[MLIR] Updates needed for general stride support (#2428)

requirements.txt

@@ -29,4 +29,4 @@ pybind/pybind11@d159a563383d10c821ba7b2a71905d1207db6de4 --build
 msgpack/msgpack-c@cpp-3.3.0 -DMSGPACK_BUILD_TESTS=Off
 sqlite3@3.43.2 -DCMAKE_POSITION_INDEPENDENT_CODE=On
 ROCmSoftwarePlatform/composable_kernel@70eefcf4f263aa5c25f3c9ff0db8f6f199ef0fb9 -DCK_BUILD_JIT_LIB=On -DCMAKE_POSITION_INDEPENDENT_CODE=On
-ROCmSoftwarePlatform/rocMLIR@9e66e8050209f03349a41b6b497f0da2b285a53b -DBUILD_FAT_LIBROCKCOMPILER=On
+ROCmSoftwarePlatform/rocMLIR@a6880f1e6daec99876cd6a4820fbc69c57216401 -DBUILD_FAT_LIBROCKCOMPILER=On

src/targets/gpu/fuse_mlir.cpp

@@ -114,10 +114,7 @@ struct mlir_op
             }
             if(ins->name() == "@return")
             {
-                auto s = ins_shapes[ins->inputs().at(0)].with_type(type);
-                if(not s.standard())
-                    MIGRAPHX_THROW("MLIR doesnt support non-standard output");
-                return s;
+                return ins_shapes[ins->inputs().at(0)].with_type(type);
             }
             std::vector<shape> input_shapes;
             input_shapes.resize(ins->inputs().size());
@@ -139,8 +136,15 @@ get_fusable_input_op_stream(instruction_ref lower_input)
 {
     instruction_ref upper_input = lower_input;
     std::vector<operation> op_stream;
-    while(
-        contains({"slice", "transpose", "contiguous", "reshape", "squeeze", "flatten", "unsqueeze"},
+    while(contains({"slice",
+                    "transpose",
+                    "multibroadcast",
+                    "broadcast",
+                    "contiguous",
+                    "reshape",
+                    "squeeze",
+                    "flatten",
+                    "unsqueeze"},
                    upper_input->name()))
     {
         operation op = upper_input->get_operator();

src/targets/gpu/mlir.cpp

@@ -37,7 +37,7 @@
 #include <mlir-c/Pass.h>
 #include <mlir-c/Support.h>
 #include <mutex>
-#if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 3
+#if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 4
 #warning "Incompatible version of rocMLIR library used, disabling"
 // Only undefine when not using cppcheck
 #ifndef CPPCHECK
@@ -319,31 +319,30 @@ struct mlir_program
         return result;
     }
 
-    MlirType make_tensor(const shape& s) const
+    MlirType make_mlir_shaped(const shape& s) const
     {
-        if(not s.standard())
-            MIGRAPHX_THROW("MLIR expects all tensors to be in standard shape");
         if(s.dynamic())
             MIGRAPHX_THROW("MLIR does not support dynamic shapes");
         std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
-        return mlirRankedTensorTypeGet(
-            lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
+        std::vector<int64_t> strides(s.strides().begin(), s.strides().end());
+        return rocmlirMIXRShapedTypeGet(
+            lens.size(), lens.data(), strides.data(), make_type(s.type()));
     }
 
     template <class Range>
-    std::vector<MlirType> make_tensors(const Range& r)
+    std::vector<MlirType> make_mlir_shapeds(const Range& r)
     {
         std::vector<MlirType> result;
         std::transform(r.begin(), r.end(), std::back_inserter(result), [&](const auto& s) {
-            return make_tensor(s);
+            return make_mlir_shaped(s);
         });
         return result;
     }
 
     MlirType make_function_type(const std::vector<shape>& inputs, const std::vector<shape>& outputs)
     {
-        auto in  = make_tensors(inputs);
-        auto out = make_tensors(outputs);
+        auto in  = make_mlir_shapeds(inputs);
+        auto out = make_mlir_shapeds(outputs);
         return mlirFunctionTypeGet(ctx.get(), in.size(), in.data(), out.size(), out.data());
     }
 
@@ -505,11 +504,7 @@ struct mlir_program
 
         mlir_operation_state& add_results(const std::vector<shape>& outputs)
         {
-            std::vector<shape> reshaped(outputs.size());
-            std::transform(outputs.begin(), outputs.end(), reshaped.begin(), [](const shape& r) {
-                return shape{r.type(), r.lens()};
-            });
-            auto x = prog->make_tensors(reshaped);
+            auto x = prog->make_mlir_shapeds(outputs);
             if(not x.empty())
             {
                 mlirOperationStateAddResults(&op_state, x.size(), x.data());
@@ -582,7 +577,7 @@ struct mlir_program
         std::vector<shape> outputs = m.get_output_shapes();
 
