warp-lang 0.10.1__py3-none-win_amd64.whl → 0.11.0__py3-none-win_amd64.whl

warp/native/cutlass/tools/library/scripts/pycutlass/test/gemm/gemm_f16_sm80.py

@@ -0,0 +1,447 @@
+import pycutlass
+from pycutlass import *
+from pycutlass.test import *
+import unittest
+
+from pycutlass.test.gemm_testbed import test_all_gemm
+from pycutlass.utils.device import device_cc
+
+
+@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
+class GemmF16Sm80(unittest.TestCase):
+    def test_SM80_Device_Gemm_f32t_f32n_f32t_tensor_op_bf16_f32_128x128x32_64x64x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 128, 32],
+            stages=3, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.BatchedIdentitySwizzle
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor,
+            direct_store=True
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+
+    def test_SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32_128x128x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 128, 64],
+            stages=3, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32_128x256x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 256, 64],
+            stages=3, warp_count=[2, 4, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32_256x128x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[256, 128, 64],
+            stages=3, warp_count=[4, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16_sliced_k_128x64x64_64x64x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float16, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 64, 64],
+            stages=3, warp_count=[2, 1, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float16
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_GemmUniversal_f16n_f16t_f32t_tensor_op_f32_64x64x32_32x32x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float16, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[64, 64, 32],
+            stages=10, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float16
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32_256x128x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[256, 128, 64],
+            stages=3, warp_count=[4, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_test_SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16_sliced_k_128x64x64_64x64x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 64, 64],
+            stages=3, warp_count=[2, 1, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32_128x256x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 256, 64],
+            stages=3, warp_count=[2, 4, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.RowMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32_128x256x64_64x64x64(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 16],
+            element_a=cutlass.float16, element_b=cutlass.float16,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 256, 64],
+            stages=3, warp_count=[2, 4, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        B = TensorDescription(
+            element=cutlass.float16, layout=cutlass.ColumnMajor,
+            alignment=8
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    
+
+if __name__ == '__main__':
+    pycutlass.get_memory_pool(2**24, 2**24)
+    unittest.main()

warp/native/cutlass/tools/library/scripts/pycutlass/test/gemm/gemm_f32_sm80.py

@@ -0,0 +1,146 @@
+import pycutlass
+from pycutlass import *
+from pycutlass.memory_manager import get_allocated_size
+from pycutlass.test import *
+import unittest
+
+from pycutlass.test.gemm_testbed import test_all_gemm
+from pycutlass.utils.device import device_cc
+
+
+@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
+class GemmF32nF32nF32nTensorOpF32Sm80(unittest.TestCase):
+    def test_SM80_Device_Gemm_f32t_f32n_f32t_tensor_op_bf16_f32_128x128x32_64x64x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 8],
+            element_a=cutlass.float32, element_b=cutlass.float32,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add_fast_bf16
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 128, 32],
+            stages=3, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float32, layout=cutlass.RowMajor,
+            alignment=4
+        )
+        B = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C,
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+
+
+    def test_SM80_Device_Gemm_f32n_f32n_f32t_tensor_op_f32_128x128x32_64x64x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 8],
+            element_a=cutlass.float32, element_b=cutlass.float32,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[128, 128, 32],
+            stages=3, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+        B = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f32n_f32n_f32t_tensor_op_fast_accurate_f32_64x64x32_32x32x32(self):
+        math_inst = MathInstruction(
+            instruction_shape=[16, 8, 8],
+            element_a=cutlass.float32, element_b=cutlass.float32,
+            element_accumulator=cutlass.float32, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add_fast_f32
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[64, 64, 32],
+            stages=3, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        A = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+        B = TensorDescription(
+            element=cutlass.float32, layout=cutlass.ColumnMajor,
+            alignment=4
+        )
+        C = TensorDescription(
+            element=cutlass.float32, layout=cutlass.RowMajor,
+            alignment=4
+        )
+
+        element_epilogue = cutlass.float32
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+
+if __name__ == '__main__':
+    pycutlass.get_memory_pool(2**24, 2**24)
+    pycutlass.compiler.load_from_cache()
+    unittest.main()

warp/native/cutlass/tools/library/scripts/pycutlass/test/gemm/gemm_f64_sm80.py

@@ -0,0 +1,102 @@
+import pycutlass
+from pycutlass import *
+from pycutlass.test import *
+import unittest
+
+from pycutlass.test.gemm_testbed import test_all_gemm
+from pycutlass.utils.device import device_cc
+
+
+@unittest.skipIf(device_cc() < 80, "Device compute capability is insufficient for SM80 tests.")
+class GemmF64TensorOpSm80(unittest.TestCase):
+    def test_SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64_32x32x16_16x16x16(self):
+        math_inst = MathInstruction(
+            instruction_shape=[8, 8, 4],
+            element_a=cutlass.float64, element_b=cutlass.float64,
+            element_accumulator=cutlass.float64, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[32, 32, 16],
+            stages=4, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        # alignment 1 restricted for double
+        A = TensorDescription(
+            element=cutlass.float64, layout=cutlass.ColumnMajor,
+            alignment=1
+        )
+        B = TensorDescription(
+            element=cutlass.float64, layout=cutlass.RowMajor,
+            alignment=1
+        )
+        C = TensorDescription(
+            element=cutlass.float64, layout=cutlass.RowMajor,
+            alignment=1
+        )
+
+        element_epilogue = cutlass.float64
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+    
+    def test_SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64_64x64x16_32x32x16(self):
+        math_inst = MathInstruction(
+            instruction_shape=[8, 8, 4],
+            element_a=cutlass.float64, element_b=cutlass.float64,
+            element_accumulator=cutlass.float64, opcode_class=cutlass.OpClass.TensorOp,
+            math_operation=MathOperation.multiply_add
+        )
+
+        tile_description = TileDescription(
+            threadblock_shape=[64, 64, 16],
+            stages=4, warp_count=[2, 2, 1],
+            math_instruction=math_inst
+        )
+
+        # alignment 1 restricted for double
+        A = TensorDescription(
+            element=cutlass.float64, layout=cutlass.RowMajor,
+            alignment=1
+        )
+        B = TensorDescription(
+            element=cutlass.float64, layout=cutlass.ColumnMajor,
+            alignment=1
+        )
+        C = TensorDescription(
+            element=cutlass.float64, layout=cutlass.RowMajor,
+            alignment=1
+        )
+
+        element_epilogue = cutlass.float64
+
+        epilogue_functor = LinearCombination(
+            C.element, C.alignment, 
+            math_inst.element_accumulator, element_epilogue)
+        
+        swizzling_functor = cutlass.IdentitySwizzle1
+
+        operation = GemmOperationUniversal(
+            arch=80, tile_description=tile_description,
+            A=A, B=B, C=C, 
+            epilogue_functor=epilogue_functor, swizzling_functor=swizzling_functor
+        )
+
+        self.assertTrue(test_all_gemm(operation, "universal"))
+
+if __name__ == '__main__':
+    pycutlass.get_memory_pool(2**24, 2**24)
+    unittest.main()