = "", enable_splitk_for_tuning  =  true, kernel = "mixr", num_cu = 0 : i64} {
+  func.func @mlir_dot_where(%arg0: !migraphx.shaped<1x5x4xf32, 20x4x1>, %arg1: !migraphx.shaped<1x4x3xf32, 12x3x1>, %arg2: !migraphx.shaped<1x5x3xi8, 15x3x1>, %arg3: !migraphx.shaped<1x5x3xf32, 15x3x1>) -> !migraphx.shaped<1x5x3xf32, 15x3x1> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
     %0 = migraphx.dot %arg0, %arg1 : <1x5x4xf32, 20x4x1>, <1x4x3xf32, 12x3x1> -> <1x5x3xf32, 15x3x1>
     %1 = migraphx.where %arg2, %0, %arg3 : <1x5x3xi8, 15x3x1>, <1x5x3xf32, 15x3x1>, <1x5x3xf32, 15x3x1> -> <1x5x3xf32, 15x3x1>
     return %1 : !migraphx.shaped<1x5x3xf32, 15x3x1>