         std::vector<MlirLocation> arg_locs(inputs.size(), location);
-        auto body_inputs   = make_tensors(inputs);
+        auto body_inputs   = make_mlir_shapeds(inputs);
         mlir_region region = mlirRegionCreate();
         mlir_block fbody = mlirBlockCreate(body_inputs.size(), body_inputs.data(), arg_locs.data());
         MlirBlock result = fbody.get();
@@ -608,7 +603,7 @@ struct mlir_program
             return "func.return";
         if(ins->name() == "@literal")
         {
-            return "tosa.const";
+            return "migraphx.literal";
         }
         return "migraphx." + ins->name();
     }
@@ -667,7 +662,8 @@ struct mlir_program
             if(ins->name() == "@literal")
             {
                 literal r            = ins->get_literal();
-                MlirType tensor_type = make_tensor(ins->get_shape());
+                MlirType shaped_type = make_mlir_shaped(ins->get_shape());
+                MlirType tensor_type = rocmlirMIXRShapedTypeAsTensor(shaped_type);
                 MlirAttribute mlir_value_attr =
                     mlirDenseElementsAttrRawBufferGet(tensor_type, r.get_shape().bytes(), r.data());
                 ops.add_attributes({{"value", mlir_value_attr}});
@@ -945,35 +941,7 @@ void adjust_param_shapes(module& m, const std::vector<shape>& inputs)
         auto param        = m.get_parameter(name);
         if(input.standard())
             continue;
-        auto lens    = input.lens();
-        auto strides = input.strides();
-        std::vector<operation> ops;
-        if(input.transposed())
-        {
-            auto perm  = find_permutation(input);
-            auto iperm = invert_permutation(perm);
-            lens       = reorder_dims(lens, iperm);
-            strides    = reorder_dims(strides, iperm);
-            ops.push_back(make_op("transpose", {{"permutation", perm}}));
-        }
-        if(input.broadcasted())
-        {
-            std::transform(lens.begin(),
-                           lens.end(),
-                           strides.begin(),
-                           lens.begin(),
-                           [](auto len, auto stride) -> std::size_t {
-                               if(stride == 0)
-                                   return 1;
-                               return len;
-                           });
-            ops.push_back(make_op("multibroadcast", {{"out_lens", input.lens()}}));
-        }
-        auto new_param =
-            std::accumulate(ops.begin(),
-                            ops.end(),
-                            m.add_parameter(name + ".0", shape{input.type(), lens}),
-                            [&](auto x, auto op) { return m.insert_instruction(param, op, x); });
+        auto new_param = m.add_parameter(name + ".0", input);
         m.replace_instruction(param, new_param);
         m.remove_instruction(param);
     }

test/gpu/mlir.cpp

@@ -141,9 +141,9 @@ TEST_CASE(conv)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution(%arg0: tensor<2x8x3x3xf32>, %arg1: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> 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]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
-    return %0 : tensor<1x2x2x2xf32>
+  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>
   }
 }
 )__migraphx__";
@@ -160,15 +160,38 @@ module {
     EXPECT(verify_mlir(m));
 }
 
+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 = "", 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>
+  }
+}
+)__migraphx__";
+    migraphx::module m;
+    auto x    = m.add_parameter("x", {migraphx::shape::float_type, {1, 8, 4, 4}, {128, 1, 32, 8}});
+    auto w    = m.add_parameter("w", {migraphx::shape::float_type, {2, 8, 3, 3}, {72, 1, 24, 8}});
+    auto conv = m.add_instruction(migraphx::make_op("convolution"), x, w);
+    m.add_return({conv});
+    auto s = migraphx::gpu::dump_mlir(m);
+    // Skip test if MLIR is not enabled
+    if(s.empty())
+        return;
+    CHECK(encode(s) == encode(mlir_output));
+    EXPECT(verify_mlir(m));
+}
+
 TEST_CASE(conv_add_relu)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_convolution_add_relu(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> 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]} : (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>
-    return %2 : tensor<1x2x2x2xf32>
+  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>
+    return %2 : !migraphx.shaped<1x2x2x2xf32, 8x4x2x1>
   }
 }
 )__migraphx__";
@@ -192,10 +215,10 @@ TEST_CASE(quant_dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_quant_dot_add(%arg0: tensor<1x5x4xi8>, %arg1: tensor<1x4x3xi8>, %arg2: tensor<1x5x3xi32>) -> tensor<1x5x3xi32> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
-    %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>
+  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>
   }
 }
 )__migraphx__";
@@ -219,10 +242,10 @@ TEST_CASE(dot_add)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_add(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
-    %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>
+  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>
   }
 }
 )__migraphx__";
@@ -245,11 +268,11 @@ TEST_CASE(conv_int8_dequantize_quantize)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  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", 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]} : (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>
-      return %2 : tensor<1x2x2x2xi32>
+  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>
+      return %2 : !migraphx.shaped<1x2x2x2xi32, 8x4x2x1>
     }
 }
 )__migraphx__";
@@ -278,10 +301,10 @@ TEST_CASE(dot_convert)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_convert(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>) -> tensor<1x5x3xf16> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
-    %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>
+  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>
   }
 }
 )__migraphx__";
@@ -304,10 +327,10 @@ TEST_CASE(dot_where)
 {
     const std::string mlir_output = R"__migraphx__(
 module {
-  func.func @mlir_dot_where(%arg0: tensor<1x5x4xf32>, %arg1: tensor<1x4x3xf32>, %arg2: tensor<1x5x3xi8>, %arg3: tensor<1x5x3xf32>) -> tensor<1x5x3xf32> attributes {arch = "", kernel = "mixr", num_cu = 0 : i64} {
-    %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>
+  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>
   }
 }
 )__migraphx__";