[HOTFIX][BACKPORT] Manually add missing perf config for MI200 to avoid perf regression (#1818)

* hot fix: add missing perfConfig manually to recover performance on Mi200

* formattign

* fix test

Author: umangyadav

mlir/include/mlir/Dialect/Rock/Tuning/QuickTuningPerfconfigs.inc

@@ -233,6 +233,7 @@ const InitParamsAccel PopulateParamsXDL::initParametersForward8BitConv[PopulateP
     {128,64,16,128,16,8,1,true,true},
     {64,32,4,64,16,16,1,true,true},
     {32,16,8,16,16,16,1,true,true},
+    {64,256,4,32,16,4,1,true,true},
     {64,256,4,32,32,8,1,true,true},
     {256,32,4,64,16,16,1,true,true},
     {32,32,32,32,16,16,1,true,true},
@@ -270,7 +271,7 @@ static const InitParamsAccel initParametersI8Gemm[nInitParametersI8Gemm];
 // END_GEMM_XDL_i8_DECS
 
 // BEGIN_CONV_XDL_i8_DECS
-static constexpr size_t nInitParametersForward8BitConv = 24;
+static constexpr size_t nInitParametersForward8BitConv = 25;
 static const InitParamsAccel initParametersForward8BitConv[nInitParametersForward8BitConv];
 // END_CONV_XDL_i8_DECS
 

mlir/test/Dialect/Rock/affix_tuning_params.mlir

@@ -49,7 +49,7 @@ func.func @rock_conv_f16(%filter : memref<1x128x8x3x3xf16>, %input : memref<128x
 func.func @rock_conv_i8(%filter : memref<1x128x8x3x3xi8>, %input : memref<128x1x8x32x32xi8>, %output : memref<128x1x128x30x30xi32>) {
   // CHECK: rock.conv
   // CHECK-SAME: derivedBlockSize = 256
-  // CHECK-SAME: params = #rock.xdlops_gemm_derived_params<kpackPerBlock = 4, mPerBlock = 64, nPerBlock = 256, kpack = 8, mPerWave = 32, nPerWave = 128, mnPerXdl = 32, splitKFactor = 1, forceUnroll = true>
+  // CHECK-SAME: params = #rock.xdlops_gemm_derived_params<kpackPerBlock = 4, mPerBlock = 64, nPerBlock = 256, kpack = 4, mPerWave = 32, nPerWave = 128, mnPerXdl = 16, splitKFactor = 1, forceUnroll = true>
   // GRID: rock.gridwise_gemm
   // GRID-SAME: gridSize = 900
   rock.conv(%filter, %input, %output) features = mfma|dot|atomic_add|atomic_add_f16 {