Cleanup

src/targets/gpu/kernels/include/migraphx/kernels/ck_gemm_softmax_gemm.hpp

@@ -65,9 +65,9 @@ __device__ void ck_gemm_softmax_gemm_matrix(C c, A a, B b, B1 b1, Settings s)
                                              to_ck_tensor<C>());
 
     static_assert(desc.IsValid(), "Invalid ck gemm.");
-    const float scale = s.scale;
+    
     G::Run(desc,
-           scale,
+           s.scale,
            to_ck_const_pointer(a.data()),
            to_ck_const_pointer(b.data()),
            to_ck_const_pointer(b1.data()),

test/onnx/gen_onnx.py

@@ -7799,82 +7799,3 @@ def where_mixed_test():
                                  outputs=['z'])
 
     return ([node], [c, x, y], [z])
-
-
-@onnx_test()
-def gemm_softmax_gemm_test():
-    a = helper.make_tensor_value_info('a', TensorProto.FLOAT16, [1, 1])
-    b = helper.make_tensor_value_info('b', TensorProto.FLOAT16, [1, 1])
-    # c = helper.make_tensor_value_info('c', TensorProto.FLOAT16, [1, 1])
-    b1 = helper.make_tensor_value_info('b1', TensorProto.FLOAT16, [1, 1])
-    # bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT16, [1, 1])
-    out = helper.make_tensor_value_info('out', TensorProto.FLOAT16, [1, 1])
-
-    scale_array = np.array([1])
-    bias_array = np.array([0])
-
-    scale_tensor = helper.make_tensor(name='scale',
-                                    data_type=TensorProto.FLOAT16,
-                                    dims=[1, 1],
-                                    vals=[1])
-    bias_tensor = helper.make_tensor(name='bias',
-                                    data_type=TensorProto.FLOAT16,
-                                    dims=[1, 1],
-                                    vals=[0])
-
-    gemm1 = onnx.helper.make_node('MatMul',
-                                 inputs=['a', 'b'],
-                                 outputs=['gemm1_out'])
-    mul1 = onnx.helper.make_node('Mul',
-                                 inputs=['gemm1_out', 'scale'],
-                                 outputs=['mul1_out'])
-    add1 = onnx.helper.make_node('Add',
-                                 inputs=['mul1_out', 'bias'],
-                                 outputs=['add1_out'])
-    softmax = onnx.helper.make_node('Softmax',
-                                 inputs=['add1_out'],
-                                 outputs=['softmax_out'])
-    gemm2 = onnx.helper.make_node('MatMul',
-                                 inputs=['softmax_out', 'b1'],
-                                 outputs=['out'])
-
-
-    return ([gemm1, mul1, add1, softmax, gemm2], [a, b, b1], [out], [scale_tensor, bias_tensor])
-
-
-
-@onnx_test()
-def old_gemm_softmax_gemm_test():
-    a = helper.make_tensor_value_info('a', TensorProto.FLOAT16, [1, 1])
-    b = helper.make_tensor_value_info('b', TensorProto.FLOAT16, [1, 1])
-    c = helper.make_tensor_value_info('c', TensorProto.FLOAT16, [1, 1])
-    b1 = helper.make_tensor_value_info('b1', TensorProto.FLOAT16, [1, 1])
-    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT16, [1, 1])
-    out = helper.make_tensor_value_info('out', TensorProto.FLOAT16, [1, 1])
-
-    scale_array = np.array([(1/8)])
-
-    scale_tensor = helper.make_tensor('scale',
-                                    TensorProto.FLOAT16,
-                                    [1, 1],
-                                    [1])
-
-    gemm1 = onnx.helper.make_node('MatMul',
-                                 inputs=['a', 'b'],
-                                 outputs=['gemm1_out'])
-    mul1 = onnx.helper.make_node('Mul',
-                                 inputs=['gemm1_out', 'scale'],
-                                 outputs=['mul1_out'])
-    add1 = onnx.helper.make_node('Add',
-                                 inputs=['mul1_out', 'c'],
-                                 outputs=['add1_out'])
-    softmax = onnx.helper.make_node('Softmax',
-                                 inputs=['add1_out'],
-                                 outputs=['softmax_out'])
-    gemm2 = onnx.helper.make_node('MatMul',
-                                 inputs=['softmax_out', 'b1'],
-                                 outputs=['out'])
-
-
-    return ([gemm1, mul1, add1, softmax, gemm2], [a, b, c, b1, bias], [out], [scale_tensor])
-

test/onnx/old_gemm_softmax_gemm_test.onnx

-old_gemm_softmax_gemm_test:
-
-
a
-
b	gemm1_out"MatMul
-!
-	gemm1_out
-scalemul1_out"Mul
-
-mul1_out
-
cadd1_out"Add
- 
-add1_outsoftmax_out"Softmax
-
-softmax_out
-b1out"MatMulold_gemm_softmax_gemm_test*


-*

BscaleZ
-
a
-
-
-

-
Z
-
b
-
-
-

-
Z
-
c
-
-
-

-
Z
-b1
-
-
-

-
Z
-bias
-
-
-

-
b
-out
-
-
-

-
B
\ No newline at end of file