MaskRCNN fails on MI Series

[richagadgil]

convolution_backwards compiled by MIOpen fails on MI300x but not on Navi. I have attached a unit test to replicate the issue:

struct test_convolution_backwards_miopen : verify_program<test_convolution_backwards_miopen>
{
    migraphx::program create_program() const
    {
        migraphx::program p;
        auto* mm = p.get_main_module();
        auto input =
            mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {100, 256, 14, 14}});
        auto weights =
            mm->add_parameter("w", migraphx::shape{migraphx::shape::float_type, {256, 256, 2, 2}});
        mm->add_instruction(
            migraphx::make_op("convolution_backwards",
                              {{"padding", {0, 0}}, {"stride", {2, 2}}, {"dilation", {1, 1}}, {"padding_mode", 0}, {"group", 1}}),
            input,
            weights);
        return p;
    }
};

w = @param:w -> float_type, {256, 256, 2, 2}, {1024, 4, 2, 1}
x = @param:x -> float_type, {100, 256, 14, 14}, {50176, 196, 14, 1}
@2 = allocate[shape=float_type, {100, 256, 28, 28}, {200704, 784, 28, 1},buf_type=nullopt] -> float_type, {100, 256, 28, 28}, {200704, 784, 28, 1}
@3 = gpu::miopen_op[op=gpu::convolution_backwards[op={padding={0, 0},stride={2, 2},dilation={1, 1},padding_mode=0,group=1},solution_object={binary_object: 0},algo=0,solution_id=0]](x,w,@2) -> float_type, {100, 256, 28, 28}, {200704, 784, 28, 1}

MI300X:

Memory access fault by GPU node-2 (Agent handle: 0x7ff7ac016300) on address 0x7fb3a6e0c000. Reason: Unknown.
GPU core dump created: gpucore.33142
Aborted (core dumped)

Navi31:

[ COMPLETE ] test_convolution_backwards_maskrcnn